fix(export): bind keyword inputs and add OWLv2 recipes

[ssss141414]

Summary

This adds CPU fp32/fp16 support for google/owlv2-base-patch16-finetuned, an open-vocabulary zero-shot object detector, together with a class-wide exporter repair for keyword-generated multi-input models. The shipped Effort/Outcome is L2, and both required CPU tuples reached L2 PASS with semantic-input performance and all-four-output parity against the pinned PyTorch checkpoint. winml eval remains CLI-BLOCKED for both tuples because no zero-shot-object-detection evaluator exists; no task metric is fabricated.

Model metadata

What the model does

OWLv2 is an open-vocabulary, text-conditioned object detector. It accepts an image and one or more text queries and returns per-patch query logits, normalized bounding boxes, and text/image embeddings for locating objects not limited to a fixed training label set.

• Evidence: Hugging Face model card for google/owlv2-base-patch16-finetuned pinned at d69b086b07123308a4e198343ab2824e1af7774a; pinned config.json with architectures=["Owlv2ForObjectDetection"] and model_type="owlv2"; Transformers Owlv2ForObjectDetection.forward source.

• Confidence: verified.

Primary user stories

• A user supplies an image and natural-language candidate object labels to obtain scored bounding boxes for zero-shot text-conditioned object localization. Evidence: pinned model-card usage example and intended-use section. Confidence: verified.

• A researcher supplies images and previously unseen category descriptions to study open-vocabulary detector robustness, generalization, bias, and constraints. Evidence: pinned model-card intended-use section. Confidence: verified.

Supported tasks

| Task | Support surfaces | Evidence | Confidence |

|---|---|---|---|

| zero-shot-object-detection | checkpoint; Transformers; Optimum ONNX; WinML | Pinned Hub pipeline_tag; Transformers Owlv2ForObjectDetection; Optimum TasksManager vendor registration for owlv2; winml inspect resolution to OwlV2OnnxConfig | verified |

| feature-extraction | Transformers; Optimum ONNX | Transformers Owlv2Model/Owlv2TextModel/Owlv2VisionModel; Optimum TasksManager vendor registration | mapped |

Model architecture

Owlv2ForObjectDetection

├── OWLv2 multimodal backbone

│   ├── Causal text encoder: token + position embeddings → 12 layers (512 hidden, 8 heads, 2048 FFN) → 512 projection

│   └── ViT-B/16 image encoder: 960×960 image → 60×60 patches + class token → 12 layers (768 hidden, 12 heads, 3072 FFN)

├── Patch-text fusion: class-token-conditioned patch map + normalized text embeddings

├── Class head: patch/query similarity logits + learned shift/scale

├── Objectness head: MLP → 1 score per patch

└── Box head: MLP + grid bias → normalized cxcywh per patch

• Source/confidence: pinned checkpoint configuration and resolved nested text/vision configs, Transformers Owlv2ForObjectDetection source, and the exported ONNX hierarchy (verified).

Validation and support evidence

Baseline

Fresh baseline validation used current main commit 38767add6f91c7b10b6394fae3af6f437e02effd with WinML 0.2.0. Optimum advertised feature-extraction and zero-shot-object-detection both before and after WinML registration, with no WinML-added task: VENDOR-ONLY.

• Auto-config resolved zero-shot-object-detection, AutoModelForZeroShotObjectDetection, and OwlV2OnnxConfig, but emitted /export/input_tensors as [pixel_values,input_ids,attention_mask].

• Recipe-free build returned process exit 0 in 122.981 s, but failed the L0 floor: the graph bound pixel_values to INT32 [1,16] and input_ids to FLOAT [1,3,960,960], opposite Owlv2ForObjectDetection.forward(input_ids, pixel_values, attention_mask). An exit-zero export was therefore not valid task coverage.

• Semantic named-input perf failed with exit 1: Invalid rank for input: pixel_values Got: 4 Expected: 2. Random graph-shaped perf returned p50 3034.894 ms, throughput 0.33 samples/s, and RSS total delta 1426.47 MB, but is invalid task evidence because it follows the swapped graph bindings.

