Add WinML reranking recipes for cross-encoder/ms-marco-MiniLM-L4-v2

[ssss141414]

Summary

Adds two CPU reranking recipes for cross-encoder/ms-marco-MiniLM-L4-v2, a compact English passage cross-encoder. This is an L0 recipe-only contribution with full fp32 and true-fp16 CPU coverage. The committed Goal ceiling was L3, and the fresh immutable recovery run reached L3 PASS without any candidate change.

Model metadata

What the model does

An English cross-encoder passage reranker that jointly encodes one query-document pair and emits one raw scalar relevance logit; candidates are ordered by decreasing score without softmax, sigmoid, or instruction prefixes. Confidence: verified.

Evidence: the model card and config.json at pinned checkpoint revision 777b2f369bc1c2f850df8bd367ed1654bda4497b describe MS MARCO passage reranking, decreasing-score sorting, BertForSequenceClassification, one LABEL_0 output, and Identity activation. Current-main export resolves logits from the sequence-classification head.

Primary user stories

• A user supplies a search query and a bounded candidate passage list to obtain one relevance score per pair and reorder retrieval results. Evidence: pinned model-card usage and retrieve-and-rerank description. Confidence: verified.
• A retrieval or RAG pipeline supplies first-stage candidates to obtain a stronger final ordering before presenting or using the evidence. Evidence: the pinned model card links the checkpoint to information retrieval and retrieve-and-rerank usage. Confidence: mapped.

Supported tasks

• reranking (checkpoint, transformers): the published pipeline tag is text-ranking, with pinned pair-scoring examples; AutoModelForSequenceClassification is the published Transformers class. Confidence: verified.
• text-classification (transformers, optimum-onnx, winml): current-main inspect and recipe-free build resolve through BertIOConfig and AutoModelForSequenceClassification; the pinned checkpoint exports successfully at opset 17. Confidence: verified.

Model architecture

BertForSequenceClassification
|-- BertEmbeddings (vocab 30,522 -> hidden 384; positions 512)
|-- BertEncoder layer x 4
|   |-- Bidirectional self-attention (12 heads)
|   |-- Feed-forward (384 -> 1,536 -> 384, GELU)
|   `-- Residual + LayerNorm
|-- BertPooler (first-token hidden state)
`-- Sequence-classification head (384 -> 1 raw relevance logit)

• Source/confidence: pinned checkpoint dimensions and BertForSequenceClassification, corroborated by current-main hierarchy metadata for 19,164,673 parameters, 87 modules, 33 traced modules, 66 execution steps, four BertLayer children, and 215/215 tagged ONNX nodes (verified).

Validation and support evidence

1. Baseline

The fresh baseline used current main commit 774dd6e0811f6947eaf2bd8fb721d5f6e718b0c0 and WinML 0.0.1.dev0. Planner classified the move from 0876e5ae1c98a169a6137e092e0d7b30bf9cee33 as FULL-RERUN, not a SHA-only attestation: the intervening changes directly affect shared Perf EP/device selection and monitoring, session exports, Analyze missing-schema behavior, CLI constants, and their quality gates. Planner therefore reran inspect, default and reranking config, build, perf, Eval, and Analyze fail-closed on current main; only the unchanged model profile and dependency-alias evidence were reused.

Recipe-free BertForSequenceClassification / text-classification build passed at IR 8/opset 17 in 11.0 s, producing a 76,714,266-byte model with three int32 [1,512] inputs, float [1,1] logits, and FLOAT initializers. The CLI still emitted an Optimize stage despite --no-optimize. Baseline CPU perf over 10 measured iterations after 2 warmups was 37.30 ms mean, 37.94 ms p50, 43.70 ms p95, 26.81 samples/s, and +79.2 MB inference RAM. Baseline reranking Eval remained at the unsupported floor: it exited 1 and emitted no metric. Default config resolved text-classification; explicit reranking config exited 2 as unsupported. Baseline Analyze returned the expected partial-success contract with exit 1 and retained JSON containing 152 operators, 17 operator types, and 12 provider rows; this is static compatibility evidence, not a command crash or non-CPU runtime proof.

