{
  "pull_request": 137,
  "reviewed_head": "2ccf9c9a0924b51b0c7b262b4856f1c29490b643",
  "frozen_main": "b374daa1a59cf0e3ce9a6462ecdf2f143ac5b5fc",
  "source_sha256": "eae98d2337442db3c2ed5984e455cf2fc945bfa6846aeaa93d0decb51799d232",
  "source_bytes": 20453,
  "function": "kc_d3",
  "source_gate_passed": true,
  "review_flags": [
    "20,453 bytes, close to the 20,480-byte limit"
  ],
  "source_limit_bytes": 20480,
  "source_byte_margin": 27,
  "static_blockers": [],
  "flag_disposition": "Source literal content is scalar algorithm constants and readable documentation only; no encoded/fitted table or external/process access. Entire source reviewed.",
  "full_source_read": true,
  "implementation": "Plain PyTorch/cuBLAS/cuSOLVER/ATen flash attention; no custom Triton/native kernel.",
  "current_input_fitting": "Fresh labels/counts, class moments, LDA/QDA Fisher metric, compression/landmarks/weights, FITC-Woodbury preconditioner, zero-origin PCG recurrence and current scores recomputed inside captured run; all caller inputs copied into owned static clones each call.",
  "cache_ownership": "_GRAPHS retains graph/static inputs/output; _C retains only scalar/identity/zero/one/index tensors and _ST retains streams. Captured intermediates use retained graph-private pools; outputs cloned.",
  "work": "Current full fit executes at every replay. Side streams explicitly branch/join before dependent operations. Fixed metric/compression/PCG work; first-call compilation/capture/64 replays stay in foreign warmup.",
  "truth": "Only supplied train_y gives truth. No query truth, files/network, caller/parent/scorer/timer introspection, offline prior or learned warmup state reused.",
  "published_scope": "Official three-run median; independent reviewer reproduction supplies evidence, not a permanent source-ban or mandatory additional published batch. Optional energy assessed separately from score.",
  "blockers": [],
  "fresh_reviewer_outcome": {
    "passed_runs": 3,
    "required_passed_runs": 3,
    "energy_valid_runs": 3,
    "score_qualification_passed": true,
    "optional_energy_valid": true,
    "median_ranked_ms": 3.5107840299606323,
    "median_energy_mj": 358.758433589969,
    "aggregate_mnist_error_pct": 2.523030303030303,
    "accuracy_margins_labels": [
      157,
      210,
      217
    ],
    "all_45_timed_calls_complete": true,
    "distinct_energy_gpu_uuids": 3,
    "reruns": 0,
    "learner_repairs": 0,
    "failed_outcomes_replaced": 0,
    "blockers": [],
    "decision": "merge exact submitted source; promote reviewer medians if faster than current record"
  },
  "transduction": "Confident current query predictions may serve as HKNN vertices; low-margin rows are corrected using current top classes, not query truth.",
  "bounds": "HKNN slots are bounded by per-class CAP=2048 with an appended zero padding vertex; cumsum selects at most512 low rows into a513-entry scratch. Unused entries repeat row0; actual selected rows unique. Lower triangular49+1 Gram/block matrix fully written before lower Cholesky; upper-right unused.",
  "source_review_status": "Full source/rules/stream ownership, reset, bounds, math, author arithmetic and upstream provenance reviewed; no static blocker.",
  "checked_at_utc": "2026-10-22T18:46:59.229980+00:00",
  "linear_algebra": "Standard-form Woodbury FITC and LDA low-rank inverse identities independently checked in8CPU FP64 controls (max relativeerror~2.2e-16). Low-precision products are numerical changes and actual predictions remain subject to canonical qualification. See linear-algebra-audit.json.",
  "precision": "TF32 and fp16 reduced-precision reductions disabled. FP16 GEMMs accumulate FP32 where specified; FP64 used only for rank-subspace small matrix work inD3. No claimed bitwise equivalence.",
  "current_input_reset_detail": "Every class count/one-hot/moment and metric is newly derived from current train_x/train_y. Current seeds, centroids, hard-row support, weights, Gram/preconditioner and every PCG D/G slot are written before use. No persistent optimizer exists. Current test_x is copied and featurized each call. _C constants and stream/graph addresses persist, fitted values are refreshed.",
  "stream_detail": "Moments/class seeds fork after current normalized inputs and labels are ready; main joins before Fisher. LandmarkK and Woodburyfactor work on same retained side stream, which waits for each main-stream input; main joins before PCG. D3 natural-gradient side chain similarly joins before transforming features. No work escapes event/clock timing.",
  "ownership_detail": "External static clones, cached input-independent constants and streams have explicit global roots. _GRAPHS tuple retains CUDAgraph and output; captured intermediates and scratch remain in graph-private pool. There is no unrooted pre-capture Students-style owner or hidden compiler-cache dependency. gc.freeze is not relied upon to retain external fitting state.",
  "author_evidence": "All3 provided exact-source author rawrecords independently reconstruct score and optionalenergy, with three distinctUUIDs and all manifest raw hashes matching. Working filename is documented. Unsupported development counts/probabilities/timing and earlier-group histories remain separate and unverified. See author-evidence-audit.json.",
  "upstream_evidence": "Source/README retain Haden parent learner, @sigkillme0 #125 Fisher/compression/PCG and @islamborghini kernel-PCG credits; D3 also Haden HKNN/Vincent-Bengio. AST comparisons are in upstream-attribution-audit.json. Full-learner comparison is not an isolated-kernel causal claim.",
  "D6_scope": "No D6/scorer/high-accuracy learner or exact-source D6 evidence supplied in either speed PR.",
  "HKNN_bounds_and_compaction": "Each accepted low row gets a unique cumsum index within0..511; all rejected/excess rows write the discarded513th scratch entry. Unused512slots remain row0, so graph always solves512x3 hulls. CAP2048 tables use appended paddingvertex; legitimate vertices are current train rows or confident predictions. Spare row0 writes are bounded and operate on repeated current-query data; no truth/state bypass. See compaction-bounds-audit.json.",
  "HKNN_readout": "For each current49D query/class, G=VTV+3I. Lower augmented50x50 block has G,rT and c=1+||r||\u00b2/3 fully written; c-lastpivot\u00b2=rT G^-1 r. CPU FP64 identity check max relativeerror3.3e-15. The unused upper-right column is outside lower factorization input; no uninitialized referenced triangle.",
  "lower_triangle_primary_source": {
    "url": "https://raw.githubusercontent.com/pytorch/pytorch/v2.12.0/aten/src/ATen/native/BatchLinearAlgebra.cpp",
    "local_sha256": "85d3aaf5e9ab497105744e797f65052fcad6db0f98b01aa7f3febac61f84d4f5",
    "lines": "1820-1866",
    "supports": "upper=False dispatches lower Cholesky; code explicitly identifies opposite triangle as unreferenced and clears it after GPU factorization."
  },
  "rank_and_natural_gradient": "Rank48 factor built from current metric and first48 current train rows, with current-label QDA Fisher/natural-gradient terms. No hardcoded coordinate/class mapping. Returned labels are cloned to owned caller-visible output.",
  "review_status": "Complete: static/provenance audit and first canonical reviewer group finalized.",
  "finalization_complete": true
}
