Merge nucleic/sleek-ember-seal-uady into dev

ENERGY_AND_RESIDENCY.md

@@ -19,6 +19,13 @@ placement.
 
 ## macOS
 
+- Generate the ML Program with `convert_coreml.py`, then run the gated `eval.py
+  --coreml-model ... --coreml-compute-units cpu-and-ne --compare-onnx ...` command from
+  the README. Preserve the conversion manifest and frozen report with the model metrics.
+- Run `inspect_coreml.py` with `--compute-units cpu-and-ne`; preserve its full operation
+  report and record both `neuralEngineOperationShare` and
+  `neuralEngineEstimatedCostShare`. A gated evaluation without the ONNX comparison is
+  invalid.
 - Run the release Core ML artifact once with `.cpuAndNeuralEngine`, recording the
   `MLComputePlan` ANE operation share and the artifact's required floor.
 - Repeat with a diagnostic CPU-only configuration on the same Mac and power source.

README.md

@@ -196,6 +196,59 @@ routing-tier-drift gates pass. This is the current accuracy-qualified shipping c
 latency and energy/residency still require measurement on the target Apple and Windows
 accelerator runtimes.
 
+### Convert and validate Core ML
+
+Core ML Tools no longer maintains the legacy ONNX converter, so the Apple artifact is
+converted directly from the selected Hugging Face checkpoint. `convert_coreml.py` uses a
+fixed-shape export-only BERT forward to avoid dynamic Transformers masking helpers, checks
+that forward against Transformers before conversion, writes an ML Program package, and
+records hashes for every package file.
+
+Install the pinned converter in the macOS environment and create the package:
+
+```bash
+ml/purpose-classifier/venv/bin/python -m pip install \
+  -r ml/purpose-classifier/requirements-coreml.txt
+ml/purpose-classifier/venv/bin/python ml/purpose-classifier/convert_coreml.py \
+  --model-dir \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/model \
+  --output \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/purpose-lite-v1-fp16.mlpackage \
+  --overwrite-output
+```
+
+Run the frozen gate with CPU+Neural Engine placement and compare labels directly with the
+accepted int8 ONNX artifact. Gated Core ML evaluation fails closed without
+`--compare-onnx`, and requires at least 99.5% scorable label agreement:
+
+```bash
+ml/purpose-classifier/venv/bin/python ml/purpose-classifier/eval.py \
+  --model-dir \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/model \
+  --calibration \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/calibration.json \
+  --coreml-model \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/purpose-lite-v1-fp16.mlpackage \
+  --coreml-compute-units cpu-and-ne \
+  --compare-onnx \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/export/purpose-lite-v1-int8-qdq.onnx \
+  --report \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/ane-frozen-eval.json
+```
+
+Record the compute plan separately; this reports both operation-count and estimated-cost
+ANE shares. Repeat evaluation with `--coreml-compute-units cpu-only --no-gate` before the
+energy comparison in `ENERGY_AND_RESIDENCY.md`:
+
+```bash
+ml/purpose-classifier/venv/bin/python ml/purpose-classifier/inspect_coreml.py \
+  --model \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/purpose-lite-v1-fp16.mlpackage \
+  --compute-units cpu-and-ne \
+  --report \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/ane-compute-plan.json
+```
+
 For a wiring smoke test, use a small deterministic prefix:
 
 ```bash

convert_coreml.py

@@ -0,0 +1,354 @@
+#!/usr/bin/env python3
+"""Convert the selected purpose-lite checkpoint to a fixed-shape Core ML package."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+import shutil
+import sys
+from pathlib import Path
+from typing import Any, Sequence
+
+import numpy as np
+import torch
+import torch.nn.functional as functional
+
+from purpose_data import LABELS, DataError, load_jsonl, write_json
+from train import MAX_LENGTH, encode_fixed_shape
+
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+DEFAULT_MODEL_DIR = (
+    SCRIPT_DIR
+    / "outputs"
+    / "purpose-lite-v1-distilled-qat-mlx-4e"
+    / "model"
+)
+DEFAULT_VALIDATION = SCRIPT_DIR / ".artifacts" / "dataset-v1" / "validation.jsonl"
+DEFAULT_OUTPUT = (
+    SCRIPT_DIR
+    / "outputs"
+    / "purpose-lite-v1-distilled-qat-mlx-4e"
+    / "coreml"
+    / "purpose-lite-v1-fp16.mlpackage"
+)
+COREMLTOOLS_VERSION = "9.0"
+
+
+class FixedShapeBertForCoreML(torch.nn.Module):
+    """Conversion-only BERT forward without Transformers' dynamic mask helpers."""