• Eval was CLI-BLOCKED before data/model execution: zero-shot-object-detection was absent from TASK_SCHEMAS and _EVALUATOR_REGISTRY. The pinned COCO mirror is object-detection data, but a valid evaluator must derive candidate text queries from category names, run grounded post-processing, and score COCO mAP; the fixed-label object-detection evaluator is not semantically interchangeable. The generalized evaluator gap is filed as winml-cli #1147.

Baseline Goal floor: L0.

Goal

• Effort: L2 — repair the class-of-models keyword-input name binding, add a regression test, and add the two required recipes.

• Goal ceiling: L2.

• Outcome commitment: L2 for CPU/cpu fp32 and fp16.

• Success definition: for both tuples, perform fresh config/build with exact named ONNX structure, run processor-generated semantic image+text perf, and compare all four raw outputs against pinned PyTorch revision d69b086b07123308a4e198343ab2824e1af7774a.

• No ceiling change or re-issued charter occurred.

Outcome

• Shipped tier: L2.

• Highest Goal verdict: L2 PASS.

• Coverage: full; both required CPU/cpu precision tuples passed; deferred tuples: none.

• Supplementary eval: CLI-BLOCKED for both tuples; no task metric fabricated.

• Shipped paths: the shared exporter repair and regression test, plus fp32/fp16 recipes under examples/recipes/google_owlv2-base-patch16-finetuned/cpu/cpu/.

• Model findings: owlv2-004 records faithful fp32/fp16 CPU export/perf/parity and the fp16 box-delta caveat; owlv2-005 records the zero-shot evaluator gap, why fixed-label evaluation is invalid, and the filed issue winml-cli #1147.

• Methodology finding: _meta-084 distinguishes recipe ordering defects from generalized keyword-export binding defects and requires recipe-free semantic acceptance while preserving explicit positional export protocols.

• Lane A provenance: these findings and the corresponding producer/tester/reviewer contract refinements are in source-branch commit 7cdfad2a05fc20d2e7ed96db93e666df20689485. This exact post-merge source-branch commit updates owlv2-005.feature_gaps_filed to winml-cli #1147. It remains on the source branch because ModelKitArtifacts PR #159 is already merged; no Lane A files are included in this model PR.

Per-EP/device/precision results — including perf and eval data

Goal ladder

| Tier | Verdict | Evidence |

|---|---|---|

| L0 | PASS | Both exact CPU/cpu precision tuples were freshly configured, built, and structurally validated. fp32/fp16 build times were 100.8660367 s / 116.974988 s; both expose input_ids INT32 [1,16], pixel_values FLOAT [1,3,960,960], and attention_mask INT32 [1,16], plus logits, pred_boxes, text_embeds, and image_embeds. |

| L1 | PASS | Both tuples completed 10 warmups and 100 measured processor-generated semantic-input iterations with memory capture. |

| L2 | PASS | Both tuples compared logits, pred_boxes, text_embeds, and image_embeds against pinned PyTorch on identical semantic inputs. |

Performance

| Tier | EP / Device | Precision | Verdict | Mean | p50 | p90 / p95 / p99 | Std | Warmup mean | Throughput | RAM Δ (load + inference) | VRAM local/shared Δ |

|---|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|

| L1 | CPUExecutionProvider / cpu | fp32 | PASS | 2864.275 ms | 2868.799 ms | 2990.744 / 3090.821 / 4458.42 ms | 227.563 ms | 3129.52 ms | 0.35 samples/s | 1423.95 MB (504.39 + 919.56) | 0.0 / 0.0 MB |

| L1 | CPUExecutionProvider / cpu | fp16 | PASS | 3958.256 ms | 3943.525 ms | 4091.145 / 4127.269 / 4566.801 ms | 94.429 ms | 4021.603 ms | 0.25 samples/s | 3634.66 MB (618.49 + 3016.17) | 0.0 / 0.0 MB |

