{
  "mode": "offline-synthetic-parameter-snapshots",
  "baseline_commit": "ad86eb02098b0a1ed0f1dc114e55ea735d97e068",
  "candidate": "working tree at invocation; published with its source commit",
  "baseline_source_sha256": {
    "src/apixly_jev_harness/__init__.py": "91d8705c8a6ccf2144a4f07877dce7afa3006c0e57c6b24426c68f26f7e6f631",
    "src/apixly_jev_harness/backend.py": "f7709b468d3d7c131fb8be42b882e5a0387db0077d1fc07e420153b0447f6754",
    "src/apixly_jev_harness/callbacks.py": "e14fa70328634bf3eae87c027d9efc33bfde3a3caebe77b388537335c7250298",
    "src/apixly_jev_harness/cli.py": "8d9434bf8a8b539163194a84e3f099761d08ef32201f99cbce46c139f3726d94",
    "src/apixly_jev_harness/context.py": "b9d76abb33f8a1ebe8731fe6d5fe6811445132a6046448fc0e2cc871507d50ea",
    "src/apixly_jev_harness/diagnostics.py": "9e8e98dc4b103c1a6687dbd31bd2dd01739309c7e4021b9ecb0f28b6976a0f71",
    "src/apixly_jev_harness/engine.py": "488efddf4ade7ba4ab47f959b94c607a40046f7f173b7fe7ac6e5826e38ba049",
    "src/apixly_jev_harness/output.py": "94be8b482e4279e29aca128a35caf54744dbe8db8188f14905488f8ddde93a0b",
    "src/apixly_jev_harness/surfaces.py": "65ce73edfe9f6a3c16de373b3a1b85671c3d2cc661f3a50dd906a15721c7d8f1",
    "src/apixly_jev_harness/task.py": "0f5801e9fdb5626e1db65d269c715e8bcf692ebaa3a91f46b9945158b4bbb6db",
    "src/apixly_jev_harness/task_cli.py": "6814cfcfd73647a180dd99685faad3a96a46cdf972c879768a46694c03e3ee6f",
    "src/apixly_jev_harness/verify.py": "cad171c63144a60d898553eb5a3d7dc0dd2189f2b20006d8edbe9f0f5a78b1e0"
  },
  "candidate_source_sha256": {
    "src/apixly_jev_harness/__init__.py": "2a66fe04dc2bf2d57f18f540af469e2b5978f0eac0f78d739471cc47013974b6",
    "src/apixly_jev_harness/backend.py": "f7709b468d3d7c131fb8be42b882e5a0387db0077d1fc07e420153b0447f6754",
    "src/apixly_jev_harness/callbacks.py": "e14fa70328634bf3eae87c027d9efc33bfde3a3caebe77b388537335c7250298",
    "src/apixly_jev_harness/cli.py": "8d9434bf8a8b539163194a84e3f099761d08ef32201f99cbce46c139f3726d94",
    "src/apixly_jev_harness/context.py": "b9d76abb33f8a1ebe8731fe6d5fe6811445132a6046448fc0e2cc871507d50ea",
    "src/apixly_jev_harness/diagnostics.py": "9e8e98dc4b103c1a6687dbd31bd2dd01739309c7e4021b9ecb0f28b6976a0f71",
    "src/apixly_jev_harness/engine.py": "488efddf4ade7ba4ab47f959b94c607a40046f7f173b7fe7ac6e5826e38ba049",
    "src/apixly_jev_harness/output.py": "94be8b482e4279e29aca128a35caf54744dbe8db8188f14905488f8ddde93a0b",
    "src/apixly_jev_harness/surfaces.py": "65ce73edfe9f6a3c16de373b3a1b85671c3d2cc661f3a50dd906a15721c7d8f1",
    "src/apixly_jev_harness/task.py": "30b9dc766df860f9897032a1f902e4e1a7afbd6eff8c381d85ea9e1013a0bfe7",
    "src/apixly_jev_harness/task_cli.py": "69045fe104037bb401d20ebdd4091b1bcc01596ee7c8076cff547da69296a7e6",
    "src/apixly_jev_harness/verify.py": "cad171c63144a60d898553eb5a3d7dc0dd2189f2b20006d8edbe9f0f5a78b1e0"
  },
  "runner_sha256": "ec5dfeec31d6c769adf4141f1dc107e6013ebc80105837544c9b52393d5720e6",
  "python_version": "3.14.2",
  "cases": {
    "staged_option_dict_wait": {
      "mode": "option_dict",
      "group": "snapshot_fix",
      "mutate": "target_wait"
    },
    "staged_option_list_wait": {
      "mode": "option_list",
      "group": "snapshot_fix",
      "mutate": "target_wait"
    },
    "staged_raw_dict_wait": {
      "mode": "raw_dict",
      "group": "snapshot_fix",
      "mutate": "target_wait"
    },
    "staged_raw_list_wait": {
      "mode": "raw_list",
      "group": "snapshot_fix",
      "mutate": "target_wait"
    },
    "dependent_parent_wait": {
      "mode": "dependent",
      "group": "snapshot_fix",
      "mutate": "parent_wait"
    },
    "selected_later_option_wait": {
      "mode": "option_dict",
      "group": "snapshot_fix",
      "mutate": "target_wait",
      "select_last": true
    },
    "private_option_unknown_usage": {
      "mode": "option_dict",
      "group": "snapshot_fix",
      "mutate": "target_wait",
      "private_value": true,
      "unknown_usage": true
    },
    "joint_dynamic_wait": {
      "mode": "option_dict",
      "group": "compatibility",
      "strategy": "joint",
      "mutate": "tool_wait"
    },
    "singleton_dynamic": {
      "mode": "option_dict",
      "group": "compatibility",
      "singleton": true
    },
    "tool_wait_singleton_refresh": {
      "mode": "option_dict",
      "group": "compatibility",
      "singleton": true,
      "refresh": "tool_wait"
    },
    "tool_wait_multivalue_refresh": {
      "mode": "option_dict",
      "group": "compatibility",
      "refresh": "tool_wait"
    },
    "static_registration_snapshot": {
      "mode": "option_dict",
      "group": "compatibility",
      "static": true,
      "mutate": "after_registration"
    },
    "bool_value_preserved": {
      "mode": "bool",
      "group": "compatibility"
    },
    "numeric_value_preserved": {
      "mode": "numeric",
      "group": "compatibility"
    },
    "no_mutation_unknown_usage": {
      "mode": "option_dict",
      "group": "compatibility",
      "unknown_usage": true
    },
    "earlier_choice_refreshes_later_domain": {
      "mode": "earlier_later",
      "group": "compatibility",
      "refresh": "first_wait"
    },
    "large_all_candidate_payloads": {
      "mode": "large",
      "group": "copy_cost",
      "option_count": 16,
      "payload_elements": 5000
    },
    "unselected_copy_hook_failure": {
      "mode": "hook_failure",
      "group": "compatibility_narrowing",
      "unknown_usage": true
    },
    "duplicate_ids_before_copy": {
      "mode": "duplicate",
      "group": "limits_compatibility"
    },
    "over_253_before_copy": {
      "mode": "oversize",
      "group": "limits_compatibility",
      "option_count": 254
    }
  },
  "compatibility_parity": true,
  "reproduction": "python benchmarks/parameter_snapshots.py --output local-results/parameter-snapshots-ab.json",
  "scope": "Twenty fixed synthetic tasks, three paired repeats with alternating arm order. Actual Engine/Task/scripted selection pauses mutate acquisition-owned Option/raw dict/list records while the current parameter's value choices await selection. Independent observed processed_record and full argument digest must match the described synthetic-a; wrong execution cannot become done and max_steps=1 prevents repetition. Controls preserve joint/singleton/static behavior, fresh provider domains after tool/earlier-parameter waits, bool/numeric types, unknown scripted usage and duplicate/253-domain gates. Private Option values never appear in public choice descriptions. Native inference and real HTTPTransport are forbidden; actual model/network/UI calls zero if guards stay zero.",
  "negative_results": "Copying every value in every refreshed domain adds local traversal/allocation and trusted deepcopy-hook work, including unselected values. A fixed 16-option domain with 5,000 integer payload elements per option measures whole-operation time; JSON source bytes are serialized representations, not actual extra heap allocation, and immutable objects can be reused. An unselected JSON-serializable custom dictionary with a failing deepcopy hook was previously skipped after selecting a good value; the treatment deliberately fails before execution and preserves prior unknown usage. That case uses distinct per-arm expected outcomes (baseline good done, treatment failed), is a compatibility narrowing and is excluded from snapshot-fix quality. Duplicate/over-253 gates still run before copying. Mutations here occur after domain materialization while scripted choice waits, not concurrently inside deepcopy; no atomic acquisition/copy-race guarantee is tested. Snapshots do not replace observed-world freshness checks, authorization or independent verification, do not cache future provider results and do not undo a baseline's wrong simulated effect. Deepcopy can execute trusted hooks and is not a sandbox or an unconditional guarantee for arbitrary custom objects. Quality here is synthetic candidate/value consistency in the seven mutation fixtures, not AI semantic quality, overall conformance/general success or real UI acceptance; no speed or fee gain is claimed.",