+
+    def __init__(self, model: Any) -> None:
+        super().__init__()
+        self.model = model
+        config = model.config
+        self.num_heads = int(config.num_attention_heads)
+        self.head_size = int(config.hidden_size) // self.num_heads
+        if int(config.hidden_size) % self.num_heads:
+            raise DataError("BERT hidden size must be divisible by attention heads")
+        if int(config.max_position_embeddings) < MAX_LENGTH:
+            raise DataError("BERT checkpoint cannot represent the fixed 128-token input")
+        self.register_buffer(
+            "fixed_position_ids",
+            torch.arange(MAX_LENGTH, dtype=torch.int64).reshape(1, MAX_LENGTH),
+            persistent=False,
+        )
+
+    def forward(
+        self,
+        input_ids: Any,
+        attention_mask: Any,
+        token_type_ids: Any,
+    ) -> Any:
+        bert = self.model.bert
+        input_ids = input_ids.to(torch.int64)
+        token_type_ids = token_type_ids.to(torch.int64)
+        value = (
+            bert.embeddings.word_embeddings(input_ids)
+            + bert.embeddings.position_embeddings(self.fixed_position_ids)
+            + bert.embeddings.token_type_embeddings(token_type_ids)
+        )
+        value = bert.embeddings.LayerNorm(value)
+        zero = torch.zeros((), dtype=value.dtype, device=value.device)
+        hidden = torch.full((), -10000.0, dtype=value.dtype, device=value.device)
+        additive_mask = torch.where(
+            attention_mask[:, None, None, :] != 0,
+            zero,
+            hidden,
+        )
+
+        for layer in bert.encoder.layer:
+            self_attention = layer.attention.self
+
+            def split_heads(projected: Any) -> Any:
+                return projected.reshape(
+                    1,
+                    MAX_LENGTH,
+                    self.num_heads,
+                    self.head_size,
+                ).permute(0, 2, 1, 3)
+
+            queries = split_heads(self_attention.query(value))
+            keys = split_heads(self_attention.key(value))
+            values = split_heads(self_attention.value(value))
+            scores = torch.matmul(queries, keys.transpose(-1, -2)) / math.sqrt(
+                self.head_size
+            )
+            probabilities = torch.softmax(scores + additive_mask, dim=-1)
+            context = torch.matmul(probabilities, values)
+            context = context.permute(0, 2, 1, 3).reshape(
+                1,
+                MAX_LENGTH,
+                -1,
+            )
+            attention_output = layer.attention.output.dense(context)
+            value = layer.attention.output.LayerNorm(value + attention_output)
+            intermediate = functional.gelu(
+                layer.intermediate.dense(value),
+                approximate="none",
+            )
+            value = layer.output.LayerNorm(
+                value + layer.output.dense(intermediate)
+            )
+
+        pooled = torch.tanh(bert.pooler.dense(value[:, 0]))
+        return self.model.classifier(pooled)
+
+
+def _checkpoint_config(model_dir: Path) -> dict[str, Any]:
+    config_path = model_dir / "config.json"
+    try:
+        config = json.loads(config_path.read_text(encoding="utf-8"))
+    except (OSError, UnicodeError, json.JSONDecodeError) as exc:
+        raise DataError(f"{config_path}: cannot load model config: {exc}") from exc
+    configured_labels = [
+        config.get("id2label", {}).get(
+            str(index),
+            config.get("id2label", {}).get(index),
+        )
+        for index in range(len(LABELS))
+    ]
+    if (
+        config.get("model_type") != "bert"
+        or config.get("hidden_size") != 384
+        or config.get("num_hidden_layers") != 6
+    ):
+        raise DataError("Core ML conversion requires the purpose-lite BERT architecture")
+    if configured_labels != list(LABELS):
+        raise DataError("Core ML checkpoint label order does not match purpose-lite")
+    return config
+
+
+def _sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _package_manifest(package: Path) -> dict[str, Any]:
+    files = []
+    for path in sorted(item for item in package.rglob("*") if item.is_file()):
+        files.append(
+            {
+                "path": str(path.relative_to(package)),
+                "bytes": path.stat().st_size,
+                "sha256": _sha256(path),
+            }
+        )
+    return {
+        "schemaVersion": 1,