L2 all-output parity

Parity is against PyTorch checkpoint revision d69b086b07123308a4e198343ab2824e1af7774a on the same processor-generated inputs, before post_process_grounded_object_detection.

| Precision | Output | Shape | Cosine similarity | Max abs diff | Mean abs diff | RMSE | Additional task-facing measure |

|---|---|---|---:|---:|---:|---:|---|

| fp32 | logits | [1,3600,1] | 0.9999999999973036 | 0.0002536773681640625 | 0.00003874407874213324 | 0.00005066664507799361 | sigmoid probability max Δ 1.0459867096952605e-09 |

| fp32 | pred_boxes | [1,3600,4] | 0.9999999999978494 | 0.000020384788513183594 | 6.614572420302365e-07 | 1.2354488877881363e-06 | mean raw normalized cxcywh IoU 0.9999946354690188; center L2 4.988567341990188e-07; size L2 1.6024903863345032e-06 |

| fp32 | text_embeds | [1,1,512] | 0.9999999999998044 | 1.1920928955078125e-07 | 2.1396012961361066e-08 | 2.7659271996813866e-08 | row cosine min/mean 0.9999999999998042 |

| fp32 | image_embeds | [1,60,60,768] | 0.9999999999992097 | 0.00016033649444580078 | 7.003135723185441e-07 | 1.258130226559189e-06 | row cosine min 0.9999999999611946; mean 0.9999999999992424 |

| fp16 | logits | [1,3600,1] | 0.9999985234500899 | 0.4004535675048828 | 0.03721526781717936 | 0.047675061767347246 | sigmoid probability max Δ 4.985980597621861e-07 |

| fp16 | pred_boxes | [1,3600,4] | 0.9999983128330443 | 0.016967952251434326 | 0.0005751266507690565 | 0.001102209982320831 | mean raw normalized cxcywh IoU 0.9954116316743082; center L2 0.00043634357334799317; size L2 0.0013853381709355733 |

| fp16 | text_embeds | [1,1,512] | 0.9999997517780107 | 0.0001614093780517578 | 0.00002433118910527554 | 0.000031142131733688554 | row cosine min/mean 0.9999997517780105 |

| fp16 | image_embeds | [1,60,60,768] | 0.99999922301875 | 0.18775832653045654 | 0.0006581207882101428 | 0.0012534093656211827 | row cosine min 0.9999292538054624; mean 0.9999992292296191 |

The fp16 pred_boxes max-absolute delta 0.016967952251434326 fails the strict auxiliary 0.005 absolute check. L2 remains PASS because all outputs are present and the task-facing box criterion passes: mean raw normalized-box IoU is 0.9954116316743082 against the 0.99 threshold. This does not claim dataset mAP, query derivation, threshold stability, target-size rescaling, or final detection-list equivalence.

Eval

| EP / Device | Precision | Verdict | Dataset | Revision | Subset | Exit | Metric |

|---|---|---|---|---|---|---:|---|

| CPUExecutionProvider / cpu | fp32 | CLI-BLOCKED | detection-datasets/coco | cf0b22332314a937e9dc8a1957b21725430bb41d | none | 1 | none |

| CPUExecutionProvider / cpu | fp16 | CLI-BLOCKED | detection-datasets/coco | cf0b22332314a937e9dc8a1957b21725430bb41d | none | 1 | none |

Both tuples returned the same error before dataset/model execution:

Error: Evaluation failed: Task 'zero-shot-object-detection' is not supported. Supported tasks: compare-tensor, depth-estimation, feature-extraction, fill-mask, image-classification, image-feature-extraction, image-segmentation, image-to-text, keypoint-detection, mask-generation, next-sentence-prediction, object-detection, question-answering, sentence-similarity, sequence-classification, text-classification, token-classification, zero-shot-classification, zero-shot-image-classification.

Block reason: no zero-shot-object-detection schema or evaluator exists; fixed-label object detection is semantically non-interchangeable. Filed gap: winml-cli #1147.