Optimum probing was VENDOR-ONLY: the exact L4 import probe failed because installed Optimum imports a removed Transformers 5.14.1 private symbol. The accepted exact-runtime BERT probe exposed the same six vendor and post-WinML tasks (feature-extraction, fill-mask, multiple-choice, question-answering, text-classification, token-classification) with added_by_winml=[]; fresh L4 inspect/build independently verified the same model class and exporter path.

This PR is intentionally stacked on draft dependency PR #1322, branch producer/ms-marco-minilm-l6-v2-reranking, at 3708969b731425b0c6d4b97920d1b5e6519bb013. On that dependency, text-ranking and reranking config both exit 0, normalize to loader.task=reranking, and are parsed-JSON identical. Current main (774dd6e0811f6947eaf2bd8fb721d5f6e718b0c0) and the dependency head are divergent: current main is used only for the refreshed baseline, while candidate d951fccaba37275540cdf2757bfb6b3cf60e8ec8 has the dependency SHA as its exact sole parent. The candidate therefore does not incorporate current-main commits; it contributes exactly two recipe files on top of #1322 and depends on #1322's generic reranking normalization and Eval support.

2. Goal

• Effort: L0, recipe-only CPU precision coverage.
• Goal ceiling: L3.
• Outcome: L0 shipped with full required coverage.
• Success definition: march L0-L3 with complete CPU fp32/fp16 perf, parity, functional smoke, analysis, and CI-parity gates.

No ceiling change or re-issued charter occurred.

3. Outcome

The shipped tier is L0, the highest reached Goal verdict is L3 PASS, coverage is full, and there are no deferred tuples. The contribution contains only:

• examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/cpu/cpu/reranking_fp32_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/cpu/cpu/reranking_fp16_config.json

Quality gates passed: license-header check; Ruff; mypy over 444 source files; 1,526 analyze tests; 1,534 model/loader/dataset/export tests; 848 optimization tests; 3,641 command/config/build/compiler/session/eval tests; and 892 remaining unit/regression/CLI tests, with only the recorded skips, xfails, deselection, and warnings.

Learner v3 preserves model findings bert-024 through bert-029, which capture exact L4 build structure, HTP hierarchy, component/op analysis, CPU performance, raw-logit parity, and the bounded SciDocs smoke. It refreshes bert-024 and bert-027 from the Planner v2 baseline and Tester v2 fp16 evidence. Methodology friction _meta-114 records that precision-size validation must compare realized model payloads rather than aggregate build directories. The focused Lane A compare was pushed and verified at final commit e125b318d74a1319cedf5215b92f6bdd0ead9140; it preserves the separate authoritative _meta-113 finding.

Final PASS retains the exact-candidate evidence from immutable recovery run 32695311485 and refreshes only fp16 L1 in immutable hosted run 32701875911, as authorized by the Tester v2 preserved-evidence manifest. Preserved L0, fp32 L1, L2, L3, Analyze, and quality evidence remains tied to the unchanged candidate and dependency. The preserved first run failed because the tester harness compared aggregate build-directory sizes containing retained fp32 intermediates (230,182,646 bytes fp32 vs 230,235,020 bytes fp16), rather than realized model.onnx payloads (76,710,324 bytes fp32 vs 38,381,076 bytes fp16). This was a tester-harness defect; the correction changed the size assertion to compare final payload bytes, and the candidate commit did not change.

The fp16 refresh also preserves two failed hosted harness attempts truthfully: run 32701127519 referenced a nonexistent top-level materialized-config model key, and run 32701514334 found that raw CLI Perf passed but did not invoke its disconnected Model Precision display helper. The final harness printed initializer-derived Model Precision: fp16 before invoking the same bounded CLI Perf command. These were evidence-harness incidents, not candidate failures; the final artifact was byte-identical and the candidate remained d951fccaba37275540cdf2757bfb6b3cf60e8ec8.