+        "modelVersion": "purpose-lite-v1",
+        "package": package.name,
+        "bytes": sum(item["bytes"] for item in files),
+        "files": files,
+    }
+
+
+def _verify_wrapper_parity(
+    torch_model: Any,
+    wrapper: FixedShapeBertForCoreML,
+    tokenizer: Any,
+    records: Sequence[dict[str, Any]],
+) -> float:
+    encoded = encode_fixed_shape(
+        tokenizer,
+        [record["prompt"] for record in records],
+        torch,
+    )
+    maximum_error = 0.0
+    with torch.inference_mode():
+        for index in range(len(records)):
+            item = {key: value[index : index + 1] for key, value in encoded.items()}
+            reference = torch_model(**item).logits
+            candidate = wrapper(
+                item["input_ids"].to(torch.int32),
+                item["attention_mask"].to(torch.int32),
+                item.get("token_type_ids", torch.zeros_like(item["input_ids"])).to(
+                    torch.int32
+                ),
+            )
+            error = float(torch.max(torch.abs(reference - candidate)).item())
+            maximum_error = max(maximum_error, error)
+            torch.testing.assert_close(candidate, reference, rtol=1e-5, atol=2e-5)
+            if int(candidate.argmax(dim=-1).item()) != int(
+                reference.argmax(dim=-1).item()
+            ):
+                raise DataError("conversion wrapper changed a parity prediction")
+    return maximum_error
+
+
+def convert(args: argparse.Namespace) -> dict[str, Any]:
+    _checkpoint_config(args.model_dir)
+    checkpoint = args.model_dir / "model.safetensors"
+    if not checkpoint.is_file():
+        raise DataError(f"{checkpoint}: checkpoint is missing")
+    if args.output.suffix != ".mlpackage":
+        raise DataError("Core ML output must end in .mlpackage")
+    if args.output.exists():
+        if not args.overwrite_output:
+            raise DataError(
+                f"{args.output}: output exists; pass --overwrite-output intentionally"
+            )
+        if args.output.is_dir():
+            shutil.rmtree(args.output)
+        else:
+            args.output.unlink()
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+
+    try:
+        import coremltools as ct
+        from transformers import AutoModelForSequenceClassification, AutoTokenizer
+    except ImportError as exc:
+        raise DataError(
+            "Core ML conversion requires requirements-coreml.txt on macOS"
+        ) from exc
+    if ct.__version__ != COREMLTOOLS_VERSION:
+        raise DataError(
+            f"expected coremltools {COREMLTOOLS_VERSION}, found {ct.__version__}"
+        )
+
+    model = AutoModelForSequenceClassification.from_pretrained(
+        args.model_dir,
+        local_files_only=True,
+    ).eval()
+    tokenizer = AutoTokenizer.from_pretrained(args.model_dir, local_files_only=True)
+    wrapper = FixedShapeBertForCoreML(model).eval()
+    validation = load_jsonl(args.validation)
+    if len(validation) < args.parity_records:
+        raise DataError("validation split is smaller than --parity-records")
+    maximum_error = _verify_wrapper_parity(
+        model,
+        wrapper,
+        tokenizer,
+        validation[: args.parity_records],
+    )
+    print(f"conversion-wrapper parity: max_abs_error={maximum_error:.3g}")
+
+    example = (
+        torch.zeros((1, MAX_LENGTH), dtype=torch.int32),
+        torch.ones((1, MAX_LENGTH), dtype=torch.int32),
+        torch.zeros((1, MAX_LENGTH), dtype=torch.int32),
+    )
+    with torch.inference_mode():
+        traced = torch.jit.trace(wrapper, example, strict=True)
+        traced = torch.jit.freeze(traced)
+    deployment_target = getattr(ct.target, args.minimum_deployment_target, None)
+    if deployment_target is None:
+        raise DataError(
+            f"coremltools does not support {args.minimum_deployment_target}"
+        )
+    coreml_model = ct.convert(
+        traced,
+        convert_to="mlprogram",
+        minimum_deployment_target=deployment_target,
+        compute_precision=ct.precision.FLOAT16,
+        inputs=[
+            ct.TensorType(
+                name="input_ids",
+                shape=(1, MAX_LENGTH),
+                dtype=np.int32,
+            ),
+            ct.TensorType(
+                name="attention_mask",
+                shape=(1, MAX_LENGTH),
+                dtype=np.int32,
+            ),
+            ct.TensorType(
+                name="token_type_ids",