Delta

Recipe versus fresh auto-config

| Recipe | JSON pointer | Auto-config baseline | Shipped | Reason |

|---|---|---|---|---|

| fp32 | /export/input_tensors | [pixel_values,input_ids,attention_mask] | [input_ids,pixel_values,attention_mask] | Match the resolved Owlv2ForObjectDetection.forward order and remain independently safe for positional ONNX naming. |

| fp16 | /export/input_tensors | [pixel_values,input_ids,attention_mask] | [input_ids,pixel_values,attention_mask] | Same ordering repair. |

| fp16 | /quant | null | generated mode=fp16 configuration with fp16_keep_io_types=true | Produce the required exact fp16 CPU candidate while retaining fp32 I/O. |

All checkpoint dimensions and loader fields remain checkpoint-derived. The production recipe README is untouched.

General exporter change

HTPExporter._resolve_keyword_input_names and HTPExporter._convert_model_to_onnx now derive ONNX input_names for keyword-generated multi-input exports from the inspected forward signature filtered to the configured/generated input set. If inspection is unavailable or incomplete, configured order is preserved without guessing. Models with explicit get_export_args() retain their configured positional protocol; a dedicated regression test covers that behavior.

This is consistent with the charter's class-of-models fix. Recipe-free acceptance passed in 106.191 s with exact semantic bindings (input_ids INT32 [1,16], pixel_values FLOAT [1,3,960,960], attention_mask INT32 [1,16]), so acceptance is not dependent on the model-specific recipe ordering.

Analyze summary — component level and op level

ANALYZE-PARTIAL-SUCCESS: public static rule analysis produced all 11 classification rows for each artifact, but the process exited 1 after host OpenVINO plugin registration failed with Error 126 (onnxruntime_providers_shared.dll missing). These classifications are static rule analysis, not accelerator runtime execution or fresh accelerator support.

Rules provenance: public WinML CLI release v0.2.0, rules-v0.2.0.zip, containing 1,746 parquet rule files for TensorRT RTX GPU, OpenVINO CPU/GPU/NPU, and QNN GPU/NPU.

Component-level summary

| Artifact | Architecture coverage | Mapping | Actionable EP findings |

|---|---|---|---|

| fp32 | 12x causal text Transformer; 12x ViT-B/16 image Transformer; patch/text fusion; class, objectness, and box heads | 1007 mapped, 0 partial, 0 unmapped; confidence mapped | TensorRT RTX and OpenVINO CPU/GPU/NPU unknown: Cast, Einsum, Sqrt. QNN GPU/NPU partial: Add, Div, Erf, Mul; unknown: Cast, Einsum, Sqrt; none unsupported. |

| fp16 | same regions | 1012 mapped, 0 partial, 0 unmapped; confidence mapped | TensorRT RTX and OpenVINO CPU/GPU/NPU unknown: Cast, Einsum, Sqrt. QNN GPU/NPU partial: Add, Div, Erf, Expand, Mul; unknown: Cast, Einsum, Sqrt; none unsupported. |

Op-level summary

| Artifact | Graph | Dominant operators | EP roll-up |

|---|---|---|---|

| fp32 | 1007 operators / 27 types | Reshape 235; Add 231; MatMul 199; Transpose 121; Mul 81; LayerNormalization 52; Sigmoid 25; Softmax 24 | TensorRT RTX/OpenVINO/QNN findings as above; no fully-supported EP group established by the rules. |

| fp16 | 1012 operators / 27 types | Reshape 235; Add 231; MatMul 199; Transpose 121; Mul 81; LayerNormalization 52; Sigmoid 25; Softmax 24 | TensorRT RTX/OpenVINO/QNN findings as above; no fully-supported EP group established by the rules. |

Rule-less results for CPUExecutionProvider/CPU, CUDAExecutionProvider/GPU, DmlExecutionProvider/GPU, MIGraphXExecutionProvider/GPU, and VitisAIExecutionProvider/NPU classify all operator types as unknown.

