Bind native decision scoring and calibration provenance

.github/workflows/ci.yml

@@ -336,7 +336,7 @@ jobs:
       - name: Verify specialist references without model downloads
         run: |
           pip install pytest safetensors pyte onnx==1.23.2 onnxruntime==1.31.0 onnxscript==0.7.2 'transformers>=4.51,<5' 'timestral @ git+https://github.com/mistralai/timestral.git@a534bdbd244d64f3d9d743f46c7bf6fc330ed4ff'
-          python -m pytest tests/test_specialist_labs.py tests/test_specialist_workflow.py tests/test_specialist_terminal_portability.py tests/test_specialist_native_export.py tests/test_specialist_forecast_portability.py tests/test_specialist_decision_portability.py -q
+          python -m pytest tests/test_specialist_labs.py tests/test_specialist_workflow.py tests/test_specialist_terminal_portability.py tests/test_specialist_native_export.py tests/test_specialist_native_scoring.py tests/test_specialist_forecast_portability.py tests/test_specialist_decision_portability.py -q
 
       - name: Upload coverage
         uses: actions/upload-artifact@v7.0.1

docs/factory/specialist-labs.md

@@ -335,6 +335,41 @@ not numerical or calibration parity. Refit temperature and rerun primary and
 regression gates before using native confidence. No model is promoted by this
 conversion.
 
+For a source-bound native score run, re-export the adapter with training
+identities retained, then use the existing local executable and model files:
+
+```sh
+uv run --project scripts/specialists python scripts/specialists/decision.py score-native --binary <existing-native-cli> --base <local-base> --adapter <new-adapter.lora> --source-revision <declared-full-commit> --data <frozen-split.jsonl> --seconds 60 --out <new-native-scores.jsonl>
+```
+
+The scorer observes the executable, base/tokenizer, adapter and sidecar hashes,
+and binds tokenizer versions and the exact prompt implementation into a distinct
+native model identity. The source revision is operator-declared, not verified
+build provenance. Every exact-token response must match its request. Changed
+files, unknown adapter training provenance, malformed output and expired budgets
+fail closed. Native adapter sidecars exported before training identities were
+preserved require re-export from their original source adapter.
+Merged bases retain their own training exclusions too. Standalone tokenizer
+files and the resolved chat template are part of this identity. Checksums run at
+the start and before final output; per-row file fingerprints detect intervening
+changes. Every row cold-loads the native model in a separate bounded process.
+The total time budget includes those loads; failure produces no final score
+file. Use a small split and retain failed-run diagnostics separately.
+All rows are checked against the native 256-token cap before any model process.
+Metal libraries, bundles and dynamic libraries beside the executable or in its
+packaged app's `Contents/Resources` and `Contents/Frameworks` are hashed too.
+Missing Metal libraries block scoring. This does not attest external system
+libraries or drivers. Exports with unknown
+training provenance remain loadable diagnostic artifacts; re-exporting such a
+source cannot reconstruct its training identities or qualify confidence.
+
+Use `decision.py calibrate` on native calibration scores only, with train,
+primary and regression scores supplied through `--exclude`; score those splits
+separately. Existing calibration and eval leakage checks remain authoritative.
+MLX and browser calibration cannot qualify this different runtime identity.
+The subprocess fixtures test the scoring boundary; actual native runs and a
+frontier-qualified frozen primary/regression comparison remain separate proof.
+
 ## Portable Qwen decisions
 
 Use a local tied-embedding Qwen3 MLX affine q4/group64 base and, optionally,

scripts/specialists/decision.py

@@ -535,6 +535,14 @@ def argument_parser():
     child.add_argument("--base", required=True)
     child.add_argument("--adapter", required=True)
     child.add_argument("--out", required=True)
+    child = sub.add_parser("score-native")
+    child.add_argument("--binary", required=True)
+    child.add_argument("--base", required=True)
+    child.add_argument("--source-revision", required=True)
+    child.add_argument("--adapter")
+    child.add_argument("--data", required=True)
+    child.add_argument("--out", required=True)
+    child.add_argument("--seconds", type=float, default=60)
     child = sub.add_parser("export-browser")
     child.add_argument("--base", required=True)
     child.add_argument("--adapter")
@@ -572,12 +580,32 @@ def run_model(args):
         print(json.dumps(budget.measurements()))
 
 
-def main():
-    args = argument_parser().parse_args()
+def run_native(args):
     if args.command == "export-native":
         from decision_native import export
 
         print(json.dumps(export(args.base, args.adapter, args.out)))
+        return
+    from decision_native_score import NativeDecisionModel
+
+    budget = Budget(args.seconds)
+    rows = read_rows(args.data)
+    if Path(args.out).exists():
+        raise FileExistsError(args.out)
+    model = NativeDecisionModel(
+        args.binary, args.base, args.source_revision, args.adapter, budget
+    )
+    model.preflight(rows)
+    scores = [model.score(row) for row in rows]
+    model.finish()
+    write_rows(args.out, scores)
+    print(json.dumps(budget.measurements()))
+
+
+def main():
+    args = argument_parser().parse_args()
+    if args.command in ("score-native", "export-native"):
+        run_native(args)
     elif args.command == "export-browser":
         from decision_browser import export
 

scripts/specialists/decision_native.py

@@ -136,6 +136,7 @@ def export(base, adapter, out):
         "base_identity": metadata["base_identity"],
         "source_adapter_sha256": metadata["adapter_sha256"],
         "native_adapter_sha256": file_hash(out),
+        "train_identities": metadata.get("train_identities"),
         "entries": len(entries),
         "quality": "unqualified; recalibrate after runtime conversion",
     }