+                shape=(1, MAX_LENGTH),
+                dtype=np.int32,
+            ),
+        ],
+        outputs=[ct.TensorType(name="logits", dtype=np.float32)],
+    )
+    coreml_model.author = "Nucleic"
+    coreml_model.short_description = "purpose-lite-v1 prompt classifier"
+    coreml_model.version = "purpose-lite-v1"
+    coreml_model.user_defined_metadata["com.nucleic.model.version"] = (
+        "purpose-lite-v1"
+    )
+    coreml_model.user_defined_metadata["com.nucleic.model.labels"] = json.dumps(
+        list(LABELS),
+        separators=(",", ":"),
+    )
+    coreml_model.user_defined_metadata["com.nucleic.model.sourceSha256"] = _sha256(
+        checkpoint
+    )
+    coreml_model.user_defined_metadata["com.nucleic.model.fixedShape"] = "1x128"
+    coreml_model.user_defined_metadata["com.nucleic.model.minimumDeploymentTarget"] = (
+        args.minimum_deployment_target
+    )
+    coreml_model.save(str(args.output))
+
+    manifest = _package_manifest(args.output)
+    manifest.update(
+        {
+            "sourceCheckpoint": str(checkpoint),
+            "sourceCheckpointSha256": _sha256(checkpoint),
+            "coremltoolsVersion": ct.__version__,
+            "minimumDeploymentTarget": args.minimum_deployment_target,
+            "computePrecision": "float16",
+            "wrapperMaximumAbsoluteError": maximum_error,
+        }
+    )
+    manifest_path = args.output.with_name(f"{args.output.stem}-manifest.json")
+    write_json(manifest_path, manifest)
+    print(f"Core ML package: {args.output} ({manifest['bytes']} bytes)")
+    return manifest
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model-dir", type=Path, default=DEFAULT_MODEL_DIR)
+    parser.add_argument("--validation", type=Path, default=DEFAULT_VALIDATION)
+    parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
+    parser.add_argument("--parity-records", type=int, default=16)
+    parser.add_argument(
+        "--minimum-deployment-target",
+        default="macOS15",
+        choices=("macOS15", "macOS26"),
+    )
+    parser.add_argument("--overwrite-output", action="store_true")
+    return parser
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    parser = build_parser()
+    args = parser.parse_args(argv)
+    if args.parity_records <= 0:
+        parser.error("--parity-records must be positive")
+    try:
+        convert(args)
+    except (AssertionError, DataError, OSError, RuntimeError, ValueError) as exc:
+        print(f"error: {exc}", file=sys.stderr)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

eval.py

@@ -13,6 +13,8 @@ from collections import Counter
 from pathlib import Path
 from typing import Any, Sequence
 
+import numpy as np
+
 from purpose_data import (
     HARD_SLICES,
     LABELS,
@@ -100,6 +102,16 @@ def _synchronize(torch: Any, device: Any) -> None:
         torch.mps.synchronize()
 
 
+def _coreml_compute_unit(coremltools: Any, requested: str) -> Any:
+    values = {
+        "all": coremltools.ComputeUnit.ALL,
+        "cpu-only": coremltools.ComputeUnit.CPU_ONLY,
+        "cpu-and-gpu": coremltools.ComputeUnit.CPU_AND_GPU,
+        "cpu-and-ne": coremltools.ComputeUnit.CPU_AND_NE,
+    }
+    return values[requested]
+
+
 def routing_tier_drift(
     records: Sequence[dict[str, Any]],
     actual: Sequence[int],
@@ -245,8 +257,62 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
     label_to_id = {label: index for index, label in enumerate(LABELS)}
     tokenizer = AutoTokenizer.from_pretrained(args.model_dir, local_files_only=True)
     onnx_session = None
+    coreml_model = None
     reference_model = None
-    if args.onnx_model is not None:
+    reference_onnx_session = None
+    reference_onnx_input_names: set[str] = set()
+    if args.compare_pytorch and args.compare_onnx is not None:
+        raise DataError("choose only one parity reference")
+    if args.coreml_model is not None:
+        try:
+            import coremltools as ct
+        except ImportError as exc:
+            raise DataError(
+                "Core ML evaluation requires requirements-coreml.txt on macOS"
+            ) from exc
+        if args.device != "auto":
+            raise DataError(
+                "Core ML compute placement uses --coreml-compute-units, not --device"
+            )
+        coreml_model = ct.models.MLModel(
+            str(args.coreml_model),
+            compute_units=_coreml_compute_unit(ct, args.coreml_compute_units),
+        )