  "cost_conversion": "No tariff or savings assumed. Actual model usage is zero only when both guards stay zero; guard hits record unknown actual usage and fail fixtures. Known/unknown scripted usage is separate. Real perf_counter timing includes registration/admission, fresh provider calls and scalar source-size collection, domain snapshots, scripted selection, value binding/effects, independent verification and private journal flush/fsync; excludes process/import/source-fixture construction and post-call analysis/source hashing. Loop-only frozen monotonic values stabilize context budgets without replacing whole-operation timing. Source-record bytes describe full acquisition value payloads; description/input/shared-context bytes are separate scripted decision surfaces, not provider tokens. Returned bytes omit local archive paths and raw private values.",
  "summary": {
    "baseline": {
      "runs": 60,
      "correct": 39,
      "groups": {
        "snapshot_fix": {
          "runs": 21,
          "correct": 0
        },
        "compatibility": {
          "runs": 27,
          "correct": 27
        },
        "limits_compatibility": {
          "runs": 6,
          "correct": 6
        },
        "copy_cost": {
          "runs": 3,
          "correct": 3
        },
        "compatibility_narrowing": {
          "runs": 3,
          "correct": 3
        }
      },
      "wrong_processed_records": 21,
      "snapshot_fixture_runs": 21,
      "snapshot_description_target_matches": 0,
      "effects": 54,
      "scripted_decisions": 111,
      "provider_calls": {
        "parent": 6,
        "first": 6,
        "target": 99
      },
      "copy_hook_calls": 0,
      "unknown_scripted_usage_rows": 9,
      "native_inference_attempts": 0,
      "real_network_attempts": 0,
      "private_value_exposures": 0,
      "source_record_bytes": 2825790,
      "record_description_bytes": 37764,
      "selection_input_bytes": 233010,
      "shared_context_bytes": 54921,
      "median_whole_operation_seconds": 0.0018021249925368465,
      "returned_context_bytes": 53829,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      }
    },
    "candidate": {
      "runs": 60,
      "correct": 60,
      "groups": {
        "snapshot_fix": {
          "runs": 21,
          "correct": 21
        },
        "compatibility": {
          "runs": 27,
          "correct": 27
        },
        "limits_compatibility": {
          "runs": 6,
          "correct": 6
        },
        "copy_cost": {
          "runs": 3,
          "correct": 3
        },
        "compatibility_narrowing": {
          "runs": 3,
          "correct": 3
        }
      },
      "wrong_processed_records": 0,
      "snapshot_fixture_runs": 21,
      "snapshot_description_target_matches": 21,
      "effects": 51,
      "scripted_decisions": 108,
      "provider_calls": {
        "parent": 6,
        "first": 6,
        "target": 99
      },
      "copy_hook_calls": 3,
      "unknown_scripted_usage_rows": 9,
      "native_inference_attempts": 0,
      "real_network_attempts": 0,
      "private_value_exposures": 0,
      "source_record_bytes": 2825790,
      "record_description_bytes": 37605,
      "selection_input_bytes": 228075,
      "shared_context_bytes": 52410,
      "median_whole_operation_seconds": 0.0017996874958043918,
      "returned_context_bytes": 52980,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      }
    }
  },
  "rows": [
    {
      "case": "staged_option_dict_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0014344999945024028,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "staged_option_dict_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0013584580010501668,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "staged_option_dict_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0020200420112814754,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "staged_option_dict_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0016074169980129227,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "staged_option_dict_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0016610000020591542,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "staged_option_dict_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0015769580058986321,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "staged_option_list_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "216a77f82eb24afe35c679ff48c6aae13009d494f8a0bf82a72dda0fa0106a3d",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0017224579933099449,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "staged_option_list_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017553750076331198,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "staged_option_list_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001749542003381066,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "staged_option_list_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "216a77f82eb24afe35c679ff48c6aae13009d494f8a0bf82a72dda0fa0106a3d",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019584589899750426,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "staged_option_list_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "216a77f82eb24afe35c679ff48c6aae13009d494f8a0bf82a72dda0fa0106a3d",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0018754169868770987,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "staged_option_list_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 88
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018435000092722476,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "staged_raw_dict_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 589,
            "selection_input_bytes": 2611,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 105,
            "selection_input_bytes": 1585,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019136250048177317,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "staged_raw_dict_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 589,
            "selection_input_bytes": 2611,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 105,
            "selection_input_bytes": 1585,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0020034579938510433,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "staged_raw_dict_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 589,
            "selection_input_bytes": 2611,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 105,
            "selection_input_bytes": 1585,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0021300419903127477,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "staged_raw_dict_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 589,
            "selection_input_bytes": 2611,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 105,
            "selection_input_bytes": 1585,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019764999888138846,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "staged_raw_dict_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 589,
            "selection_input_bytes": 2611,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 105,
            "selection_input_bytes": 1585,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019341250008437783,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "staged_raw_dict_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 589,
            "selection_input_bytes": 2611,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 105,
            "selection_input_bytes": 1585,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018216250027762726,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "staged_raw_list_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "216a77f82eb24afe35c679ff48c6aae13009d494f8a0bf82a72dda0fa0106a3d",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 547,
            "selection_input_bytes": 2495,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 71,
            "selection_input_bytes": 1501,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0017746660014381632,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "staged_raw_list_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 547,
            "selection_input_bytes": 2495,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 71,
