{
  "pull_request": 119,
  "reviewed_head": "4f2ea071a78da77a84e1de0f5551ec5f71f6ceec",
  "source_sha256": "6f0e1a109ece2679f8ee1e0527cd98abc6a58e664d11fc6033740765ba4a990f",
  "source_bytes": 10883,
  "source_gate_passed": true,
  "review_flags": [],
  "imports": [
    "os",
    "torch",
    "torch.nn.functional",
    "mnist"
  ],
  "static_blockers": [],
  "fresh_fitting": "Module _S caches CUDA graphs, compiled callables and allocated buffers, never a reusable fitted solution. Every classify calls _load(current supplied data), recomputes current same-class tangents, then _reset initializes weights/biases/gains, zeros optimizer moments/step tensors/Muon buffers/pseudo labels/confidences/schedule index and resets EMA/full query pool before 2000 updates.",
  "initial_graph_capture": "The first call in each worker uses current input copies to compile/capture warm-up steps; classify immediately reloads inputs and resets all learned state before its actual training loop. The subsequent timed call overwrites those inputs and resets again. Graph-captured random sampling produces per-step randomness; persistent RNG progression is not retained fitted data.",
  "query_truth": "No test_y argument, scorer memory access or dataset IO. Current graph readout uses only current train_y and EMA predictions. Pseudo-labels and confidence weights derive from current predictions of the opposite model half, refreshed at updates 400, 800, 1200 and 1600. Both halves sample all queries; mutual/self-teaching is explicit and no held-out-vote claim remains in this submission.",
  "input_integrity": "_load copies caller inputs into owned tx/ty/qx tensors. _tangents reads owned tx/ty and creates an owned tangent array. All model/optimizer/EMA/scoring/graph writes act on owned storage; no in-place caller-input operation or output alias exists.",
  "import_gate": "os used only for getuid and sandbox writable compile-cache environment defaults. torch and torch.nn.functional used for training, linear algebra, compilation/CUDA graphs. No filesystem/dataset reads, network, downloaded weights or introspection. __main__ mnist import is guarded and inert in scorer import.",
  "allowed_operations": "From-scratch batched MLPs, tangent jitter from same-class training neighbours, mixup, Muon/AdamW, EMA, label propagation and pseudo-label cross-teaching operate exclusively on supplied arrays. torch.compile and CUDA graph caches amortize compilation but all required learner fitting repeats each call.",
  "input_shape_scope": "Cache assumes canonical fixed TRAIN/TEST/DIMS input shapes; this matches the competition protocol.",
  "source_delta": {
    "upstream118_head": "436b80bb40e4c346792ddad71670c7a5dbc8c5e9",
    "upstream118_sha256": "a3b18bac64dddc3f7aa8c614e2279932aa74040720cb861e8dfa40eaa67923c3",
    "ast_unchanged_functions": [
      "_capture",
      "_cross_teach",
      "_forward",
      "_load",
      "_loss",
      "_orth",
      "_readout",
      "_step",
      "_tangents",
      "_update",
      "classify"
    ],
    "ast_changed_functions": [
      "_build",
      "_reset"
    ],
    "removed_functions": [],
    "added_functions": []
  },
  "attribution": "Source docstring, primary README and manifests credit @likelytobelaura PR #118 pinned at 436b80bb40e4c346792ddad71670c7a5dbc8c5e9, joint graph readout to @sjbaebae, and Muon coefficients to Keller Jordan et al. Upstream frozen source hash matches a3b18bac64dddc3f7aa8c614e2279932aa74040720cb861e8dfa40eaa67923c3.",
  "author_manifests": {
    "primary_source_identity": true,
    "primary_source_bytes": true,
    "comparison_source_identity": true,
    "upstream_source_identity": true
  },
  "author_primary_records": "Independent frozen canonical judge and exact full energy reconstruction finds 3/3 passes; median 2580.8455810546875 ms and 536878.610466496 mJ, pooled error 1.8481818181818181%, with margins 80, 83 and 8 labels. Execution name mlp_d5_tangent_refresh400_2000.py was renamed to packaged submission.py; exact source hash matches.",
  "separate_comparison": "Only the step count differs (2500 instead of 2000), with the same every-400 refresh and full query pool; source hash 86d7caab67e4c9de30a698652fab6195599e5184113c359a83872f5135851018. Independent saved records pass 3/3, median 3385.8267822265625 ms and 674307.0533432962 mJ. No extra reviewer GPU batch was launched for this secondary variant.",
  "failed_candidate_retained": "One-refresh 2000-step source 92e90c1ffcc0e66f4af06a290655b25cbc9dcb0a7305a412f4bdd3ffcc99804b and all three raw records are retained. Independent saved judge finds 2/3 passes; run 2 fails by 60 labels. Passing runs are never pooled into the primary periodic-refresh source; no qualifying median is reported.",
  "limitations": "Author executable-AST/development claims are consistent with frozen sources and supplied arrays but are not independently repeated. Fresh canonical hidden-draw sandbox batch supplies independent qualification; no new secondaryvariant GPU work.",
  "qualification_scope": "Static audit finds no blockers. Fresh initial reviewer batch passed only 1/3: two later foreign warm-ups exceeded the canonical 90-second watchdog after seven timed calls. Those incomplete records cannot receive canonical accuracy/timing qualification; leave the PR unmerged. No retries or source repairs.",
  "canonical_process_scope": "Each of the 15 canonical timed calls gets a fresh sandbox worker. That worker imports the source, fits once on an untimed foreign warm-up draw, then fits on the timed current draw and is killed. Module graph/buffer caches therefore persist only within a worker between warm-up and timed fitting, and reset/load correctness prevents warm-up learned state or data from carrying into the timed fit."
}