+        coreml_input_names = {
+            item.name for item in coreml_model.get_spec().description.input
+        }
+        device = torch.device("cpu")
+        model = None
+        onnx_input_names = set()
+        runtime_name = f"coreml-{args.coreml_compute_units}"
+        if args.compare_pytorch:
+            reference_model = AutoModelForSequenceClassification.from_pretrained(
+                args.model_dir,
+                local_files_only=True,
+            ).to(device)
+            reference_model.eval()
+        if args.compare_onnx is not None:
+            try:
+                import onnxruntime as ort
+            except ImportError as exc:
+                raise DataError(
+                    "Core ML↔ONNX parity requires onnxruntime"
+                ) from exc
+            reference_options = ort.SessionOptions()
+            reference_options.graph_optimization_level = (
+                ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+            )
+            reference_onnx_session = ort.InferenceSession(
+                str(args.compare_onnx),
+                sess_options=reference_options,
+                providers=["CPUExecutionProvider"],
+            )
+            reference_onnx_input_names = {
+                item.name for item in reference_onnx_session.get_inputs()
+            }
+    elif args.onnx_model is not None:
+        if args.compare_onnx is not None:
+            raise DataError("--compare-onnx requires --coreml-model")
         try:
             import onnxruntime as ort
         except ImportError as exc:
@@ -273,8 +339,12 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
             ).to(device)
             reference_model.eval()
     else:
+        if args.compare_onnx is not None:
+            raise DataError("--compare-onnx requires --coreml-model")
         if args.compare_pytorch:
-            raise DataError("--compare-pytorch requires --onnx-model")
+            raise DataError(
+                "--compare-pytorch requires --onnx-model or --coreml-model"
+            )
         device = _device(torch, args.device)
         model = AutoModelForSequenceClassification.from_pretrained(
             args.model_dir, local_files_only=True
@@ -284,6 +354,14 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
         runtime_name = "pytorch"
 
     def predict_logits(encoded: dict[str, Any]) -> Any:
+        if coreml_model is not None:
+            inputs = {
+                key: value.numpy().astype(np.int32, copy=False)
+                for key, value in encoded.items()
+                if key in coreml_input_names
+            }
+            output = np.asarray(coreml_model.predict(inputs)["logits"])
+            return torch.from_numpy(output.reshape(1, len(LABELS)))
         if onnx_session is not None:
             inputs = {
                 key: value.numpy()
@@ -300,7 +378,9 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
     confidences: list[str] = []
     margins: list[float] = []
     reference_predictions: list[int] = []
-    inference_batch_size = 1 if onnx_session is not None else args.batch_size
+    inference_batch_size = (
+        1 if onnx_session is not None or coreml_model is not None else args.batch_size
+    )
     with torch.inference_mode():
         for start in range(0, len(records), inference_batch_size):
             batch = records[start : start + inference_batch_size]
@@ -315,6 +395,19 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
                 reference_predictions.extend(
                     reference_logits.argmax(dim=-1).tolist()
                 )
+            elif reference_onnx_session is not None:
+                reference_inputs = {
+                    key: value.numpy()
+                    for key, value in encoded.items()
+                    if key in reference_onnx_input_names
+                }
+                reference_logits = reference_onnx_session.run(
+                    ["logits"],
+                    reference_inputs,
+                )[0]
+                reference_predictions.extend(
+                    np.asarray(reference_logits).argmax(axis=-1).tolist()
+                )
             distribution = torch.softmax(logits, dim=-1)
             top = torch.topk(distribution, k=2, dim=-1)
             batch_probabilities = top.values[:, 0].tolist()
@@ -465,7 +558,7 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
