{
  "schema": "glee.public_research_edition.v1",
  "edition_date": "2026-10-12",
  "producer": "GLEE engineering team / AI-assisted public edition (Codex)",
  "boundary": "Public summaries of a bounded engineering pilot and dated engineering cases. No general efficacy, independent replication, scientific validation or current runtime-health claim.",
  "entries": [
    {
      "id": "agent-steering-20261012",
      "question": "Can guidance help an AI repair code?",
      "kind": "engineering pilot",
      "observation_date": "2026-10-12",
      "outcome": "inconclusive",
      "summary": "A matched comparison produced no accepted repairs in either arm. Most runs stopped before candidate evaluation, so it cannot establish whether evidence guidance improves repair quality.",
      "method": {
        "tasks": 8,
        "repositories": 1,
        "matched_pairs": 16,
        "runs": 32,
        "decisions_per_run": 4,
        "arms": [
          {
            "name": "Self review",
            "runs": 16,
            "accepted": 0,
            "candidate_evaluations": 0
          },
          {
            "name": "Evidence guidance",
            "runs": 16,
            "accepted": 0,
            "candidate_evaluations": 1
          }
        ]
      },
      "observations": {
        "response_schema_failed": 8,
        "source_hash_protocol_exhausted": 23,
        "candidate_tests_failed": 1,
        "accepted_repairs": 0
      },
      "limits": [
        "Thirty-one runs never reached candidate evaluation; the one evaluated candidate failed six assertions.",
        "The apparatus intentionally exposed an older source hash while a current read remained available. This recovery challenge dominated the pilot.",
        "The result does not establish general model ability, steering efficacy, false-acceptance specificity, or economic advantage. Full cost remains unknown.",
        "This is an authored public summary of an internal frozen trial and continuation audit. Private trajectories and test artifacts are not downloadable here; independent reproduction is not established."
      ],
      "provenance": {
        "source_study_producer": "GLEE engineering team",
        "continuation_audit_producer": "GLEE engineering team / AI-assisted audit (Codex)",
        "source_report_title": "Experiment003 continuation audit",
        "source_report_sha256": "56092df3baceef0b95c6bcf02d59bd56acb21e502e80fcdf7464f7ec67967407",
        "audit_observed_at": "2026-10-12T20:17:04.074688+00:00"
      }
    },
    {
      "id": "status-label-vs-capability",
      "kind": "dated engineering case",
      "observation_date": "2026-07-24",
      "question": "Does a status label tell us what a system can actually do?",
      "claim": "A health field answered a different question than the one callers were asking, and the label was accurate the whole time.",
      "evidence": [
        "Local inference pool unloads its backend after 300s idle and reports status down with no resident model.",
        "It swaps a model back in on the next request: a live call returned content with swap_occurred true and 501ms swap latency while health still read down.",
        "A fuel check gated on status == healthy, so five minutes of quiet flipped the whole estate to no-local-reasoning.",
        "Downstream jobs fell through to a cloud path with no key and skipped. Skips were recorded as ordinary outcomes, so nothing alarmed for 34 hours."
      ],
      "decision": "Gate on capability rather than the label: reachable, swap-capable, models registered, queue not saturated. Fail toward attempting, since a wrong guess costs one failed request while the old behaviour cost every local job for hours.",
      "limits": "Down remains correct as a description of residency. The fix is on our side of the contract, not theirs. Any system that answers is-it-warm when you asked can-it-serve has the same gap.",
      "source_url": "/data/field-notes/GLEE-FIELD-NOTES-v0.json",
      "source_sha256": "27fb8df506f4a3e566fbfc05303bece99875476d1bef5521cbcff38243d73bce"
    },
    {
      "id": "guard-inversion",
      "kind": "dated engineering case",
      "observation_date": "2026-07-24",
      "question": "Can a safeguard make a wrong answer look more convincing?",
      "claim": "A constraint a weak generator can satisfy without being correct will be satisfied without being correct.",
      "evidence": [
        "To stop a model altering our figures we required every grounded number to appear verbatim in generated copy.",
        "The model stopped altering numbers and began inventing the sentences around them: a fabricated version, a fabricated user count, a fabricated success rate, all wrapped around genuine figures.",
        "Before the guard, ungrounded output showed as wrong numbers, which are trivially detectable. After it, ungrounded output showed as correct numbers in false context, which passes automated checks and reads as more credible."
      ],
      "decision": "Add a check for figures appearing with no claim attached, and keep generated drafts human-reviewed rather than auto-published.",
      "limits": "This catches recitation, not falsehood. A confident lie in clean prose still passes. Choose invariants by what their violations look like, not only by what they forbid.",
      "source_url": "/data/field-notes/GLEE-FIELD-NOTES-v0.json",
      "source_sha256": "27fb8df506f4a3e566fbfc05303bece99875476d1bef5521cbcff38243d73bce"
    }
  ]
}