Run Engraph research evidence gates in CI

.github/workflows/research-integrity.yml

@@ -0,0 +1,71 @@
+name: Engraph research integrity
+
+on:
+  pull_request:
+    branches:
+      - feat/engraph-p3-perf
+    paths:
+      - models/**
+      - research-evidence/**
+      - scripts/research_receipt.py
+      - scripts/test_research_receipt.py
+      - scripts/test_verify_*.py
+      - scripts/verify_*.py
+      - examples/p3_actor_parity.rs
+      - src/llm.rs
+      - src/serve.rs
+      - Cargo.lock
+      - target/release/examples/p3_actor_parity
+      - .github/workflows/research-integrity.yml
+  push:
+    branches:
+      - feat/engraph-p3-perf
+    paths:
+      - models/**
+      - research-evidence/**
+      - scripts/research_receipt.py
+      - scripts/test_research_receipt.py
+      - scripts/test_verify_*.py
+      - scripts/verify_*.py
+      - examples/p3_actor_parity.rs
+      - src/llm.rs
+      - src/serve.rs
+      - Cargo.lock
+      - target/release/examples/p3_actor_parity
+      - .github/workflows/research-integrity.yml
+
+permissions:
+  contents: read
+
+concurrency:
+  group: engraph-research-integrity-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  verify:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    env:
+      PYTHONDONTWRITEBYTECODE: "1"
+      PYTHONPATH: scripts
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
+        with:
+          python-version: "3.12"
+      - name: Verify shared manifest contract
+        run: python3 research-evidence/contracts/v2/verify_contract.py
+      - name: Verify model provenance
+        run: python3 scripts/verify_model_provenance.py
+      - name: Verify claim durability
+        run: python3 scripts/verify_claim_durability.py
+      - name: Verify authentic P3 correctness receipt
+        run: python3 scripts/research_receipt.py verify-record research-evidence/runs/p3-actor-parity-v2-20260717-a2.receipt.json
+      - name: Audit the authentic P3 attempt
+        run: python3 scripts/research_receipt.py audit-plans research-evidence/runs --attempt-id 076b69d2-3a1f-40d6-9555-9c7cc6c69302 --record-only
+      - name: Run integrity tests
+        run: python3 -m unittest discover -s scripts -p "test_*.py" -q
+      - name: Run integrity tests without assertions
+        run: python3 -O -m unittest discover -s scripts -p "test_*.py" -q

research-evidence/README.md

@@ -86,6 +86,22 @@ complete fail-closed envelope and globally unique event identity. Use
 `--attempt-id <id>` to evaluate one new attempt without allowing an unrelated
 historical failure to become either its pass or its failure.
 
+Clean clones and CI can validate the committed receipt envelope and repository
+artifacts without claiming custody of unavailable model or tokenizer bytes:
+
+```sh
+python3 scripts/research_receipt.py verify-record \
+  research-evidence/runs/p3-actor-parity-v2-20260717-a2.receipt.json
+python3 scripts/research_receipt.py audit-plans research-evidence/runs \
+  --attempt-id 076b69d2-3a1f-40d6-9555-9c7cc6c69302 \
+  --record-only
+```
+
+Record-only verification filters only explicitly classified unavailable
+external artifacts. Malformed, changed, unsafe, or hash-mismatched evidence
+still fails. The strict `verify` and default `audit-plans` commands remain the
+custody checks and must not be replaced by record-only results.
+
 Receipts contain a canonical self-digest for accidental corruption detection.
 The digest is stored beside editable metadata and is not authentication. They
 are not signed, externally timestamped, or stored on immutable media, so

scripts/research_receipt.py

@@ -1422,6 +1422,15 @@ def verify_receipt(path: Path) -> list[str]:
     return errors
 