         "modelVersion": calibration.get("modelVersion", args.model_dir.name),
         "device": str(device),
         "runtime": runtime_name,
-        "artifact": str(args.onnx_model or args.model_dir),
+        "artifact": str(args.coreml_model or args.onnx_model or args.model_dir),
         "fixedInputShape": [1, MAX_LENGTH],
         "overall": metrics,
         "scoredClassification": scored_metrics,
@@ -508,13 +601,21 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
         },
         "misclassifications": misclassifications,
     }
-    if reference_model is not None:
+    if reference_predictions:
-        report["pytorchParity"] = prediction_agreement(
+        parity_name = (
+            "onnxParity" if reference_onnx_session is not None else "pytorchParity"
+        )
+        parity = prediction_agreement(
             records,
             actual,
             reference_predictions,
             predicted,
         )
+        report[parity_name] = parity
+        if reference_onnx_session is not None:
+            report["gates"]["onnxLabelAgreementAtLeast99_5Percent"] = (
+                parity["scoredLabelAgreement"] >= 0.995
+            )
     write_json(args.report, report)
     return report
 
@@ -522,15 +623,31 @@ def evaluate(args: argparse.Namespace) -> dict[str, Any]:
 def build_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--model-dir", type=Path, default=DEFAULT_MODEL_DIR)
-    parser.add_argument(
+    runtime = parser.add_mutually_exclusive_group()
+    runtime.add_argument(
         "--onnx-model",
         type=Path,
         help="score a fixed-shape ONNX artifact instead of the PyTorch checkpoint",
     )
+    runtime.add_argument(
+        "--coreml-model",
+        type=Path,
+        help="score a fixed-shape Core ML package instead of the PyTorch checkpoint",
+    )
     parser.add_argument(
         "--compare-pytorch",
         action="store_true",
-        help="include label-level drift from --model-dir when scoring ONNX",
+        help="include label-level drift from --model-dir when scoring an artifact",
+    )
+    parser.add_argument(
+        "--compare-onnx",
+        type=Path,
+        help="include Core ML label-level drift from this ONNX reference",
+    )
+    parser.add_argument(
+        "--coreml-compute-units",
+        choices=("all", "cpu-only", "cpu-and-gpu", "cpu-and-ne"),
+        default="cpu-and-ne",
     )
     parser.add_argument("--calibration", type=Path, default=DEFAULT_CALIBRATION)
     parser.add_argument("--test", type=Path, default=DEFAULT_TEST)
@@ -552,6 +669,11 @@ def main(argv: Sequence[str] | None = None) -> int:
     args = parser.parse_args(argv)
     if args.batch_size <= 0 or args.latency_samples < 0:
         parser.error("batch size must be positive and latency samples non-negative")
+    if args.coreml_model is not None and args.compare_onnx is None and not args.no_gate:
+        parser.error(
+            "gated Core ML evaluation requires --compare-onnx; use --no-gate only "
+            "for diagnostics"
+        )
     try:
         report = evaluate(args)
     except (DataError, OSError, ValueError) as exc:

inspect_coreml.py

@@ -0,0 +1,153 @@
+#!/usr/bin/env python3
+"""Inspect Core ML operation placement and estimated accelerator cost share."""
+
+from __future__ import annotations
+
+import argparse
+import platform
+import sys
+from collections import Counter, defaultdict
+from pathlib import Path
+from typing import Any, Sequence
+
+from purpose_data import DataError, write_json
+
+
+def _device_category(device: Any) -> str:
+    name = type(device).__name__.lower()
+    description = str(device).lower()
+    combined = f"{name} {description}"
+    if "neural" in combined:
+        return "neuralEngine"
+    if "gpu" in combined:
+        return "gpu"
+    if "cpu" in combined:
+        return "cpu"
+    return "unknown"
+
+
+def _compute_unit(coremltools: Any, requested: str) -> Any:
+    values = {
+        "all": coremltools.ComputeUnit.ALL,
+        "cpu-only": coremltools.ComputeUnit.CPU_ONLY,
+        "cpu-and-gpu": coremltools.ComputeUnit.CPU_AND_GPU,
+        "cpu-and-ne": coremltools.ComputeUnit.CPU_AND_NE,
+    }
+    return values[requested]
+
+
+def inspect(args: argparse.Namespace) -> dict[str, Any]:
+    if not args.model.exists():
+        raise DataError(f"{args.model}: Core ML model is missing")
+    try:
+        import coremltools as ct
+    except ImportError as exc:
+        raise DataError(