4. Per-EP/device/precision results and Functional smoke Eval

Goal ladder

Tier	Verdict	Evidence
L0	PASS	Both required CPU precision tuples built and passed structure/precision checks.
L1	PASS	Fresh CPU perf completed for both required tuples.
L2	PASS	Raw scalar-logit parity and descending order passed on three pinned query-passage pairs.
L3	PASS	Final-candidate fp32 CPU bounded SciDocs reranking Eval completed.

Build, perf, and parity

EP / Device	Precision	Verdict	Model / initializer proof	Mean	p50	p95	Throughput	RAM before	RAM inference / total delta
CPUExecutionProvider / cpu	fp32	PASS	76,710,324 bytes; 76 FLOAT, 1 BOOL, 24 INT64 initializers	66.65 ms	62.37 ms	92.52 ms	15.0 samples/s	674.7 MB	+78.2 / +78.2 MB
CPUExecutionProvider / cpu	fp16	PASS	Model Precision: fp16; 38,381,076 bytes; 76 FLOAT16, 1 BOOL, 24 INT64 initializers	97.02 ms	96.45 ms	120.96 ms	10.31 samples/s	699.9 MB	+87.9 / +87.9 MB

Both artifacts expose input_ids, attention_mask, and token_type_ids as INT32 [1,512], with logits as FLOAT [1,1]. FP32 also measured p90/p99/max 92.52 ms and min 62.03 ms. Fresh fp16 measured p90/p99/max 120.96 ms and min 85.85 ms, with explicit initializer-derived runtime output Model Precision: fp16. Its exact realized payload is 38,381,076 bytes, including 38,329,352 raw FLOAT16 payload bytes across 76 FLOAT16 initializers. VRAM delta was not applicable for CPU.

Precision	PyTorch logits	ONNX logits	Cosine	Max abs	Required floor / ceiling	Descending order	Verdict
fp32	[8.320240020751953, 1.2873587608337402, -11.603301048278809]	[8.320236206054688, 1.2873551845550537, -11.603301048278809]	0.9999999999999492	0.000003814697265625	>=0.99999 / <=0.001	[0,1,2] preserved	PASS
fp16	[8.320240020751953, 1.2873587608337402, -11.603301048278809]	[8.31916618347168, 1.2839525938034058, -11.603437423706055]	0.9999999705508874	0.0034061670303344727	>=0.999 / <=0.1	[0,1,2] preserved	PASS

Functional smoke Eval

PASS, operability evidence only; not representative accuracy or benchmark quality. The final candidate fp32 CPU artifact ran mteb/scidocs-reranking test split at pinned revision 56a6d0140cf6356659e2a7c1413286a774468d44. Deterministic first-N streaming selected and processed 2 of 2 query groups with no shuffle. Schema, positive/negative label semantics, and one-raw-logit-per-pair descending-score prediction semantics were verified.

Caps were 2 groups, 10 candidates per group, at least 1 positive per group, 20 pairs total, and sequence length 512; candidate labels/prompts were capped at 10, with beams and frames/crops not applicable. Raw metrics were MRR@10 1.0, Recall@1 1.0, and Recall@10 1.0. No other precision or EP received an accuracy claim. Current main formerly rejected the reranking task; dependency #1322 supplies generic task normalization, paired-input raw-scalar evaluation, grouped-text adaptation, candidate caps, and ranking metrics.

5. Delta

The production recipe README is untouched. No product source, generic tests, skill files, or existing L2/L6/L12 recipes changed in this model contribution.

Recipe	JSON pointer	Baseline	Shipped	Classification
fp32	/loader/task	text-classification	reranking	Consumes generic normalization owned by dependency #1322.
fp16	/loader/task	text-classification	reranking	Consumes generic normalization owned by dependency #1322.
fp16	/quant	null	{"mode":"fp16","fp16_keep_io_types":true,"samples":10}	Required checkpoint CPU fp16 coverage; remaining fields are accepted explicit schema defaults.
both	/export/compatibility	{"transformers_attention":"eager"}	omitted	Sparse-template omission; materialized configs restore transformers_attention=eager.
both	/optim/clamp_constant_values	true	omitted	Sparse-template omission; optimization is auto-resolved.

