{
  "updated": "2026-10-12",
  "sources": [
    {
      "file": "fml-draft-only-interpretation-20261008.json",
      "sha256": "82946ce3600256dec9b12a35abb1ccc97466ce9532d368ddb8ac2964c781b6b2"
    },
    {
      "file": "fml-focused-drafting-interpretation-20261008.json",
      "sha256": "15e64b71b28dd4e0ab0f980cd15e7cd486624fad41ce1b7a565c25eeb252956a"
    },
    {
      "file": "fml-swarm-repair-smoke-acceptance-20261010.json",
      "sha256": "4cd09dcdea59709b073693c3668ccd27f574a68e6b537a7c5fb46f09a6fe0fae"
    }
  ],
  "draft_only": {
    "date": "2026-10-08",
    "attempts_per_run": 8,
    "pairs": [
      {
        "task": "Causality_gcastle",
        "seed": 1,
        "g0_shd": 73,
        "draft_only_shd": 53,
        "draft_only_advantage": 20,
        "state": "observed_pair"
      },
      {
        "task": "Causality_gcastle",
        "seed": 2,
        "g0_shd": 71,
        "draft_only_shd": 64,
        "draft_only_advantage": 7,
        "state": "observed_pair"
      }
    ],
    "cost": {
      "actual_all_in_micro_usd": null,
      "by_arm": {
        "gnosys_g0": {
          "research_seconds": 8132.607,
          "estimated_direct_compute_micro_usd": 1490000,
          "recorded_model_micro_usd": 132041
        },
        "gnosys_draft_only": {
          "research_seconds": 12197.91,
          "estimated_direct_compute_micro_usd": 2180000,
          "recorded_model_micro_usd": 214034
        }
      },
      "compute_estimate_method": "ceil each CodeBuild total elapsed minute at $0.01/min; includes queue/startup and is not invoice reconciliation",
      "excluded": [
        "storage",
        "requests",
        "transfer",
        "90-day retention",
        "shared overhead",
        "prior stages"
      ]
    },
    "limitations": [
      "one development-exposed task",
      "no cross-agent superiority claim",
      "counterbalancing is not randomized provider load",
      "equal limits are not equal realized cost",
      "shared run seeds do not guarantee matched model randomness",
      "no changes after inspecting outcomes"
    ]
  },
  "focused_drafting": {
    "date": "2026-10-08",
    "steps_per_research_run": 8,
    "metric": "structural_hamming_distance",
    "lower_is_better": true,
    "pairs": [
      {
        "seed": 1,
        "gnosys_g0": 47,
        "focused_drafting": 92,
        "focused_advantage": -45
      },
      {
        "seed": 2,
        "gnosys_g0": 81,
        "focused_drafting": 81,
        "focused_advantage": 0
      }
    ],
    "decision": "Do not promote focused drafting; retain G0 unchanged.",
    "limitations": [
      "Two run seeds are not independent datasets or powered prospective confirmation.",
      "The intervention's causal mechanism is unknown.",
      "Earlier four-step AIDE and Autoresearch results are not matched to this eight-step stage.",
      "Do not infer that a longer horizon caused differences versus prior stages.",
      "Protected final scores cannot become a prompt-tuning signal."
    ]
  },
  "swarm_smoke": {
    "date": "2026-10-10",
    "result": {
      "task": "Causality_gcastle",
      "arm": "gnosys_coordinated_swarm",
      "run_seed": 1,
      "metric": "shd_mean",
      "direction": "lower",
      "final_primary_metric": 43.0,
      "search_attempts": 6,
      "observed_search_attempts": 4,
      "failed_search_attempts": 2,
      "reproduction_attempts": 1,
      "duplicate_candidate_attempts": 0,
      "model_calls": 17,
      "accepted_proposals": 8,
      "drafts": 4,
      "improvements": 2,
      "lead_findings": 1,
      "independent_ballot_calls": 2,
      "review_batches": 1,
      "challenger_selections": 1,
      "peak_client_transport_concurrency": 4,
      "concurrency_scope": "Overlapping client transport intervals, not independent models or provider compute concurrency",
      "prompt_tokens": 73219,
      "completion_tokens": 15511
    },
    "limitations": [
      "One development-exposed task and run seed; operational qualification only.",
      "No concurrent G0 or uncoordinated control was run in this smoke; SHD 43.0 is not a relative-performance claim.",
      "Two failed search attempts remain in the denominator and costs.",
      "Smoke output is not pooled into the proposed pilot and is not supplied as search evidence.",
      "No statistical superiority, efficiency, independent-judge or default-promotion claim.",
      "Full objective remains incomplete until a separately approved complete six-run comparison and decision report."
    ]
  },
  "interpretation": "Separate development experiments, not a common leaderboard. No recursive self-improvement or general superiority established."
}
