{
  "created_utc": "2026-11-09T14:38:24.151548+00:00",
  "status": "Post-review exploratory analysis of already opened worlds; not confirmation",
  "scientific_question": "Why does joint minus action-value utility shrink with sample size?",
  "authorization": "User authorized additional reviewer-response experiments; no new approval requested.",
  "source_release": "/home/nurgaly/experiment/unionmind/articles/make_1/release/huggingface-revised",
  "code_sha256": "4a300dfc769473f5afb08e15e877f6c53dfff5f6d3a0abfe17e0043798eecbab",
  "device": "cuda:4",
  "input_sha256": {
    "data/controlled/worlds.safetensors": "e8649aeaed970fa78c3f402bbe9e775c486f3cf94ce2494cfe98a8099dbf3c7c",
    "data/controlled/worlds.json": "860e154c4ffd0690080ab8c263756d4340cc119fffa66e1cc41d1b42ac3848cb",
    "data/controlled/manifest.json": "2b50155e7f4d198449e2499dad21ba82d51e94423a1711eca8de4a01ab9ca156",
    "postconfirmation.py": "3365341c056cee9b153cc864bad864965c918e11e0cec1819d49c8875c68dcbf",
    "unionmind/decision_sufficiency/core.py": "fa2f7405a8d0dc6ec96ebae1bb2a2b810d03ef1fd2877f5ebef7df1722992cee",
    "results/postconfirmation/controlled_worlds.csv.gz": "bdb15095eadf96dff2aa316619a838b74024eece87c57143bcd44b4710e9547f",
    "results/postconfirmation/controlled_frontiers.csv.gz": "946ebf8764ee3dfed5a741fa3a45e0cb5f8476960453716d9abebed0aa617e57",
    "results/postconfirmation/controlled_timing.json": "082e8b981d9e6b13a07bfd3136d6afe08681bb8c702ba2c0c2b9a2d4966b6eea",
    "provenance/controlled/confirmation_benchmark.json": "133420e45f42ff5dbacc8d01700aeba56bf65c02f5bc0b30bb7aff45a88bc249"
  },
  "units": "All 192 archived worlds, all six families, all three original nested sample streams",
  "grid": {
    "availability": [
      6,
      3
    ],
    "records": [
      64,
      256,
      1024,
      4096
    ],
    "estimators": [
      "joint_jeffreys",
      "naive_bayes",
      "support_mixture"
    ],
    "methods": [
      "identity",
      "query_joint",
      "posterior_clusters",
      "posterior_bins",
      "future_marginal",
      "raw_information",
      "raw_entropy",
      "hash_code",
      "action_values"
    ],
    "budgets": [
      0,
      1,
      2
    ],
    "costs": [
      0.0,
      0.01,
      0.05
    ]
  },
  "primary_aggregation": "Average positive budgets, all costs and three streams within each world; equal worlds and six families",
  "interventions": [
    "Native: estimated-law encoder, estimated encoded-law exact planner, true-law evaluation",
    "Replan the identical native code under the pushed true law to separate encoding and model-estimation loss",
    "Privileged diagnostic: construct each unchanged encoder under the true law, then plan under either true or fitted pushed law; never a deployable competitor"
  ],
  "exact_decomposition": "J-A = (encoding_loss_A-encoding_loss_J) + (estimation_loss_A-estimation_loss_J)",
  "encoding_loss": "V(available true raw law)-V(true law pushed through native code)",
  "estimation_loss": "V(true law pushed through native code)-realized utility of fitted policy",
  "alias_diagnostic": "Evaluate all unordered raw-state pairs with true current-posterior difference <=1e-12 (exact) and <=0.01 (fixed near-alias threshold), weighted by 2 P(e)P(e'); compare true optimal root choices at each cost and budget and whether each native code merges disagreeing aliases",
  "intervals": "Descriptive paired 95% percentile bootstrap, 10,000 draws, seed 18701, worlds resampled within fixed families after averaging streams; intervals are not corrected and are not new confirmatory tests",
  "practical_scale": "Joint-action gain divided by remaining true available-oracle headroom of action-values; report signed gain, not just a percentage",
  "timing": "Retain published complete-cost records. Fresh warm alternating-order 11-repeat 192-world primary-batch timings include count/estimate, profile, clustering/other encoding, push, exact planning, materialization and 1/32/1024/65536 recorded-event replays. One warmup charged and separately reported. Timing is descriptive on shared hardware.",
  "prohibition": "No tuning, no model refits, no changed codes/iterations/costs, no dropped units, no original results or manuscript edits, no promotion",
  "planned_outputs": [
    "scores.npz",
    "indices.json",
    "frontier_all_conditions.csv",
    "paired_by_condition.csv",
    "sample_frontier.csv",
    "family_sample_frontier.csv",
    "native_worlds.csv.gz",
    "alias_pairs.csv.gz",
    "alias_summary.csv",
    "privileged_geometry.csv",
    "timing.json",
    "summary.json",
    "compute.json"
  ],
  "amendment_01": {
    "created_utc": "2026-11-09T14:43:15.749761+00:00",
    "reason": "The expanded execution trace completed and saved all grid cells, then strict whole-grid native replay validation failed. 946/124416 archived world means differ; primary-mixture joint/action-value/identity at all sample sizes match within 3.34e-16. Preserve all discrepancies and original code/spec/log; aggregate fresh diagnostics without altering source results. Run the released original full grid as a separate trace comparison to investigate tie/numerical sensitivity.",
    "previous_code_sha256": "db7360ee28143a5b9f1e16f373d887354f4824826d8d9b5ed64425e6b732a5dc",
    "completed_grid_code": "collection_code_01.py",
    "completed_grid_specification": "specification_01.json",
    "analysis_changes": "Validation/reporting repair only; no new encoders, thresholds, unit exclusions, tuning, or changed data. Add --resume-aggregation and --reference-only. All native values/oracles in each decomposition come from the same execution trace."
  },
  "amendment_02": {
    "created_utc": "2026-11-09T14:47:08.492706+00:00",
    "reason": "The unmodified released full-grid controlled replay also differs from archive in 968/124416 native world means while exactly reproducing the headline frontier. Test finite-precision argmax sensitivity directly.",
    "operation": "For all grid cells and fixed freshly constructed encoders, compare normal pushforward to accumulation over reversed raw-state order. This is mathematically the same law; measure maximum cell-law perturbation, fitted-optimal expected-value difference, and true-law realized-utility difference.",
    "no_scientific_comparator_change": true,
    "new_outputs": [
      "numeric_push_sensitivity.npz",
      "numeric_push_sensitivity.csv.gz",
      "numeric_diagnostic.json",
      "numeric_compute.json"
    ]
  }
}