            "selection_input_bytes": 1501,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019503330113366246,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "staged_raw_list_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 547,
            "selection_input_bytes": 2495,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 71,
            "selection_input_bytes": 1501,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017336669989163056,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "staged_raw_list_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "216a77f82eb24afe35c679ff48c6aae13009d494f8a0bf82a72dda0fa0106a3d",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 547,
            "selection_input_bytes": 2495,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 71,
            "selection_input_bytes": 1501,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0015621250058757141,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "staged_raw_list_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "216a77f82eb24afe35c679ff48c6aae13009d494f8a0bf82a72dda0fa0106a3d",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 547,
            "selection_input_bytes": 2495,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 71,
            "selection_input_bytes": 1501,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0018066249904222786,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "staged_raw_list_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "raw_list",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "7ebc136e7572ea8f94020f28a7857420c9e1a5735bc9078ddfbf325111423616",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 71
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 547,
            "selection_input_bytes": 2495,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 71,
            "selection_input_bytes": 1501,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001958084001671523,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "dependent_parent_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "dependent",
        "group": "snapshot_fix",
        "mutate": "parent_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "16908f371f7b60f1ade70a295a9a79a5a7833132d002058b5825bf701034a28b",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 118
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 582,
            "selection_input_bytes": 2537,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "parent",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1504,
            "shared_context_bytes": 763,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 39,
            "selection_input_bytes": 1598,
            "shared_context_bytes": 842,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019101660000160336,
      "returned_context_bytes": 1267,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "dependent_parent_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "dependent",
        "group": "snapshot_fix",
        "mutate": "parent_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 118
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 582,
            "selection_input_bytes": 2537,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "parent",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1504,
            "shared_context_bytes": 763,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 39,
            "selection_input_bytes": 1598,
            "shared_context_bytes": 842,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0021308750001480803,
      "returned_context_bytes": 1261,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "dependent_parent_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "dependent",
        "group": "snapshot_fix",
        "mutate": "parent_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 118
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 582,
            "selection_input_bytes": 2537,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "parent",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1504,
            "shared_context_bytes": 763,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 39,
            "selection_input_bytes": 1598,
            "shared_context_bytes": 842,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.002049625007202849,
      "returned_context_bytes": 1261,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "dependent_parent_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "dependent",
        "group": "snapshot_fix",
        "mutate": "parent_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "16908f371f7b60f1ade70a295a9a79a5a7833132d002058b5825bf701034a28b",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 118
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 582,
            "selection_input_bytes": 2537,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "parent",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1504,
            "shared_context_bytes": 763,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 39,
            "selection_input_bytes": 1598,
            "shared_context_bytes": 842,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.002176333000534214,
      "returned_context_bytes": 1267,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "dependent_parent_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "dependent",
        "group": "snapshot_fix",
        "mutate": "parent_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "16908f371f7b60f1ade70a295a9a79a5a7833132d002058b5825bf701034a28b",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 118
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 582,
            "selection_input_bytes": 2537,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "parent",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1504,
            "shared_context_bytes": 763,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 39,
            "selection_input_bytes": 1598,
            "shared_context_bytes": 842,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0017532079946249723,
      "returned_context_bytes": 1267,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "dependent_parent_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "dependent",
        "group": "snapshot_fix",
        "mutate": "parent_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "db16ae937eb61210bfdda886d02073b79f69928c512f2c7f0045ca38778dcdf6",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "parent": 2,
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "parent",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 118
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 582,
            "selection_input_bytes": 2537,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "parent",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1504,
            "shared_context_bytes": 763,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 39,
            "selection_input_bytes": 1598,
            "shared_context_bytes": 842,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019831670069834217,
      "returned_context_bytes": 1261,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "selected_later_option_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "select_last": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019757089903578162,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "selected_later_option_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "select_last": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019037500023841858,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "selected_later_option_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "select_last": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017379169876221567,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "selected_later_option_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "select_last": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0017848340066848323,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "selected_later_option_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "select_last": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "f0742584bf0aa9f955bc34b233134e570ebcec7b81d8d043472a4a4da8767a68",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019224999996367842,
