diff --git a/ENERGY_AND_RESIDENCY.md b/ENERGY_AND_RESIDENCY.md
index 4dbc1d0..6a60993 100644
--- a/ENERGY_AND_RESIDENCY.md
+++ b/ENERGY_AND_RESIDENCY.md
@@ -19,7 +19,8 @@ placement.
 
 ## macOS
 
-- Generate the ML Program with `convert_coreml.py`, then run the gated `eval.py
+- Generate the float ML Program with `convert_coreml.py`, calibrate the W8A8 candidate
+  with `quantize_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
@@ -36,6 +37,13 @@ placement.
   per inference than CPU-only. On AC, a `.all` GPU retry must separately meet its declared
   GPU operation-share floor before it can activate a deep model.
 
+The first float16 candidate is a diagnostic baseline only: it achieved 94.98% scored
+accuracy and 97.97% agreement with the accepted ONNX artifact, so it fails the rollout
+accuracy and parity gates despite 1.51 ms p95 latency. Its compute plan preferred the ANE
+for 150/165 operations with placement information (90.91%) and 59.71% of estimated cost.
+Do not spend energy-measurement time on that rejected package; repeat placement, latency,
+and energy measurements on the first accuracy-qualified W8A8 package.
+
 ## Windows
 
 - Record the Windows ML execution provider and assigned device after AOT compilation.
diff --git a/README.md b/README.md
index 36a77af..fdd4420 100644
--- a/README.md
+++ b/README.md
@@ -217,6 +217,28 @@ ml/purpose-classifier/venv/bin/python ml/purpose-classifier/convert_coreml.py \
   --overwrite-output
 ```
 
+The direct float16 package is the conversion baseline, not the accepted Apple artifact.
+On the first physical Apple-Silicon run it scored 94.98% (890/937), two correct decisions
+behind the accepted ONNX graph, with 97.97% scorable label agreement. Calibrate a
+Core ML-native W8A8 candidate with the same deterministic 256-record sample and QDQ policy
+as the ONNX exporter:
+
+```bash
+ml/purpose-classifier/venv/bin/python ml/purpose-classifier/quantize_coreml.py \
+  --model \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/purpose-lite-v1-fp16.mlpackage \
+  --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-w8a8.mlpackage \
+  --overwrite-output
+```
+
+Activation calibration is grouped to keep temporary Core ML packages bounded and prints
+progress while it runs. The candidate uses per-tensor asymmetric uint8 activations,
+per-channel symmetric int8 linear weights, and per-tensor asymmetric uint8 embedding
+weights. It fails the command if the resulting package exceeds 25 MiB.
+
 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:
@@ -228,7 +250,7 @@ ml/purpose-classifier/venv/bin/python ml/purpose-classifier/eval.py \
   --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 \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/purpose-lite-v1-w8a8.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 \
@@ -243,7 +265,7 @@ 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 \
+    ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/purpose-lite-v1-w8a8.mlpackage \
   --compute-units cpu-and-ne \
   --report \
     ml/purpose-classifier/outputs/purpose-lite-v1-distilled-qat-mlx-4e/coreml/ane-compute-plan.json
diff --git a/quantize_coreml.py b/quantize_coreml.py
new file mode 100644
index 0000000..3a6008e
--- /dev/null
+++ b/quantize_coreml.py
@@ -0,0 +1,261 @@
+#!/usr/bin/env python3
+"""Calibrate a Core ML W8A8 candidate from the selected float16 ML Program."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import shutil
+import sys
+from collections import Counter
+from pathlib import Path
+from typing import Any, Sequence
+
+import numpy as np
+
+from convert_coreml import COREMLTOOLS_VERSION, _package_manifest
+from export import stratified_calibration_sample
+from purpose_data import DataError, load_jsonl, prompt_hash, write_json
+from train import encode_fixed_shape
+
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+CANDIDATE_DIR = (
+    SCRIPT_DIR / "outputs" / "purpose-lite-v1-distilled-qat-mlx-4e"
+)
+DEFAULT_MODEL = CANDIDATE_DIR / "coreml" / "purpose-lite-v1-fp16.mlpackage"
+DEFAULT_MODEL_DIR = CANDIDATE_DIR / "model"
+DEFAULT_VALIDATION = SCRIPT_DIR / ".artifacts" / "dataset-v1" / "validation.jsonl"
+DEFAULT_OUTPUT = CANDIDATE_DIR / "coreml" / "purpose-lite-v1-w8a8.mlpackage"
+SHIPPING_BUDGET_BYTES = 25 * 1024 * 1024
+
+
+def _tree_sha256(package: Path) -> str:
+    digest = hashlib.sha256()
+    for path in sorted(item for item in package.rglob("*") if item.is_file()):
+        relative = str(path.relative_to(package)).encode("utf-8")
+        digest.update(relative)
+        digest.update(b"\0")
+        with path.open("rb") as handle:
+            for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+                digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _optimization_configs(optimize: Any) -> tuple[Any, Any]:
+    """Return the Core ML analogue of the accepted ONNX QDQ policy."""