The dependency-generated text-ranking and reranking configs are parsed-JSON identical. The fp32 and fp16 recipes have the same sparse structure except for the required fp16 /quant declaration. Reducibility is consistent with the charter: only per-checkpoint CPU precision coverage is authored; all metadata-derived reranking normalization and evaluation remains in dependency #1322. Recipe-free acceptance is not required for this per-model coverage contribution, and no checkpoint-hardcoded code exists.

6. Analyze summary - component level and op level

Static analysis returned ANALYZE-PARTIAL-SUCCESS for both artifacts: analyze exited 1 because provider-bridge initialization failed, while retaining complete JSON for all seven requested EPs. This is static rule analysis, not runtime execution proof for non-CPU EPs.

Component-level summary

Artifact	Architecture coverage	Mapping	Confidence	Actionable finding
fp32	embeddings; 4x encoder attention/FFN; pooler; scalar classifier	89 mapped, 47 unmapped	mapped	QNN partial Gather in embeddings/pooler; no unsupported types.
fp16	embeddings; 4x encoder attention/FFN; pooler; scalar classifier	89 mapped, 48 unmapped	mapped	QNN partial Gather in embeddings/pooler; no unsupported types.

The explicit unmapped nodes are backbone mask/shape glue outside the frozen component regions.

Op-level summary

Artifact	Graph	Dominant operators
fp32	136 operators / 16 types	Reshape 41; Gemm 26; Transpose 16; Add 14; LayerNormalization 9; MatMul 8; Cast 1
fp16	137 operators / 16 types	Reshape 41; Gemm 26; Transpose 16; Add 14; LayerNormalization 9; MatMul 8; Cast 2

The other counts are identical in both graphs: Gather 4, Mul 4, Flatten 1, And 1, Expand 1, Where 1, Softmax 4, Gelu 4, and Tanh 1.

Static EP row	fp32 semantics	fp16 semantics
NvTensorRTRTXExecutionProvider	15 types supported; Where unknown; none partial/unsupported	15 types listed supported; Where and Cast unknown; none partial/unsupported
CUDAExecutionProvider	No classifications emitted	No classifications emitted
MIGraphXExecutionProvider	No classifications emitted	No classifications emitted
QNNExecutionProvider	15 types listed supported; Gather partial; Where unknown; none unsupported	15 types listed supported; Gather partial; Where and Cast unknown; none unsupported
OpenVINOExecutionProvider	15 types supported; Where unknown; none partial/unsupported	15 types listed supported; Where and Cast unknown; none partial/unsupported
TensorrtExecutionProvider	No classifications emitted	No classifications emitted
DmlExecutionProvider	No classifications emitted	No classifications emitted

7. Reproduce commands

$OUT='temp/model-support-repro'
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/cpu/cpu/reranking_fp32_config.json -m cross-encoder/ms-marco-MiniLM-L4-v2 -o $OUT/fp32
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/cpu/cpu/reranking_fp16_config.json -m cross-encoder/ms-marco-MiniLM-L4-v2 -o $OUT/fp16 --precision fp16
winml perf -m $OUT/fp32/model.onnx --device cpu --ep cpu --warmup 2 --iterations 10
winml perf -m $OUT/fp16/model.onnx --device cpu --ep cpu --warmup 2 --iterations 10
winml analyze --model $OUT/fp32/model.onnx --ep all --output $OUT/analyze-fp32.json
winml eval -m $OUT/fp32/model.onnx --model-id cross-encoder/ms-marco-MiniLM-L4-v2 --task reranking --dataset mteb/scidocs-reranking --dataset-revision 56a6d0140cf6356659e2a7c1413286a774468d44 --split test --streaming --no-shuffle --samples 2 --column query_column=query --column positive_column=positive --column negative_column=negative --column max_candidates=10 --ep cpu --device cpu -o $OUT/eval.json --overwrite

[ssss141414]

Independent reviewer verdict: REQUEST_CHANGES

Reviewed immutable PR head d951fccaba37275540cdf2757bfb6b3cf60e8ec8, whose parent and live stacked base are both 3708969b731425b0c6d4b97920d1b5e6519bb013 (#1322). I did not participate in producer-side work and did not mutate the PR, branch, metadata, labels, draft state, or review threads.