Reproduce commands

The checkpoint and dataset are pinned below. The exact Hub revision is first materialized into $SNAPSHOT with huggingface_hub.snapshot_download; every WinML checkpoint load, processor load, and PyTorch reference then consumes that local snapshot. $INPUTS is a named semantic NPZ generated from a constant 960×960 RGB image and the text query a photo of a cat; this preserves the validation surface used for perf and parity.

$MODEL_ID = 'google/owlv2-base-patch16-finetuned'
$REV = 'd69b086b07123308a4e198343ab2824e1af7774a'
$OUT = 'artifacts/google_owlv2-base-patch16-finetuned-repro'
$SNAPSHOT = Join-Path $OUT 'checkpoint_snapshot'
$INPUTS = Join-Path $OUT 'semantic_inputs.npz'
New-Item -ItemType Directory -Force $OUT | Out-Null

winml --version
git rev-parse HEAD
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='$MODEL_ID', revision='$REV', local_dir=r'$SNAPSHOT')"
python -c "from pathlib import Path; import numpy as np; from PIL import Image; from transformers import AutoProcessor; out=Path(r'$INPUTS'); p=AutoProcessor.from_pretrained(r'$SNAPSHOT', use_fast=False); x=p(text=[['a photo of a cat']], images=Image.new('RGB',(960,960),(127,127,127)), return_tensors='np'); np.savez(out, input_ids=x['input_ids'].astype(np.int32), pixel_values=x['pixel_values'].astype(np.float32), attention_mask=x['attention_mask'].astype(np.int32))"

# Recipe-free baseline/acceptance: proves the shared exporter fix is not recipe-dependent.
winml build -m $SNAPSHOT -o $OUT/recipe-free --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild --no-color

winml build -c examples/recipes/google_owlv2-base-patch16-finetuned/cpu/cpu/zero-shot-object-detection_fp32_config.json -m $SNAPSHOT -o $OUT/fp32 --ep cpu --device cpu -p fp32 --no-analyze --no-compile --rebuild --no-color
winml build -c examples/recipes/google_owlv2-base-patch16-finetuned/cpu/cpu/zero-shot-object-detection_fp16_config.json -m $SNAPSHOT -o $OUT/fp16 --ep cpu --device cpu -p fp16 --no-analyze --no-compile --rebuild --no-color
winml perf -m $OUT/fp32/model.onnx --ep cpu --device cpu --precision fp32 --input-data $INPUTS --iterations 3 --warmup 1 --memory --output $OUT/perf_fp32.json --overwrite --format json --no-color
winml perf -m $OUT/fp16/model.onnx --ep cpu --device cpu --precision fp16 --input-data $INPUTS --iterations 3 --warmup 1 --memory --output $OUT/perf_fp16.json --overwrite --format json --no-color

$env:SNAPSHOT = $SNAPSHOT
$env:INPUTS = $INPUTS
$env:FP32_ONNX = Join-Path $OUT 'fp32/model.onnx'
$env:FP16_ONNX = Join-Path $OUT 'fp16/model.onnx'
@'
import os
import numpy as np
import onnxruntime as ort
import torch
from transformers import AutoModelForZeroShotObjectDetection

inputs_np = dict(np.load(os.environ['INPUTS']))
model = AutoModelForZeroShotObjectDetection.from_pretrained(os.environ['SNAPSHOT']).eval()
inputs_pt = {name: torch.from_numpy(value) for name, value in inputs_np.items()}
with torch.no_grad():
    reference = model(**inputs_pt)
reference_by_name = {
    'logits': reference.logits.detach().cpu().numpy(),
    'pred_boxes': reference.pred_boxes.detach().cpu().numpy(),
    'text_embeds': reference.text_embeds.detach().cpu().numpy(),
    'image_embeds': reference.image_embeds.detach().cpu().numpy(),
}
for precision, model_path in [('fp32', os.environ['FP32_ONNX']), ('fp16', os.environ['FP16_ONNX'])]:
    session = ort.InferenceSession(model_path, providers=['CPUExecutionProvider'])
    actual = dict(zip([output.name for output in session.get_outputs()], session.run(None, inputs_np)))
    for name, expected in reference_by_name.items():
        observed = actual[name].astype(np.float64).ravel()
        expected = expected.astype(np.float64).ravel()
        cosine = np.dot(observed, expected) / (np.linalg.norm(observed) * np.linalg.norm(expected))
        print(precision, name, 'cosine=', float(cosine), 'max_abs=', float(np.max(np.abs(observed - expected))))
'@ | python -
Remove-Item Env:SNAPSHOT,Env:INPUTS,Env:FP32_ONNX,Env:FP16_ONNX