+
+    activation = optimize.coreml.OpLinearQuantizerConfig(
+        mode="linear",
+        dtype=np.uint8,
+        granularity="per_tensor",
+    )
+    activation_config = optimize.coreml.OptimizationConfig(
+        global_config=activation,
+    )
+
+    linear_weight = optimize.coreml.OpLinearQuantizerConfig(
+        mode="linear_symmetric",
+        dtype=np.int8,
+        granularity="per_channel",
+        weight_threshold=2048,
+    )
+    embedding_weight = optimize.coreml.OpLinearQuantizerConfig(
+        mode="linear",
+        dtype=np.uint8,
+        granularity="per_tensor",
+        weight_threshold=2048,
+    )
+    weight_config = optimize.coreml.OptimizationConfig(
+        op_type_configs={
+            "gather": embedding_weight,
+            "linear": linear_weight,
+            "matmul": linear_weight,
+        }
+    )
+    return activation_config, weight_config
+
+
+def _validate_args(args: argparse.Namespace) -> None:
+    if not args.model.is_dir():
+        raise DataError(f"{args.model}: source Core ML package is missing")
+    if not args.model_dir.is_dir():
+        raise DataError(f"{args.model_dir}: tokenizer directory is missing")
+    if args.output.suffix != ".mlpackage":
+        raise DataError("Core ML output must end in .mlpackage")
+    try:
+        same_output = args.model.resolve() == args.output.resolve()
+    except OSError as exc:
+        raise DataError(f"cannot resolve Core ML package paths: {exc}") from exc
+    if same_output:
+        raise DataError("W8A8 output must not overwrite its float16 source package")
+    if args.output.exists() and not args.overwrite_output:
+        raise DataError(
+            f"{args.output}: output exists; pass --overwrite-output intentionally"
+        )
+
+
+def quantize(args: argparse.Namespace) -> dict[str, Any]:
+    _validate_args(args)
+    try:
+        import coremltools as ct
+        import coremltools.optimize as cto
+        import torch
+        from transformers import AutoTokenizer
+    except ImportError as exc:
+        raise DataError(
+            "Core ML quantization requires requirements-coreml.txt on macOS"
+        ) from exc
+    if ct.__version__ != COREMLTOOLS_VERSION:
+        raise DataError(
+            f"expected coremltools {COREMLTOOLS_VERSION}, found {ct.__version__}"
+        )
+
+    validation = load_jsonl(args.validation)
+    calibration = stratified_calibration_sample(
+        validation,
+        args.calibration_records,
+        seed=args.calibration_seed,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.model_dir,
+        local_files_only=True,
+    )
+    sample_data = []
+    for record in calibration:
+        encoded = encode_fixed_shape(tokenizer, [record["prompt"]], torch)
+        sample_data.append(
+            {
+                name: value.numpy().astype(np.int32, copy=False)
+                for name, value in encoded.items()
+            }
+        )
+
+    print(
+        f"Core ML activation calibration: {len(sample_data)} records",
+        flush=True,
+    )
+    source_model = ct.models.MLModel(
+        str(args.model),
+        compute_units=ct.ComputeUnit.CPU_ONLY,
+    )
+    input_names = {item.name for item in source_model.get_spec().description.input}
+    expected_inputs = {"input_ids", "attention_mask", "token_type_ids"}
+    if input_names != expected_inputs:
+        raise DataError(
+            f"Core ML inputs changed: expected {sorted(expected_inputs)}, "
+            f"got {sorted(input_names)}"
+        )
+    activation_config, weight_config = _optimization_configs(cto)
+    activation_quantized = cto.coreml.linear_quantize_activations(
+        source_model,
+        activation_config,
+        sample_data,
+        calibration_op_group_size=args.calibration_op_group_size,
+    )
+    print("Core ML weight quantization: W8", flush=True)
+    quantized = cto.coreml.linear_quantize_weights(
+        activation_quantized,
+        weight_config,
+    )
+    quantized.user_defined_metadata["com.nucleic.model.quantization"] = "W8A8"
+    quantized.user_defined_metadata["com.nucleic.model.quantizationCalibration"] = (
+        f"stratified:{args.calibration_records}:seed={args.calibration_seed}"
+    )
+
+    if args.output.exists():
+        if args.output.is_dir():
+            shutil.rmtree(args.output)
+        else:
+            args.output.unlink()
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    quantized.save(str(args.output))
+
+    manifest = _package_manifest(args.output)
+    manifest.update(
+        {
+            "sourcePackage": str(args.model),
+            "sourcePackageSha256": _tree_sha256(args.model),
+            "coremltoolsVersion": ct.__version__,
+            "quantization": {
+                "name": "W8A8",
+                "activations": "per-tensor asymmetric uint8",
+                "linearWeights": "per-channel symmetric int8",
+                "embeddingWeights": "per-tensor asymmetric uint8",
+            },
