{
  "pull_request": 115,
  "reviewed_head": "40d1a646b4da0f5473020d985faa5b49c253dfa5",
  "source_sha256": "86a1a749f5b12c34be86c2e5025487827bda6b8b20f55fac72bb2911dcaa9066",
  "source_bytes": 20428,
  "archive_sha256": "582fe8613daf6be11175e93db7d11161475dd1a5345a4a857feb34800b017860",
  "source_archive_files": 30,
  "reviewed_at": "2026-10-12T14:33:56.261194+00:00",
  "static_blockers": [],
  "source_gate_passes": true,
  "checks": {
    "archive_source_only": true,
    "archive_paths_safe": true,
    "all_embedded_files_match_readable_PR_sources_and_manifest": true,
    "no_fitted_weights_data_or_query_truth": true,
    "fresh_per_call_training": true,
    "input_integrity_by_source_review": true,
    "same_training_schedule_for_MNIST_foreign_warmups_and_holdout": true,
    "no_scorer_introspection_or_protocol_tampering": true
  },
  "findings": [
    {
      "topic": "archive",
      "evidence": "Wrapper Base85/LZMA literal decodes to exactly30 readable Python/CUDA/header/binding sources. Exact decoded JSON archive SHA and every file hash match source_manifest.json and readable source/. Only source files; no serialized weights, datasets, binaries, offline fitted arrays or scorer state."
    },
    {
      "topic": "per-call training",
      "evidence": "candidates/mlp_bn_head16_d5_depthensemble_raw.py:11-16 runs two teacher fits (1000/250) and two student groups (3000 each), reading each current x/y/q. All active native launch_pb kernels first convert current x/q, initialize every weight/head with current data/seed, zero all momentum and SWA accumulators, then perform the complete update loop. Native binding caches retain code, CUDA graphs, workspaces and library plans; current caller data is recopied before every graph replay, so neither warmup nor previous energy call leaves fitted state for reuse."
    },
    {
      "topic": "labels and query use",
      "evidence": "Teacher K5 voting and graph clamping use only supplied y; teacher high-confidence predictions for current q become pseudo labels. No query truth is passed or accessed. Student mean fitting is over the first len(x) true training rows. Final K5 and graph prior again use only original supplied y, including the secret label permutation. Transductive embeddings/BN and query graph use current unlabeled query features only."
    },
    {
      "topic": "input integrity",
      "evidence": "Python builds torch.cat outputs and fresh indexed/concatenated x/y/q values. Native API captures into own tensors via device-to-device copies; launch_pb uses const float*/const long* caller input pointers and writes converted features, parameters, embeddings and predictions only to its own allocations. No in-place writes to the caller x/y/q were found. This is a source audit, not an extra live learner trial."
    },
    {
      "topic": "timing protocol",
      "evidence": "All computation occurs in classify after gate-approved imports. No timer/scorer monkeypatching, truth/file access, network use, process introspection, dataset recognition shortcuts or output reuse found. CUDA graph library autotuning and compilation are permitted warmup behavior. All three reviewer runs must independently complete the canonical sandbox protocol."
    },
    {
      "topic": "stale descriptive text",
      "evidence": "Wrapper docstring says not qualified and source_manifest scope says source-only prototype; README and raw author records supply later author qualification. Stale descriptive packaging strings are nonfunctional; no static blocker."
    }
  ],
  "qualification_status": "Completed fresh reviewer set: 2/3 passes. Run 2 fails D5 accuracy by 37 labels; leave unmerged. No qualifying median and no retries."
}