+            "Core ML inspection requires requirements-coreml.txt on macOS"
+        ) from exc
+
+    compiled = ct.models.utils.compile_model(str(args.model))
+    compute_plan = ct.models.compute_plan.MLComputePlan.load_from_path(
+        path=str(compiled),
+        compute_units=_compute_unit(ct, args.compute_units),
+    )
+    program = compute_plan.model_structure.program
+    if program is None or "main" not in program.functions:
+        raise DataError("Core ML package is not an ML Program with a main function")
+    operations = list(program.functions["main"].block.operations)
+    if not operations:
+        raise DataError("Core ML compute plan contains no operations")
+
+    preferred_counts: Counter[str] = Counter()
+    preferred_costs: dict[str, float] = defaultdict(float)
+    supported_counts: Counter[str] = Counter()
+    operation_reports = []
+    operations_with_usage = 0
+    operations_with_cost = 0
+    total_cost = 0.0
+    for operation in operations:
+        usage = compute_plan.get_compute_device_usage_for_mlprogram_operation(
+            operation
+        )
+        cost = compute_plan.get_estimated_cost_for_mlprogram_operation(operation)
+        preferred = "unknown"
+        supported: list[str] = []
+        if usage is not None:
+            operations_with_usage += 1
+            preferred = _device_category(usage.preferred_compute_device)
+            preferred_counts[preferred] += 1
+            supported = sorted(
+                {_device_category(device) for device in usage.supported_compute_devices}
+            )
+            supported_counts.update(supported)
+        weight = None
+        if cost is not None:
+            operations_with_cost += 1
+            weight = float(cost.weight)
+            total_cost += weight
+            preferred_costs[preferred] += weight
+        operation_reports.append(
+            {
+                "operatorName": str(operation.operator_name),
+                "preferredDevice": preferred,
+                "supportedDevices": supported,
+                "estimatedCostWeight": weight,
+            }
+        )
+
+    ane_operations = preferred_counts["neuralEngine"]
+    ane_cost = preferred_costs["neuralEngine"]
+    report = {
+        "schemaVersion": 1,
+        "model": str(args.model),
+        "coremltoolsVersion": ct.__version__,
+        "machine": platform.machine(),
+        "macOS": platform.mac_ver()[0],
+        "computeUnits": args.compute_units,
+        "operations": len(operations),
+        "operationsWithDeviceUsage": operations_with_usage,
+        "operationsWithEstimatedCost": operations_with_cost,
+        "preferredOperationCounts": dict(sorted(preferred_counts.items())),
+        "supportedOperationCounts": dict(sorted(supported_counts.items())),
+        "preferredEstimatedCosts": dict(sorted(preferred_costs.items())),
+        "neuralEngineOperationShare": (
+            ane_operations / operations_with_usage if operations_with_usage else 0.0
+        ),
+        "neuralEngineEstimatedCostShare": (
+            ane_cost / total_cost if total_cost else 0.0
+        ),
+        "operationDetails": operation_reports,
+    }
+    write_json(args.report, report)
+    print(
+        "Core ML placement: "
+        f"ANE operations={report['neuralEngineOperationShare']:.2%} "
+        f"ANE estimated cost={report['neuralEngineEstimatedCostShare']:.2%}"
+    )
+    return report
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model", type=Path, required=True)
+    parser.add_argument("--report", type=Path, required=True)
+    parser.add_argument(
+        "--compute-units",
+        choices=("all", "cpu-only", "cpu-and-gpu", "cpu-and-ne"),
+        default="cpu-and-ne",
+    )
+    return parser
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    args = build_parser().parse_args(argv)
+    try:
+        inspect(args)
+    except (DataError, OSError, RuntimeError, ValueError) as exc:
+        print(f"error: {exc}", file=sys.stderr)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

requirements-coreml.txt

@@ -0,0 +1,2 @@
+-r requirements.txt
+coremltools==9.0

tests/test_convert_coreml.py

@@ -0,0 +1,84 @@
+import json
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+import torch
+from transformers import BertConfig, BertForSequenceClassification
+
+
+MODULE_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(MODULE_DIR))
+
+import convert_coreml