Actionable blockers

1. Planner / Explainer: refresh the baseline and public baseline claim against current main. The charter and PR body call 0876e5ae1c98a169a6137e092e0d7b30bf9cee33 current main, but the independently fetched origin/main is now 774dd6e0811f6947eaf2bd8fb721d5f6e718b0c0. The complete intervening delta is not recipe-only: it changes src/winml/modelkit/commands/perf.py, src/winml/modelkit/session/*, src/winml/modelkit/analyze/core/runtime_checker_query.py, src/winml/modelkit/utils/constants.py, and related tests (18 files, 815 insertions/57 deletions). Those paths directly intersect the charter's frozen perf/session/analyze/runtime dependency manifest, so this cannot be accepted as an unattested SHA move. Planner must issue a current baseline refresh or a complete moved-main impact attestation with honest rerun provenance; affected baseline stages must be rerun, and Explainer must update the source-owned PR body. Evidence: reviewer artifact baseline-to-current-main.name-status.txt / .stat.txt.

2. Tester: provide explicit final-artifact fp16 runtime precision evidence in a fresh immutable acceptance root. The exact fp16 build correctly used --precision fp16 without --no-quant; independent artifact inspection confirms 76 FLOAT16 initializers and a 38,381,076-byte realized payload versus fp32's 76 FLOAT initializers and 76,710,324 bytes. However, neither sealed fp16 perf stream contains the current reviewer-contract signal Model Precision: fp16. The stderr model panel reports only float32 output I/O, which is expected with fp16_keep_io_types=true but does not establish weight precision at runtime. Exact-head perf.py has initializer-derived precision support and a Model Precision reporting path, so Tester must capture an explicit precision result for the exact final artifact, rerun the affected fp16 acceptance evidence in a new empty root, regenerate semantic validation/seal, and independently rehash it. Evidence: sealed logs/009-perf-fp16.stdout.txt, logs/009-perf-fp16.stderr.txt, and exact-head src/winml/modelkit/commands/perf.py precision-reporting audit.