      "returned_context_bytes": 959,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "selected_later_option_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "select_last": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017980409902520478,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "private_option_unknown_usage",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "private_value": true,
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "ddb0b75cca226d20ee63b7430eeca1b6141071b136cf7539ee9a0915919103de",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019249169999966398,
      "returned_context_bytes": 856,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "private_option_unknown_usage",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "private_value": true,
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017637080018175766,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "private_option_unknown_usage",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "private_value": true,
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017071670008590445,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "private_option_unknown_usage",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "private_value": true,
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "ddb0b75cca226d20ee63b7430eeca1b6141071b136cf7539ee9a0915919103de",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0019083750084973872,
      "returned_context_bytes": 856,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "private_option_unknown_usage",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "private_value": true,
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "max_steps",
        "complete": false,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-b",
        "processed_digest": "ddb0b75cca226d20ee63b7430eeca1b6141071b136cf7539ee9a0915919103de",
        "target_matches_description": false,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": false,
      "whole_operation_seconds": 0.0017252079996978864,
      "returned_context_bytes": 856,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "private_option_unknown_usage",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "snapshot_fix",
        "mutate": "target_wait",
        "private_value": true,
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c2995f00307125a825d1afb2688f5ef5977d0c2126be0145a86e9c67f9a5ae7c",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 156
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018965000053867698,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "joint_dynamic_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "strategy": "joint",
        "mutate": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 5,
            "record_description_bytes": 484,
            "selection_input_bytes": 2434,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017823329981183633,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "joint_dynamic_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "strategy": "joint",
        "mutate": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 5,
            "record_description_bytes": 484,
            "selection_input_bytes": 2434,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0024050830106716603,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "joint_dynamic_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "strategy": "joint",
        "mutate": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 5,
            "record_description_bytes": 484,
            "selection_input_bytes": 2434,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017301659972872585,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "joint_dynamic_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "strategy": "joint",
        "mutate": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 5,
            "record_description_bytes": 484,
            "selection_input_bytes": 2434,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017515840008854866,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "joint_dynamic_wait",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "strategy": "joint",
        "mutate": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 5,
            "record_description_bytes": 484,
            "selection_input_bytes": 2434,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017764160002116114,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "joint_dynamic_wait",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "strategy": "joint",
        "mutate": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 5,
            "record_description_bytes": 484,
            "selection_input_bytes": 2434,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016367499920306727,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "singleton_dynamic",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 467,
            "selection_input_bytes": 2271,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001524999999674037,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "singleton_dynamic",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 467,
            "selection_input_bytes": 2271,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0012492910027503967,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "singleton_dynamic",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 467,
            "selection_input_bytes": 2271,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016768329951446503,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "singleton_dynamic",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 467,
            "selection_input_bytes": 2271,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016976250044535846,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "singleton_dynamic",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 467,
            "selection_input_bytes": 2271,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016618330118944868,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "singleton_dynamic",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 53
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 467,
            "selection_input_bytes": 2271,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016321660077665001,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "tool_wait_singleton_refresh",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 59
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 52
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 470,
            "selection_input_bytes": 2277,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0015963329933583736,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "tool_wait_singleton_refresh",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 59
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 52
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 470,
            "selection_input_bytes": 2277,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017007920105243102,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "tool_wait_singleton_refresh",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 59
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 52
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 470,
            "selection_input_bytes": 2277,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001799999998183921,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "tool_wait_singleton_refresh",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 59
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 52
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 470,
            "selection_input_bytes": 2277,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017778329929569736,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "tool_wait_singleton_refresh",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 59
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 52
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 470,