 
+def verify_receipt_record(path: Path) -> list[str]:
+    """Verify a committed receipt without claiming external-byte custody."""
+    return [
+        error
+        for error in verify_receipt(path)
+        if error not in TERMINAL_UNAVAILABLE_ERRORS
+    ]
+
+
 def _write_exclusive_json(path: Path, value: dict[str, Any]) -> None:
     path = path.resolve()
     path.parent.mkdir(parents=True, exist_ok=True)
@@ -1586,6 +1595,7 @@ def incomplete_attempts(
     directory: Path,
     *,
     attempt_id_filter: str | None = None,
+    record_only: bool = False,
 ) -> list[dict[str, Any]]:
     resolved = directory.resolve()
     try:
@@ -1619,7 +1629,12 @@ def incomplete_attempts(
                     errors.append("planned event unexpectedly has a parent")
             planned.setdefault(attempt_id, []).append((path, errors))
         elif value.get("record_type") == "attempt_result":
-            terminal.setdefault(attempt_id, []).append((path, verify_receipt(path)))
+            receipt_errors = (
+                verify_receipt_record(path)
+                if record_only
+                else verify_receipt(path)
+            )
+            terminal.setdefault(attempt_id, []).append((path, receipt_errors))
             terminal_plan_required[attempt_id] = (
                 terminal_plan_required.get(attempt_id, False)
                 or value.get("execution_provenance") == "tool_executed"
@@ -1801,9 +1816,16 @@ def common(run_parser: argparse.ArgumentParser) -> None:
     run.add_argument("measured_command", nargs=argparse.REMAINDER)
     verify = sub.add_parser("verify")
     verify.add_argument("receipt", type=Path)
+    verify_record = sub.add_parser("verify-record")
+    verify_record.add_argument("receipt", type=Path)
     audit = sub.add_parser("audit-plans")
     audit.add_argument("directory", type=Path)
     audit.add_argument("--attempt-id")
+    audit.add_argument(
+        "--record-only",
+        action="store_true",
+        help="do not claim custody of unavailable external artifacts",
+    )
     args = parser.parse_args()
     if args.action == "create":
         print(json.dumps(create_receipt(args), sort_keys=True))
@@ -1814,10 +1836,16 @@ def common(run_parser: argparse.ArgumentParser) -> None:
         if errors:
             raise SystemExit("\n".join(errors))
         print("OK")
+    elif args.action == "verify-record":
+        errors = verify_receipt_record(args.receipt)
+        if errors:
+            raise SystemExit("\n".join(errors))
+        print("OK: receipt record verified; external byte custody not claimed")
     else:
         issues = incomplete_attempts(
             args.directory,
             attempt_id_filter=args.attempt_id,
+            record_only=args.record_only,
         )
         print(json.dumps(issues, indent=2, sort_keys=True))
         if issues:

scripts/test_research_receipt.py

@@ -14,6 +14,22 @@
 
 
 class ReceiptTests(unittest.TestCase):
+    def test_committed_a2_record_is_clone_verifiable(self) -> None:
+        receipt = (
+            rr.ROOT
+            / "research-evidence/runs/"
+            "p3-actor-parity-v2-20260717-a2.receipt.json"
+        )
+        self.assertEqual(rr.verify_receipt_record(receipt), [])
+        self.assertEqual(
+            rr.incomplete_attempts(
+                rr.ROOT / "research-evidence/runs",
+                attempt_id_filter="076b69d2-3a1f-40d6-9555-9c7cc6c69302",
+                record_only=True,
+            ),
+            [],
+        )
+
     def test_v2_verifier_rejects_rights_and_artifact_overclaims(self) -> None:
         representatives = json.loads(
             (
@@ -101,8 +117,9 @@ def test_command_evidence_rejects_secrets_without_false_tokenizer_match(
             rr._safe_command_argv(["tool", "--tokenizer", "embedded"]),
             ["tool", "--tokenizer", "embedded"],
         )
+        home_input = Path.home() / "private" / "file"
         self.assertEqual(
-            rr._safe_command_argv(["tool", "--input=/Users/d/private/file"])[1],
+            rr._safe_command_argv(["tool", f"--input={home_input}"])[1],
             "--input=$HOME/private/file",
         )
         with self.assertRaisesRegex(ValueError, "secret-bearing"):