{
  "aggregates": [
    {
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "comparative_delta": null,
      "condition_id": "no-stele",
      "dispersion": null,
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "interval_high": 1.0,
      "interval_low": 0.20654931437723745,
      "metric_name": "reward",
      "n": 1,
      "value": 1.0
    },
    {
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "comparative_delta": null,
      "condition_id": "faithful",
      "dispersion": null,
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "interval_high": 1.0,
      "interval_low": 0.20654931437723745,
      "metric_name": "reward",
      "n": 1,
      "value": 1.0
    },
    {
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "comparative_delta": null,
      "condition_id": "faithful",
      "dispersion": null,
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "interval_high": null,
      "interval_low": null,
      "metric_name": "investigative_tool_calls_paired_difference",
      "n": 1,
      "value": -1.0
    },
    {
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "comparative_delta": null,
      "condition_id": "faithful",
      "dispersion": null,
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "interval_high": null,
      "interval_low": null,
      "metric_name": "session_investigative_tool_calls_paired_difference",
      "n": 1,
      "value": -1.0
    },
    {
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "comparative_delta": null,
      "condition_id": "faithful",
      "dispersion": null,
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "interval_high": null,
      "interval_low": null,
      "metric_name": "tokens_input_paired_difference",
      "n": 1,
      "value": 83520.0
    },
    {
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "comparative_delta": null,
      "condition_id": "faithful",
      "dispersion": null,
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "interval_high": null,
      "interval_low": null,
      "metric_name": "trial_wall_seconds_paired_difference",
      "n": 1,
      "value": 24.042671999999996
    }
  ],
  "artifact_sha256": "5deb3a6046ad0faff880536d2489368f25015bf2ab406eacb3e1e4a1099fd2e6",
  "attempts": [
    {
      "attempt_id": "package-cache-is-empty-v2-attempt-01-control",
      "attempt_index": 0,
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "condition_id": "no-stele",
      "condition_kind": "no-stele",
      "cost_metrics": {
        "authoritative_receipt": null,
        "calculated_cross_check_cost": "0.584051",
        "calculated_cross_check_currency": "USD"
      },
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "graph_arm": null,
      "memory_evidence": [],
      "outcome": "passed",
      "provenance_hashes": {
        "artifact": "5deb3a6046ad0faff880536d2489368f25015bf2ab406eacb3e1e4a1099fd2e6",
        "challenge_manifest": "00da98b69521e2ac1b8a7f9baad1debf974c0da60c397358520f38c40d6dd3fc",
        "verifier": "704dc9ee5f56a1c19fb8a106a6073f133583e1ae9ee12b5865763727d9556b7e"
      },
      "quality_evidence": [],
      "timings": {
        "active_seconds": 78.638923,
        "execution_seconds": 45.367563,
        "queue_seconds": 0.0
      },
      "token_metrics": {
        "cache_read_tokens": 216544,
        "cache_write_tokens": 31079,
        "input_tokens": 288525,
        "output_tokens": 3081
      }
    },
    {
      "attempt_id": "package-cache-is-empty-v2-attempt-01-treatment",
      "attempt_index": 0,
      "challenge_id": "package-cache-is-empty",
      "cohort_id": "dotstate-faithful-20260723-v1",
      "condition_id": "faithful",
      "condition_kind": "graph",
      "cost_metrics": {
        "authoritative_receipt": null,
        "calculated_cross_check_cost": "0.710177",
        "calculated_cross_check_currency": "USD"
      },
      "experiment_id": "package-cache-is-empty-v2",
      "fixture_id": "dotstate-faithful-20260723-v1",
      "graph_arm": "faithful",
      "memory_evidence": [],
      "outcome": "passed",
      "provenance_hashes": {
        "artifact": "5deb3a6046ad0faff880536d2489368f25015bf2ab406eacb3e1e4a1099fd2e6",
        "challenge_manifest": "00da98b69521e2ac1b8a7f9baad1debf974c0da60c397358520f38c40d6dd3fc",
        "verifier": "704dc9ee5f56a1c19fb8a106a6073f133583e1ae9ee12b5865763727d9556b7e"
      },
      "quality_evidence": [],
      "timings": {
        "active_seconds": 102.681595,
        "execution_seconds": 67.075885,
        "queue_seconds": 0.0
      },
      "token_metrics": {
        "cache_read_tokens": 288772,
        "cache_write_tokens": 48301,
        "input_tokens": 372045,
        "output_tokens": 3562
      }
    }
  ],
  "attempts_analysed": 1,
  "attempts_run": 1,
  "comparability": {
    "challenge_id": "package-cache-is-empty",
    "challenge_manifest_sha256": "00da98b69521e2ac1b8a7f9baad1debf974c0da60c397358520f38c40d6dd3fc",
    "control_condition": "no-stele",
    "fixture_id": "dotstate-faithful-20260723-v1",
    "model": "anthropic/claude-haiku-4.5",
    "source_commit": "a4a5986eafa496dee73a93141588cffaed559f80",
    "treatment_condition": "faithful",
    "verifier_sha256": "704dc9ee5f56a1c19fb8a106a6073f133583e1ae9ee12b5865763727d9556b7e"
  },
  "comparability_digest": "d48398d4cd0776e22f4759898043e159c3f2cb7b340c9927591e8b35e41a4237",
  "memory_uptake": {
    "control": {
      "attempts": 1,
      "injected": 0,
      "marker": 0
    },
    "treatment": {
      "attempts": 1,
      "injected": 0,
      "marker": 0
    }
  },
  "recorded_at": "2026-08-11T11:43:52.436616Z",
  "run_id": "package-cache-is-empty-v2",
  "run_shape": "repeat",
  "schema_version": 1,
  "status": "completed",
  "publication": {
    "decision": "accepted",
    "reason_code": "accepted-preliminary",
    "reason": "The corrected N=1 low-memory pair passed its validity, correctness, mutation, and verifier rules."
  },
  "verdicts": {
    "investigative_tool_calls": {
      "difference": -1.0,
      "fisher_one_sided_p": null
    },
    "reward": {
      "difference": 0.0,
      "fisher_one_sided_p": 1.0
    },
    "session_investigative_tool_calls": {
      "difference": -1.0,
      "fisher_one_sided_p": null
    },
    "tokens_input": {
      "difference": 83520.0,
      "fisher_one_sided_p": null
    },
    "trial_wall_seconds": {
      "difference": 24.042671999999996,
      "fisher_one_sided_p": null
    }
  }
}