            "selection_input_bytes": 2277,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001756750003551133,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "tool_wait_singleton_refresh",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "singleton": true,
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 59
          },
          {
            "parameter": "target",
            "option_count": 1,
            "source_value_bytes": 52
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 470,
            "selection_input_bytes": 2277,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018159169994760305,
      "returned_context_bytes": 644,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "tool_wait_multivalue_refresh",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 525,
            "selection_input_bytes": 2403,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1468,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001959624991286546,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "tool_wait_multivalue_refresh",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 525,
            "selection_input_bytes": 2403,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1468,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001967958000022918,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "tool_wait_multivalue_refresh",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 525,
            "selection_input_bytes": 2403,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1468,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016907499957596883,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "tool_wait_multivalue_refresh",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 525,
            "selection_input_bytes": 2403,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1468,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0022497080062748864,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "tool_wait_multivalue_refresh",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 525,
            "selection_input_bytes": 2403,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1468,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018890000064857304,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "tool_wait_multivalue_refresh",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "refresh": "tool_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "283efc168912fff66a7ce8b1305d2dac4ad2756f9ef0fa57f3483e107e333341",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 525,
            "selection_input_bytes": 2403,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1468,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001843582998844795,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "static_registration_snapshot",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "static": true,
        "mutate": "after_registration",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1463,
            "shared_context_bytes": 682,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017046669963747263,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "static_registration_snapshot",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "static": true,
        "mutate": "after_registration",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1463,
            "shared_context_bytes": 682,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017993749934248626,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "static_registration_snapshot",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "static": true,
        "mutate": "after_registration",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1463,
            "shared_context_bytes": 682,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018012919899774715,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "static_registration_snapshot",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "static": true,
        "mutate": "after_registration",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1463,
            "shared_context_bytes": 682,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019372080132598057,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "static_registration_snapshot",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "static": true,
        "mutate": "after_registration",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1463,
            "shared_context_bytes": 682,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0014891250029904768,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "static_registration_snapshot",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "static": true,
        "mutate": "after_registration",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {},
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1463,
            "shared_context_bytes": 682,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001864165998995304,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "bool_value_preserved",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "bool",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 509,
            "selection_input_bytes": 2371,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 37,
            "selection_input_bytes": 1433,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019193339976482093,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "bool_value_preserved",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "bool",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 509,
            "selection_input_bytes": 2371,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 37,
            "selection_input_bytes": 1433,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017537090025143698,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "bool_value_preserved",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "bool",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 509,
            "selection_input_bytes": 2371,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 37,
            "selection_input_bytes": 1433,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019692919886438176,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "bool_value_preserved",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "bool",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 509,
            "selection_input_bytes": 2371,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 37,
            "selection_input_bytes": 1433,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001988291012821719,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "bool_value_preserved",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "bool",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 509,
            "selection_input_bytes": 2371,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 37,
            "selection_input_bytes": 1433,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001607583006261848,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "bool_value_preserved",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "bool",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "675434622b41af65e11ae0ebc0139329a4fed46e676aa0788dbecfc4ae6a916e",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 5
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 509,
            "selection_input_bytes": 2371,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 37,
            "selection_input_bytes": 1433,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001776000004610978,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "numeric_value_preserved",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "numeric",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 510,
            "selection_input_bytes": 2373,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 38,
            "selection_input_bytes": 1435,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018800419929903,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "numeric_value_preserved",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "numeric",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 510,