+from purpose_data import LABELS, DataError
+
+
+class FixedShapeBertForCoreMLTests(unittest.TestCase):
+    def test_conversion_forward_matches_transformers(self):
+        torch.manual_seed(7)
+        config = BertConfig(
+            vocab_size=64,
+            hidden_size=16,
+            num_hidden_layers=1,
+            num_attention_heads=4,
+            intermediate_size=32,
+            max_position_embeddings=128,
+            type_vocab_size=2,
+            hidden_dropout_prob=0.0,
+            attention_probs_dropout_prob=0.0,
+            num_labels=len(LABELS),
+        )
+        model = BertForSequenceClassification(config).eval()
+        wrapper = convert_coreml.FixedShapeBertForCoreML(model).eval()
+        input_ids = torch.randint(0, config.vocab_size, (1, 128), dtype=torch.int32)
+        attention_mask = torch.zeros((1, 128), dtype=torch.int32)
+        attention_mask[:, :83] = 1
+        token_type_ids = torch.zeros((1, 128), dtype=torch.int32)
+        token_type_ids[:, 43:83] = 1
+        with torch.inference_mode():
+            reference = model(
+                input_ids=input_ids.long(),
+                attention_mask=attention_mask.long(),
+                token_type_ids=token_type_ids.long(),
+            ).logits
+            candidate = wrapper(input_ids, attention_mask, token_type_ids)
+            torch.testing.assert_close(candidate, reference, rtol=1e-5, atol=2e-5)
+
+            traced = torch.jit.trace(
+                wrapper,
+                (input_ids, attention_mask, token_type_ids),
+                strict=True,
+            )
+            torch.testing.assert_close(
+                traced(input_ids, attention_mask, token_type_ids),
+                reference,
+                rtol=1e-5,
+                atol=2e-5,
+            )
+
+
+class CheckpointConfigTests(unittest.TestCase):
+    def test_rejects_changed_label_order(self):
+        config = {
+            "model_type": "bert",
+            "hidden_size": 384,
+            "num_hidden_layers": 6,
+            "id2label": {
+                str(index): label
+                for index, label in enumerate(reversed(LABELS))
+            },
+        }
+        with tempfile.TemporaryDirectory() as temp:
+            model_dir = Path(temp)
+            (model_dir / "config.json").write_text(
+                json.dumps(config),
+                encoding="utf-8",
+            )
+            with self.assertRaisesRegex(DataError, "label order"):
+                convert_coreml._checkpoint_config(model_dir)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_eval.py

@@ -9,6 +9,29 @@ sys.path.insert(0, str(MODULE_DIR))
 import eval as purpose_eval
 
 
+class CoreMLComputeUnitTests(unittest.TestCase):
+    class CoreMLTools:
+        class ComputeUnit:
+            ALL = "all-value"
+            CPU_ONLY = "cpu-value"
+            CPU_AND_GPU = "gpu-value"
+            CPU_AND_NE = "ne-value"
+
+    def test_maps_cli_compute_policies(self):
+        expected = {
+            "all": "all-value",
+            "cpu-only": "cpu-value",
+            "cpu-and-gpu": "gpu-value",
+            "cpu-and-ne": "ne-value",
+        }
+        for requested, value in expected.items():
+            with self.subTest(requested=requested):
+                self.assertEqual(
+                    value,
+                    purpose_eval._coreml_compute_unit(self.CoreMLTools, requested),
+                )
+
+
 class TierDriftTests(unittest.TestCase):
     def test_current_routing_matrix_bounds_every_label_pair(self):
         records = [{"prompt": f"prompt {index}"} for index in range(8 * 8)]

tests/test_inspect_coreml.py

@@ -0,0 +1,26 @@
+import sys
+import unittest
+from pathlib import Path
+
+
+MODULE_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(MODULE_DIR))
+
+import inspect_coreml
+
+
+class CoreMLDeviceCategoryTests(unittest.TestCase):
+    def test_classifies_compute_device_types(self):
+        NeuralEngineDevice = type("MLNeuralEngineComputeDevice", (), {})
+        GPUDevice = type("MLGPUComputeDevice", (), {})
+        CPUDevice = type("MLCPUComputeDevice", (), {})
+        self.assertEqual(
+            "neuralEngine",
+            inspect_coreml._device_category(NeuralEngineDevice()),
+        )
+        self.assertEqual("gpu", inspect_coreml._device_category(GPUDevice()))
+        self.assertEqual("cpu", inspect_coreml._device_category(CPUDevice()))
+
+
+if __name__ == "__main__":
+    unittest.main()