$RULES_ZIP = Join-Path $OUT 'rules-v0.2.0.zip'
$RULES = Join-Path $OUT 'rules-v0.2.0'
Invoke-WebRequest 'https://github.com/microsoft/winml-cli/releases/download/v0.2.0/rules-v0.2.0.zip' -OutFile $RULES_ZIP -UseBasicParsing
Expand-Archive -Path $RULES_ZIP -DestinationPath $RULES -Force
$env:WINMLCLI_RULES_DIR = (Resolve-Path $RULES).Path
winml analyze --model $OUT/fp32/model.onnx --ep all --device all --htp-metadata $OUT/fp32/export_htp_metadata.json --output $OUT/analyze_fp32.json --overwrite --format json --no-color
winml analyze --model $OUT/fp16/model.onnx --ep all --device all --htp-metadata $OUT/fp16/export_htp_metadata.json --output $OUT/analyze_fp16.json --overwrite --format json --no-color
Remove-Item Env:WINMLCLI_RULES_DIR
winml eval -m $OUT/fp32/model.onnx --model-id $SNAPSHOT --task zero-shot-object-detection --dataset detection-datasets/coco --dataset-revision cf0b22332314a937e9dc8a1957b21725430bb41d --split val --samples 1 --no-shuffle --streaming --ep cpu --device cpu --output $OUT/eval_fp32.json --overwrite --format json --no-color
winml eval -m $OUT/fp16/model.onnx --model-id $SNAPSHOT --task zero-shot-object-detection --dataset detection-datasets/coco --dataset-revision cf0b22332314a937e9dc8a1957b21725430bb41d --split val --samples 1 --no-shuffle --streaming --ep cpu --device cpu --output $OUT/eval_fp16.json --overwrite --format json --no-color

[ssss141414]

Independent reviewer verdict: REQUEST_CHANGES

Reviewed exact PR head 3c94de2ef79fc9310150bfc498da157c95622578 against current origin/main 38767add6f91c7b10b6394fae3af6f437e02effd from a fresh detached worktree.

Blocking findings and routing

1. Explainer — the public pinned reproduction block cannot execute. The PR body constructs $MODEL="$MODEL_ID@$REV" and passes it to both winml build commands. I ran the fp32 command literally with google/owlv2-base-patch16-finetuned@d69b086b07123308a4e198343ab2824e1af7774a; winml build exited 2 before export: is not a valid HuggingFace model identifier (expected 'name' or 'org/name'). winml build --help exposes no revision option. Replace this with an executable pinned-revision strategy supported by the CLI, update downstream --model-id use as needed, and run the entire public block literally before re-review. This fails the required reproducibility gate even though unrevisioned independent builds succeed.

2. Learner, then explainer — the evaluator feature gap was described but not filed. Lane A finding owlv2-005.feature_gaps_filed[] contains bare prose, neither an issue URL nor a FILE: entry. Because this Outcome-L1+ PR changes shared exporter code, the reviewer contract requires feature-gap entries to resolve to filed issues; a FILE:-only state would still remain REQUEST_CHANGES. File the zero-shot-object-detection evaluator gap, update the pushed Lane A finding with the issue URL, and update the PR's public provenance/evidence to cite it.

