[
  {
    "schema_version": 1,
    "run_id": "2026-07-30-v30-validation",
    "model": {
      "model_id": "v30_poly_indgm_scored_yos_avrate_gaussvol_mcmc_deep_i2500_w1000_ad095_td13",
      "source_fit": "data/cache/skill_model_fit_v30_poly_indgm_scored_yos_avrate_gaussvol_mcmc_deep_i2500_w1000_ad095_td13.rds"
    },
    "release": {
      "release_id": "2026-07-30-v30-deep-avrepair",
      "publication_date": "2026-07-30",
      "publication_status": "validated_release_candidate",
      "source_commit": "84c88a3b1a886b61e718fc3db2a5236740b987b6"
    },
    "run": {
      "date": "2026-07-30",
      "type": "post-refit-validation",
      "owner": "Andy + Claude Code session",
      "status": "completed_with_material_ppc_flags_requires_methodological_review",
      "headline_fit_regenerated": true,
      "headline_artifact_bytes_changed": true,
      "public_release_status_changed": false,
      "deployment_performed": false,
      "code_base_head": "84c88a3b1a886b61e718fc3db2a5236740b987b6",
      "working_tree_state": "uncommitted candidate; archive identity becomes immutable only when committed"
    },
    "scope": {
      "posterior_predictive": "raw active-v30 seven-response likelihood, refit 2026-07-30 on AV-repaired and attribution-audited inputs",
      "public_composite": "exact maturity-layer reconstruction and parameter sensitivity against the refit leaderboard",
      "temporal_screen": "five-year estimate stability against the resweep support-route trajectory, not truth calibration",
      "production_route": "R/02_skill_model.R",
      "public_promotion_route": "R/03_composite_and_leaderboard.R",
      "experimental_route": "R/02e_latent_skill_model.R remains comparison-only"
    },
    "reproduction": {
      "commands": [
        "GOODGM_MODEL_CHECK_RUN_ID=2026-07-30-v30-validation GOODGM_PPC_DRAWS=1000 GOODGM_PPC_SEED=42 GOODGM_MODEL_CHECK_GENERATED_AT_UTC=2026-07-30T18:46:37Z Rscript R/04d_production_posterior_predictive_checks.R",
        "GOODGM_MODEL_CHECK_RUN_ID=2026-07-30-v30-validation Rscript R/04e_public_composite_validation.R",
        "node scripts/model-check-archive.mjs --write 2026-07-30-v30-validation",
        "node scripts/model-check-archive.mjs --verify 2026-07-30-v30-validation"
      ],
      "ppc_draws": 1000,
      "ppc_seed": 42,
      "ppc_generated_at_utc": "2026-07-30T18:46:37Z",
      "software": {
        "R": "4.5.0",
        "brms": "2.23.0",
        "posterior": "1.6.1",
        "cmdstanr": "0.9.0",
        "CmdStan": "2.38.0",
        "readr": "2.2.0",
        "dplyr": "1.2.1"
      },
      "entrypoint_sha256": {
        "R/04d_production_posterior_predictive_checks.R": "f2992191be6c698ed33ab9b3cfa09ab1af1aa3d13d1db27bbf6e63589961dc94",
        "R/04e_public_composite_validation.R": "19013e804442a74d711d1d5eae2b845ce183a9fc403ad0777fd180963a26ebc1",
        "scripts/model-check-archive.mjs": "529c74f26dd0640a50573af93e4d162921d5dfd54064eb3991976f0b2cfc5644"
      }
    },
    "compatibility": {
      "previous_ppc": "The prior 2026-07-19-v30-validation PPC checked the same spec on pre-repair, pre-audit inputs; outcome definitions changed (on-team games, honors credit fields), so per-cell flags are directionally but not numerically comparable.",
      "previous_calibration": "The v18 Laplace simulation screen remains not numerically comparable to this temporal stability screen.",
      "trajectory": "R/06 is support-only; its default fixed-effect prior and per-GM applicability handling differ from the deep headline route, and this archive does not add a per-snapshot convergence gate. The trajectory was fully resweeped 2026-07-30 on the same repaired, audited inputs using resolve-once truncated-horizon attribution.",
      "headline": "The deep-v30 fit was regenerated 2026-07-30 on repaired AV and audited attribution inputs; all public headline bytes changed and are re-pinned in the release manifest.",
      "public_contract": "The release still represents 190 unique GMs in 209 GM-team career rows; public intervals remain maturity-adjusted 95% credible intervals."
    },
    "claim_limits": [
      "PPC flags diagnose raw likelihood replication and do not measure maturity-layer calibration or ranking movement.",
      "The temporal screen compares estimates with a later support-route estimate, not with observed true GM skill; its inside-interval rates are not coverage.",
      "Sensitivity results apply only to the tested post-fit maturity settings, not to alternate likelihoods, brms priors, outcome blocks, attribution rules, or experimental models.",
      "The material PPC flags require methodological review; this run documents the refit evidence and does not itself authorize deployment.",
      "The archive is an uncommitted candidate until its files are committed; ordinary verification does not require the ignored source-fit cache."
    ]
  }
]