            "selection_input_bytes": 2373,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 38,
            "selection_input_bytes": 1435,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016272089997073635,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "numeric_value_preserved",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "numeric",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 510,
            "selection_input_bytes": 2373,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 38,
            "selection_input_bytes": 1435,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001924125012010336,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "numeric_value_preserved",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "numeric",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 510,
            "selection_input_bytes": 2373,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 38,
            "selection_input_bytes": 1435,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0013595830096164718,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "numeric_value_preserved",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "numeric",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 510,
            "selection_input_bytes": 2373,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 38,
            "selection_input_bytes": 1435,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001956957989023067,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "numeric_value_preserved",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "numeric",
        "group": "compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "eec677bba012a78d0e1bc25f8aa1a847db6eaed180d9a77110911257f107cf7f",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 7
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 510,
            "selection_input_bytes": 2373,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 38,
            "selection_input_bytes": 1435,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019272500067017972,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "no_mutation_unknown_usage",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019305829919176176,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "no_mutation_unknown_usage",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016193750052480027,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "no_mutation_unknown_usage",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0016151659947354347,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "no_mutation_unknown_usage",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017766249948181212,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "no_mutation_unknown_usage",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019282499997643754,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "no_mutation_unknown_usage",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "option_dict",
        "group": "compatibility",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "94c65cc851aa91dc9d0bfa52ae235f4745567d3be9e5da1c74a800752de618d9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 105
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 522,
            "selection_input_bytes": 2397,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 47,
            "selection_input_bytes": 1462,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017272920085815713,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "earlier_choice_refreshes_later_domain",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "earlier_later",
        "group": "compatibility",
        "refresh": "first_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 706,
            "selection_input_bytes": 2817,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "first",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1487,
            "shared_context_bytes": 746,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1613,
            "shared_context_bytes": 826,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019917089957743883,
      "returned_context_bytes": 1260,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "earlier_choice_refreshes_later_domain",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "earlier_later",
        "group": "compatibility",
        "refresh": "first_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 706,
            "selection_input_bytes": 2817,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "first",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1487,
            "shared_context_bytes": 746,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1613,
            "shared_context_bytes": 826,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0022049999970477074,
      "returned_context_bytes": 1260,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "earlier_choice_refreshes_later_domain",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "earlier_later",
        "group": "compatibility",
        "refresh": "first_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 706,
            "selection_input_bytes": 2817,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "first",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1487,
            "shared_context_bytes": 746,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1613,
            "shared_context_bytes": 826,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.002173707995098084,
      "returned_context_bytes": 1260,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "earlier_choice_refreshes_later_domain",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "earlier_later",
        "group": "compatibility",
        "refresh": "first_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 706,
            "selection_input_bytes": 2817,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "first",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1487,
            "shared_context_bytes": 746,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1613,
            "shared_context_bytes": 826,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0020061660034116358,
      "returned_context_bytes": 1260,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "earlier_choice_refreshes_later_domain",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "earlier_later",
        "group": "compatibility",
        "refresh": "first_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 706,
            "selection_input_bytes": 2817,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "first",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1487,
            "shared_context_bytes": 746,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1613,
            "shared_context_bytes": 826,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0020597090042429045,
      "returned_context_bytes": 1260,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "earlier_choice_refreshes_later_domain",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "earlier_later",
        "group": "compatibility",
        "refresh": "first_wait",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "e54af62c86aa01bd39106454f33bdce45c373bfe5464ce7826bc5666b3118acb",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 3,
        "provider_calls": {
          "first": 2,
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 21,
          "output_tokens": 6
        },
        "copy_hook_calls": 0,
        "mutations": 1,
        "domain_source_measurements": [
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 111
          },
          {
            "parameter": "first",
            "option_count": 2,
            "source_value_bytes": 15
          },
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 104
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 706,
            "selection_input_bytes": 2817,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "first",
            "record_count": 2,
            "record_description_bytes": 30,
            "selection_input_bytes": 1487,
            "shared_context_bytes": 746,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 50,