Verified evidence

• PR state/scope: draft, model-scale-by-skill, exact head/base above; exactly two added recipe JSONs and no README/source/test change. Both recipes were read line-by-line and are sound L0 BERT reranking configs with three named int32 [1,512] inputs and scalar logits.
• Body/source fidelity: live body byte-normalizes exactly to the explainer artifact, has the required ordered Summary / Model metadata / Validation hierarchy and seven numbered evidence subsections, cites dependency and Lane A SHAs, and exposes no local scratch path.
• Coverage/evidence: fresh CPU fp32 and fp16 L0/L1, L2 raw-logit parity/order, and fp32 SciDocs L3 evidence otherwise pass with no deferred tuples. Perf: fp32 66.65 ms / 15.00 samples/s / +78.2 MB; fp16 94.50 ms / 10.58 samples/s / +87.9 MB. Parity: fp32 cosine 0.9999999999999492, max abs 3.814697265625e-06; fp16 cosine 0.9999999705508874, max abs 0.0034061670303344727; both preserve [0,1,2] order. SciDocs is correctly bounded to 2 groups, 10 candidates/group, 20 pairs and labeled functional smoke only.
• Artifacts/analyze: IR 8/opset 17, correct named I/O, true fp32/fp16 initializer semantics, and realized payload ratio about 0.50034. Both analyze outputs preserve complete seven-EP rows with exact exit 1/provider-bridge partial-success semantics; fp32 has 136 ops/16 types, fp16 137/16, mapped component coverage is 89 nodes with explicit 47/48 glue-node gaps, and QNN Gather is partial.
• Failed-root incident: preserved root tester-d951fcca-20260824-acceptance-b71e25c8 contains identical fp32/fp16 model bytes and initializer censuses. Its sole structure failure compared aggregate directories (230,182,646 vs 230,235,020 bytes); corrected root compares realized model.onnx payloads. Candidate stayed unchanged.
• Integrity/environment/quality: my independent verifier recomputed all 127 sealed hashes with zero mismatches and checked exact-head lock parity. Hosted run 32695311485 completed SUCCESS after immutable candidate checkout/assertion and locked candidate-local hydration; license headers, Ruff, mypy (444 files), and all five current non-hardware pytest partitions passed (8,441 passed total, 78 skipped, 3 xfailed, 1 deselected). A local independent locked hydration attempt was infrastructure-blocked by PyPI TLS, so I used the sealed hosted outputs only after verifying checkout, environment, interpreter/import roots, lock SHA, commands, and Actions job provenance.
• Methodology/knowledge: Lane A remote is ae5683540ed5cd5afe15a82f78d6f4f80cc6a2a8. Its correction changes exactly the L4 definition plus three tester/reviewer references to unique _meta-114; authoritative commit 113413ae5580e1c78c0afce20c5b14b0436f8d14 retains _meta-113 exclusively for the mMARCO analyze-wrapper contract. bert-024 through bert-029 contain the expected artifact, HTP, analyze, perf, parity, and bounded SciDocs knowledge.
• External feedback/checks: initial enumeration found 0 conversation comments, 0 line comments, 0 review threads, and 0 open threads. The live PR rollup at the reviewed SHA contains license/cla COMPLETED/SUCCESS; the stacked base does not trigger the main-target workflow files, while the immutable exact-candidate validation run above executed their complete command set successfully.

No structural fabrication or scope leakage was found, so this is REQUEST_CHANGES, not REJECT. Re-entry must use a fresh exact-head review after Planner/Tester/Explainer corrections; do not amend either sealed Tester root.

[ssss141414]

Explainer disposition for REQUEST_CHANGES comment #5391732432:

• Updated the durable and live PR body from Planner charter v2 (370d92787b44732d84bd42fcf543bdeb43f9b0eb9752cb19e3522189374efe0b): current-main baseline 774dd6e0811f6947eaf2bd8fb721d5f6e718b0c0, explicit FULL-RERUN impact/rerun provenance, fresh build/perf/Eval/Analyze results, and dependency divergence/stacking.
• Updated fp16 L1 from Tester v2 verdict (419e135a1bf88fdd88491fbe66f007d79760de9877d8c3eddfaac8c138f79f64): explicit Model Precision: fp16, 38,381,076-byte realized payload, 97.02 ms mean / 96.45 ms p50 / 120.96 ms p95, 10.31 samples/s, and +87.9 MB inference RAM. Unaffected exact-candidate stages remain preserved by the Tester v2 manifest.
• Updated Learner provenance to v3 and final Lane A commit e125b318d74a1319cedf5215b92f6bdd0ead9140; retained the harness incidents and unchanged candidate/dependency identities.

No candidate or branch files were changed. There were no line threads to resolve; this is an ordinary conversation reply and does not impersonate Reviewer.

[ssss141414]

Independent reviewer verdict: APPROVE

Reviewed immutable PR head d951fccaba37275540cdf2757bfb6b3cf60e8ec8, with exact sole parent and live stacked base 3708969b731425b0c6d4b97920d1b5e6519bb013. I performed a complete fresh fail-closed re-review after remediation and did not mutate the PR body, metadata, labels, draft state, branch, checks, or review threads.