Independently verified, not blocking

• Scope: exactly src/winml/modelkit/export/htp/exporter.py, tests/unit/export/test_pytorch_export.py, and the fp32/fp16 recipe JSON files under examples/recipes/google_owlv2-base-patch16-finetuned/cpu/cpu/; the production recipe README and Lane A files are absent from the model PR.
• Code/design: the shared exporter derives keyword input names from the filtered model.forward signature and leaves explicit get_export_args() positional protocols intact; no model-ID special case was introduced.
• Tests/lint: changed-file Ruff/format passed; focused exporter tests passed (27); full tests/unit/export passed (454 passed, 6 skipped, 1 xfailed). Repository-wide Ruff's only diagnostic reproduces on clean main in an untouched SAM script.
• Builds: recipe-free, fp32, and fp16 builds all exited 0 and printed Build complete. The semantic no-recipe path binds input_ids, pixel_values, and attention_mask correctly.
• Structure/precision: all three artifacts are opset 17 with the exact three named inputs and four outputs. fp32 has 419 FLOAT initializers and a 613,625,088-byte sidecar; fp16 has 419 FLOAT16 initializers and a 306,812,032-byte adjacent sidecar with fp32 I/O.
• Perf: independent named-semantic-input runs completed 10 warmups + 100 measured iterations with memory. fp32 reproduced p50 2747.455 ms, 0.35 samples/s, and 1422.45 MB RSS delta (published p50 delta -4.23%); fp16 reproduced p50 4018.968 ms, 0.25 samples/s, and 3635.53 MB RSS delta (published p50 delta +1.91%). Both are within the ±20% reviewer tolerance and report the claimed model precision.
• Parity: independent pinned-PyTorch comparison reproduced every published value. fp32 passed all checks. fp16 reproduced cosine/IoU checks and the disclosed auxiliary box max-absolute miss (0.016967952251434326 > 0.005, mean normalized-box IoU 0.9954116316743082 > 0.99); the PR describes this limitation rather than hiding it.
• Eval: both exact PR eval commands independently exited 1 with the documented unsupported zero-shot-object-detection error before data/model execution. CLI-BLOCKED is honest; fixed-label object detection was not substituted.
• Analysis/provenance: compact component/op evidence and architecture mapping are present. The Lane A text correctly caveats that commit b189423342cdc5c6085d8f7810fccad6488bf6ad was pushed to the source branch after ModelKitArtifacts PR P1-INFRA-003: CI/CD Pipeline — ADO + GitHub Actions + Self-hosted Runners #159 had already merged; it must not be represented as contained in that merge.
• GitHub gates: 0 issue comments, 0 inline comments, 0 reviews, 0 GraphQL review threads (0 open; no next page). PR remains open Draft with model-scale-by-skill; all 9 checks are completed successfully.

The PR should remain Draft. Re-review from a fresh checkout after both owner-routed fixes land.

[ssss141414]

REQUEST_CHANGES mapping response

Resolved both blocking findings at unchanged model PR head 3c94de2ef79fc9310150bfc498da157c95622578 (no production-code or recipe commit added).

1. Public pinned reproduction — resolved in the PR body. The unsupported $MODEL_ID@$REV construction is removed. The block now materializes exact Hub revision d69b086b07123308a4e198343ab2824e1af7774a into $SNAPSHOT with huggingface_hub.snapshot_download, then routes every WinML checkpoint load, processor load, and PyTorch reference through $SNAPSHOT. It also includes explicit winml --version, git rev-parse HEAD, a recipe-free baseline/acceptance build, fp32/fp16 recipe builds, semantic perf/parity, public pinned rules, analyze, and the expected blocked eval probes. Literal validation from a clean $OUT confirmed snapshot materialization, recipe-free/fp32/fp16 builds, semantic perf, all-eight parity values, the 1,746-file public rules layout and absolute rules path, and both expected unsupported-task eval blockers.