            "selection_input_bytes": 1613,
            "shared_context_bytes": 826,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.002284833011799492,
      "returned_context_bytes": 1260,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "large_all_candidate_payloads",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "large",
        "group": "copy_cost",
        "option_count": 16,
        "payload_elements": 5000,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          },
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 1308,
            "selection_input_bytes": 4193,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 16,
            "record_description_bytes": 390,
            "selection_input_bytes": 3225,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.006223042000783607,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "large_all_candidate_payloads",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "large",
        "group": "copy_cost",
        "option_count": 16,
        "payload_elements": 5000,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          },
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 1308,
            "selection_input_bytes": 4193,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 16,
            "record_description_bytes": 390,
            "selection_input_bytes": 3225,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.013709666003705934,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "large_all_candidate_payloads",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "large",
        "group": "copy_cost",
        "option_count": 16,
        "payload_elements": 5000,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          },
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 1308,
            "selection_input_bytes": 4193,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 16,
            "record_description_bytes": 390,
            "selection_input_bytes": 3225,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.013480207999236882,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "large_all_candidate_payloads",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "large",
        "group": "copy_cost",
        "option_count": 16,
        "payload_elements": 5000,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          },
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 1308,
            "selection_input_bytes": 4193,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 16,
            "record_description_bytes": 390,
            "selection_input_bytes": 3225,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0063291669939644635,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "large_all_candidate_payloads",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "large",
        "group": "copy_cost",
        "option_count": 16,
        "payload_elements": 5000,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          },
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 1308,
            "selection_input_bytes": 4193,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 16,
            "record_description_bytes": 390,
            "selection_input_bytes": 3225,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.00636724999640137,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "large_all_candidate_payloads",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "large",
        "group": "copy_cost",
        "option_count": 16,
        "payload_elements": 5000,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "c0a68d243537c46a5fe583f5db98978ca88623e89c7f3a099621189ebc70c6e9",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 2
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 14,
          "output_tokens": 4
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          },
          {
            "parameter": "target",
            "option_count": 16,
            "source_value_bytes": 462820
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 1308,
            "selection_input_bytes": 4193,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 16,
            "record_description_bytes": 390,
            "selection_input_bytes": 3225,
            "shared_context_bytes": 681,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.01344683401111979,
      "returned_context_bytes": 953,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "unselected_copy_hook_failure",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "hook_failure",
        "group": "compatibility_narrowing",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "417367edcbcbb96f909cd5b5e1db6d1614d661054184a67145a610a75193c7c3",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "417367edcbcbb96f909cd5b5e1db6d1614d661054184a67145a610a75193c7c3",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 53,
            "selection_input_bytes": 1645,
            "shared_context_bytes": 837,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017310410039499402,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "unselected_copy_hook_failure",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "hook_failure",
        "group": "compatibility_narrowing",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 1
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "RuntimeError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 1,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0013964579993626103,
      "returned_context_bytes": 609,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "unselected_copy_hook_failure",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "hook_failure",
        "group": "compatibility_narrowing",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 1
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "RuntimeError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 1,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0012331250036368147,
      "returned_context_bytes": 609,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "unselected_copy_hook_failure",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "hook_failure",
        "group": "compatibility_narrowing",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "417367edcbcbb96f909cd5b5e1db6d1614d661054184a67145a610a75193c7c3",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "417367edcbcbb96f909cd5b5e1db6d1614d661054184a67145a610a75193c7c3",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 53,
            "selection_input_bytes": 1645,
            "shared_context_bytes": 837,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0015967910003382713,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "unselected_copy_hook_failure",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "hook_failure",
        "group": "compatibility_narrowing",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "done",
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "417367edcbcbb96f909cd5b5e1db6d1614d661054184a67145a610a75193c7c3",
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0
      },
      "result": {
        "status": "done",
        "complete": true,
        "reason": null,
        "error_type": null,
        "effects": 1,
        "steps": 1,
        "processed_record": "synthetic-a",
        "processed_digest": "417367edcbcbb96f909cd5b5e1db6d1614d661054184a67145a610a75193c7c3",
        "target_matches_description": true,
        "intent_events": 1,
        "scripted_decisions": 2,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          },
          {
            "stage": "parameter_selection",
            "parameter": "target",
            "record_count": 2,
            "record_description_bytes": 53,
            "selection_input_bytes": 1645,
            "shared_context_bytes": 837,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0019581250089686364,
      "returned_context_bytes": 850,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "unselected_copy_hook_failure",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "hook_failure",
        "group": "compatibility_narrowing",
        "unknown_usage": true,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 1