+            "calibrationRecords": len(calibration),
+            "calibrationSeed": args.calibration_seed,
+            "calibrationOpGroupSize": args.calibration_op_group_size,
+            "calibrationSample": {
+                "strategy": "stratified by purpose, slice, and primary language",
+                "purposeCounts": dict(
+                    sorted(Counter(item["purpose"] for item in calibration).items())
+                ),
+                "sliceCounts": dict(
+                    sorted(Counter(item["slice"] for item in calibration).items())
+                ),
+                "promptHashes": sorted(
+                    prompt_hash(item["prompt"]) for item in calibration
+                ),
+            },
+            "shippingBudgetBytes": args.shipping_budget_bytes,
+            "shippingBudgetPassed": (
+                manifest["bytes"] <= args.shipping_budget_bytes
+            ),
+        }
+    )
+    manifest_path = args.output.with_name(f"{args.output.stem}-manifest.json")
+    write_json(manifest_path, manifest)
+    print(
+        f"Core ML W8A8 package: {args.output} ({manifest['bytes']} bytes)",
+        flush=True,
+    )
+    if not manifest["shippingBudgetPassed"]:
+        raise DataError(
+            f"{args.output}: {manifest['bytes']} bytes exceeds the "
+            f"{args.shipping_budget_bytes}-byte shipping budget"
+        )
+    return manifest
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model", type=Path, default=DEFAULT_MODEL)
+    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("--calibration-records", type=int, default=256)
+    parser.add_argument("--calibration-seed", type=int, default=20260730)
+    parser.add_argument("--calibration-op-group-size", type=int, default=32)
+    parser.add_argument(
+        "--shipping-budget-bytes",
+        type=int,
+        default=SHIPPING_BUDGET_BYTES,
+    )
+    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.calibration_records <= 0
+        or args.calibration_op_group_size == 0
+        or args.calibration_op_group_size < -1
+        or args.shipping_budget_bytes <= 0
+    ):
+        parser.error(
+            "calibration records/budget must be positive and op group size must be "
+            "-1 or positive"
+        )
+    try:
+        quantize(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())
diff --git a/tests/test_quantize_coreml.py b/tests/test_quantize_coreml.py
new file mode 100644
index 0000000..aab057f
--- /dev/null
+++ b/tests/test_quantize_coreml.py
@@ -0,0 +1,79 @@
+import argparse
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+import numpy as np
+
+
+MODULE_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(MODULE_DIR))
+
+import quantize_coreml
+from purpose_data import DataError
+
+
+class FakeOpLinearQuantizerConfig:
+    def __init__(self, **values):
+        self.values = values
+
+
+class FakeOptimizationConfig:
+    def __init__(self, *, global_config=None, op_type_configs=None):
+        self.global_config = global_config
+        self.op_type_configs = op_type_configs or {}
+
+
+class FakeCoreML:
+    OpLinearQuantizerConfig = FakeOpLinearQuantizerConfig
+    OptimizationConfig = FakeOptimizationConfig
+
+
+class FakeOptimize:
+    coreml = FakeCoreML
+
+
+class CoreMLQuantizationConfigTests(unittest.TestCase):
+    def test_matches_accepted_qdq_policy(self):
+        activation, weights = quantize_coreml._optimization_configs(FakeOptimize)
+        self.assertEqual("linear", activation.global_config.values["mode"])
+        self.assertIs(np.uint8, activation.global_config.values["dtype"])
+        self.assertEqual(
+            "per_tensor",
+            activation.global_config.values["granularity"],
+        )
+
+        linear = weights.op_type_configs["linear"].values
+        self.assertEqual("linear_symmetric", linear["mode"])
+        self.assertIs(np.int8, linear["dtype"])
+        self.assertEqual("per_channel", linear["granularity"])
+        self.assertIs(
+            weights.op_type_configs["linear"],
+            weights.op_type_configs["matmul"],
+        )
+
+        embedding = weights.op_type_configs["gather"].values
+        self.assertEqual("linear", embedding["mode"])
+        self.assertIs(np.uint8, embedding["dtype"])
+        self.assertEqual("per_tensor", embedding["granularity"])
+
+    def test_rejects_overwriting_source_package(self):
+        with tempfile.TemporaryDirectory() as temp:
+            root = Path(temp)
+            package = root / "model.mlpackage"
+            package.mkdir()
+            model_dir = root / "model"
+            model_dir.mkdir()
+            args = argparse.Namespace(
+                model=package,
+                model_dir=model_dir,
+                output=package,
+                overwrite_output=True,
+            )
+            with self.assertRaisesRegex(DataError, "must not overwrite"):
+                quantize_coreml._validate_args(args)
+
+
+if __name__ == "__main__":
+    unittest.main()