2. Evaluator feature gap — resolved by issue #1147. The issue describes the generalized missing zero-shot-object-detection schema/evaluator, provides portable pinned-snapshot reproduction with no local paths, and explains why fixed-label object detection is not a valid substitute. In Learner ownership, owlv2-005.feature_gaps_filed now cites that URL at exact post-merge ModelKitArtifacts source-branch commit 7cdfad2a05fc20d2e7ed96db93e666df20689485; PR P1-INFRA-003: CI/CD Pipeline — ADO + GitHub Actions + Self-hosted Runners #159 is already merged, so the PR body records this exact new source-branch provenance without adding Lane A files here.

Preserved: Draft state, model-scale-by-skill, unchanged model PR head, and successful CI.

[ssss141414]

Independent reviewer verdict: APPROVE

Reviewed exact PR head 3c94de2ef79fc9310150bfc498da157c95622578 after remediation response #1146 (comment). The PR must remain Draft.

Prior blockers independently closed

1. Pinned reproduction commands — CLOSED. The unsupported $MODEL_ID@$REV form is absent. The public block now materializes Hub revision d69b086b07123308a4e198343ab2824e1af7774a with snapshot_download, uses $SNAPSHOT in all three winml build commands and both winml eval --model-id commands, and routes AutoProcessor plus the PyTorch reference through that same snapshot. From a clean output directory I independently materialized the exact revision (621,600,870 bytes, including the 620,006,741-byte checkpoint), generated the named semantic inputs, and ran the recipe-free snapshot build: exit 0, Build complete, model present, and no invalid-Hugging-Face-identifier error. Both fp32/fp16 eval probes accepted the snapshot path and reached only the documented unsupported-task blocker.

2. Generalized evaluator issue and Lane A provenance — CLOSED. Add zero-shot object detection schema and evaluator to winml eval #1147 is a real open generalized feature gap: it specifies candidate-query derivation, grounded post-processing, query-local-to-dataset category mapping, COCO mAP, and why fixed-label object detection is not interchangeable. Pushed commit https://github.com/gim-home/ModelKitArtifacts/commit/7cdfad2a05fc20d2e7ed96db93e666df20689485 changes only model_knowledge/owlv2.json and replaces the bare gap with issue Add zero-shot object detection schema and evaluator to winml eval #1147. ModelKitArtifacts PR P1-INFRA-003: CI/CD Pipeline — ADO + GitHub Actions + Self-hosted Runners #159 merged earlier at e98a125db71d6d1421f288dff37fde0c5ce8262e; its 22 recorded commits do not contain 7cdfad2a. The model PR accurately describes 7cdfad2a as an exact post-merge source-branch commit rather than claiming it was in PR P1-INFRA-003: CI/CD Pipeline — ADO + GitHub Actions + Self-hosted Runners #159.

Independent revalidation

• Checkout/scope: exact requested head; merge base and current origin/main are 38767add6f91c7b10b6394fae3af6f437e02effd (rev-list count 0). Diff remains exactly two CPU recipe JSONs, exporter.py, and its focused unit test. Production recipe README is unchanged.
• Code and focused checks: generalized signature-derived keyword input ordering remains intact while explicit positional get_export_args() protocols remain authoritative. Changed-file Ruff passed, Ruff format passed, and focused exporter tests passed: 27 passed.
• Public contract: architecture is canonically unchanged; Delta/Analyze are canonically unchanged; perf/parity/eval values and blocker are unchanged, with the only semantic addition in that section being the filed issue Add zero-shot object detection schema and evaluator to winml eval #1147 link. The PR body contains no absolute/local scratch path leakage.
• Prior full evidence remains applicable: model head is unchanged; independent fp32/fp16 build, structure/precision, 100-iteration perf, all-output parity, eval, and analyze review findings therefore remain valid. The fresh pinned-snapshot build additionally validates the remediated command surface.
• External review surfaces: 2 issue comments enumerated, 0 inline comments, 0 submitted reviews, 0 GraphQL review threads, 0 open threads.
• GitHub gates: OPEN, MERGEABLE, Draft, model-scale-by-skill present, exact head preserved, and all 9/9 CI checks successful.

No remaining reviewer-owned or role-routed blocker was found. APPROVE, while preserving Draft state and the label.