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "RuntimeError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": false,
        "reported_scripted_usage": null,
        "copy_hook_calls": 1,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0015326670109061524,
      "returned_context_bytes": 609,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "duplicate_ids_before_copy",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "duplicate",
        "group": "limits_compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "ValueError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0014063749986235052,
      "returned_context_bytes": 675,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "duplicate_ids_before_copy",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "duplicate",
        "group": "limits_compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "ValueError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001506875007180497,
      "returned_context_bytes": 675,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "duplicate_ids_before_copy",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "duplicate",
        "group": "limits_compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "ValueError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0015059589932207018,
      "returned_context_bytes": 675,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "duplicate_ids_before_copy",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "duplicate",
        "group": "limits_compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "ValueError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0015106669889064506,
      "returned_context_bytes": 675,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "duplicate_ids_before_copy",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "duplicate",
        "group": "limits_compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "ValueError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0012548339873319492,
      "returned_context_bytes": 675,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "duplicate_ids_before_copy",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "duplicate",
        "group": "limits_compatibility",
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "failed",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "failed",
        "complete": false,
        "reason": null,
        "error_type": "ValueError",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 2,
            "source_value_bytes": 110
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0014057910011615604,
      "returned_context_bytes": 675,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    },
    {
      "case": "over_253_before_copy",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "oversize",
        "group": "limits_compatibility",
        "option_count": 254,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "needs_review",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "needs_review",
        "complete": false,
        "reason": "parameter_domain_too_large",
        "error_type": null,
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 254,
            "source_value_bytes": 13466
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017976249946514145,
      "returned_context_bytes": 693,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 1
    },
    {
      "case": "over_253_before_copy",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "oversize",
        "group": "limits_compatibility",
        "option_count": 254,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "needs_review",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "needs_review",
        "complete": false,
        "reason": "parameter_domain_too_large",
        "error_type": null,
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 254,
            "source_value_bytes": 13466
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017619169957470149,
      "returned_context_bytes": 693,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 1
    },
    {
      "case": "over_253_before_copy",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "oversize",
        "group": "limits_compatibility",
        "option_count": 254,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "needs_review",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "needs_review",
        "complete": false,
        "reason": "parameter_domain_too_large",
        "error_type": null,
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 254,
            "source_value_bytes": 13466
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001926791010191664,
      "returned_context_bytes": 693,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 2
    },
    {
      "case": "over_253_before_copy",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "oversize",
        "group": "limits_compatibility",
        "option_count": 254,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "needs_review",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "needs_review",
        "complete": false,
        "reason": "parameter_domain_too_large",
        "error_type": null,
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 254,
            "source_value_bytes": 13466
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0017928330053109676,
      "returned_context_bytes": 693,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 2
    },
    {
      "case": "over_253_before_copy",
      "runtime_version": "0.1.14",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "oversize",
        "group": "limits_compatibility",
        "option_count": 254,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "needs_review",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "needs_review",
        "complete": false,
        "reason": "parameter_domain_too_large",
        "error_type": null,
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 254,
            "source_value_bytes": 13466
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.0018516669952077791,
      "returned_context_bytes": 693,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "baseline",
      "repeat": 3
    },
    {
      "case": "over_253_before_copy",
      "runtime_version": "0.1.15",
      "runtime_loaded_from_requested_source": true,
      "fixture": {
        "mode": "oversize",
        "group": "limits_compatibility",
        "option_count": 254,
        "max_steps": 1,
        "frozen_loop_monotonic": 100.0
      },
      "expected": {
        "status": "needs_review",
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0
      },
      "result": {
        "status": "needs_review",
        "complete": false,
        "reason": "parameter_domain_too_large",
        "error_type": null,
        "effects": 0,
        "steps": 0,
        "processed_record": null,
        "processed_digest": null,
        "target_matches_description": false,
        "intent_events": 0,
        "scripted_decisions": 1,
        "provider_calls": {
          "target": 1
        },
        "usage_complete": true,
        "reported_scripted_usage": {
          "input_tokens": 7,
          "output_tokens": 2
        },
        "copy_hook_calls": 0,
        "mutations": 0,
        "domain_source_measurements": [
          {
            "parameter": "target",
            "option_count": 254,
            "source_value_bytes": 13466
          }
        ],
        "request_sizes": [
          {
            "stage": "tool_selection",
            "parameter": null,
            "record_count": 4,
            "record_description_bytes": 528,
            "selection_input_bytes": 2413,
            "shared_context_bytes": 306,
            "private_value_exposures": 0
          }
        ],
        "native_inference_attempts": 0,
        "real_network_attempts": 0,
        "terminal_private_value_exposures": 0
      },
      "correct": true,
      "whole_operation_seconds": 0.001842499987105839,
      "returned_context_bytes": 693,
      "actual_model_usage": {
        "requests": 0,
        "models": {},
        "input_tokens": 0,
        "output_tokens": 0,
        "usage_complete": true
      },
      "arm": "candidate",
      "repeat": 3
    }
  ]
}