verified_configs.json

{
  "schema": "verified_configs/v1",
  "policy": "append-only. Every row carries N, pass rate WITH 95% Wilson CI, and a run hash. One verified row beats six plausible ones. A row is a claim the instrument can defend; never seed from memory.",
  "configs": [
    {
      "model": "google/gemma-4-31b",
      "quant": "GGUF (LM Studio)",
      "channel": "local-api",
      "battery": "64 distinct tests (DB-verified: 64 distinct test_ids, 192 trials, N=3 each). ONE display name — 'AUX-APPROVAL-01 Benign Command Classification' — is shared by two different tests, so name-keyed exports show 63 unique names. Count what the primary key counts.",
      "n": 192,
      "passed": 189,
      "pass_rate": 0.9844,
      "wilson95": [
        0.9551,
        0.9947
      ],
      "residual_failures": "LIT-12 x3 (sound control, stochastic over-call \u2014 model behavior, not instrument)",
      "date": "2026-07-26",
      "run_hash": "sha3-512:39cdd76dec386a4 (run 953)",
      "notes": "Grader bug #3 fixed (ac37cad+6ca610a): verdict extraction + normalization. First run under the fixed grader. Claude Science pre-hoc prediction (LIT-12-only residuals) matched exactly.",
      "verify": "docker exec archetype-postgres psql -U careybalboa -d calibration_scope -c \"SELECT COUNT(*) AS trials, COUNT(DISTINCT tr.test_id) AS distinct_tests, COUNT(DISTINCT te.name) AS distinct_names, COUNT(*) FILTER (WHERE tr.passed) AS passed, COUNT(*) FILTER (WHERE tr.is_infra_error) AS infra FROM trial_results tr JOIN tests te ON te.id=tr.test_id WHERE tr.run_id=953\" -- expect 192 | 64 | 63 | 189 | 0"
    }
  ]
}