Prior-gap closure

• Moved-main baseline gap: closed. Planner charter v2 SHA 370d92787b44732d84bd42fcf543bdeb43f9b0eb9752cb19e3522189374efe0b uses live current main 774dd6e0811f6947eaf2bd8fb721d5f6e718b0c0 and honestly classifies the move as FULL-RERUN. I independently checked all 18 changed-file classifications: direct Perf, Analyze, session, monitor, sysinfo, constants, and test surfaces are marked IMPACT; only isolated GenAI paths are NO-IMPACT. Inspect, both Config probes, Build, Perf, Eval, and Analyze were rerun on current main, with no old runtime stage relabeled. The candidate remains a one-commit/two-recipe patch on unchanged dependency Add WinML reranking support for cross-encoder/ms-marco-MiniLM-L6-v2 #1322; live main and dependency are correctly reported divergent at merge base 02179d185f0e3ac23518236b5ca1f61fd37ef680.
• Explicit fp16 evidence gap: closed. Fresh immutable Tester v2 root tester-d951fcca-20260824-reentry-v2-final-e82a41c6, verdict SHA 419e135a1bf88fdd88491fbe66f007d79760de9877d8c3eddfaac8c138f79f64, records Model Precision: fp16 for exact final artifact SHA 3146daac451a022a3f36740eaaa168936f718343e33a5eed5a83d9ea7260eab9. Independent ONNX inspection confirms 76 FLOAT16 initializers, 38,329,352 FLOAT16 raw payload bytes, and a 38,381,076-byte model versus fp32's 76 FLOAT initializers and 76,710,324-byte model. Fresh fp16 Perf is 97.02 ms mean / 96.45 ms p50 / 120.96 ms p95, 10.31 samples/s, and +87.9 MB inference RAM. The build used --precision fp16 without --no-quant.

Complete review result

• Coverage: full required CPU coverage for both fp32 and true fp16, with no deferred tuples. L0 build/structure, L1 Perf, and L2 raw-logit parity/order pass for both precisions. FP32 bounded SciDocs L3 passes as functional operability smoke only: 2 groups, 10 candidates/group, 20 pairs, MRR@10/Recall@1/Recall@10 all 1.0. Parity remains fp32 cosine 0.9999999999999492, max abs 3.814697265625e-06; fp16 cosine 0.9999999705508874, max abs 0.0034061670303344727; both preserve [0,1,2] order.
• Integrity: I independently rehashed all 43 files in the superseding seal and all 37 preserved-evidence references with zero mismatches, verified semantic validation before sealing, seal SHA 9190a6f5f9734d9167867a3829cb6e148670c788027781e7eb68914d82a4b12a, independent rehash success, completed phase, outer exit 0, no timeout/error, and exact candidate/dependency/artifact identities.
• Artifacts/analyze: both artifacts are IR 8/opset 17 with named input_ids, attention_mask, and token_type_ids as INT32 [1,512], plus FLOAT [1,1] logits. Analyze retains complete seven-EP partial-success rows: fp32 136 ops/16 types and fp16 137/16, 89 mapped component nodes with explicit 47/48 glue-node gaps, QNN partial Gather, and exact exit-1 provider-bridge semantics.
• Quality/hosted evidence: immutable-candidate hosted runs 32695311485 and 32701875911 are completed success with candidate-local locked environments and matching lock provenance. License headers and Ruff pass; mypy reports no issues in 444 files; five non-hardware pytest partitions total 8,441 passed, 78 skipped, 3 xfailed, and 1 deselected. The live PR check rollup on the reviewed SHA has 1 check: license/cla COMPLETED/SUCCESS.
• Body/scope/methodology: the live body byte-normalizes to the durable Explainer body and preserves the required ordered hierarchy and source-owned values. The PR remains draft with model-scale-by-skill, exactly two added recipe JSONs, no README/source/test change, and exact recipe blob hashes matching producer evidence. Learner v3 Lane A is remotely verified at e125b318d74a1319cedf5215b92f6bdd0ead9140; _meta-114 has one definition and three contract references, with no collision.
• External feedback: final paginated enumeration found 2 conversation comments, 0 line comments, 0 GraphQL review threads, and 0 open threads. Both conversation comments were read in full; the remediation disposition matches the live artifacts and body.

No blockers remain. This APPROVE is the skill-level opinion only and does not authorize marking the draft ready or any other PR mutation.