Add the A5 self-certification kit (L0-L3)

TODO.sota/65-b4-urdu-sota.md

@@ -8,3 +8,24 @@ negative: correlated arms transfer nothing); morph/lexicon features;
 VRR. G2P — owner selects licensed corpus (option drafts in WO50);
 retrain; relabel artifacts non-research-only.
 Acceptance: external protocol numbers; license gate green on G2P.
+
+
+## CLE corpus acquisition (2026-10-29, owner-reported)
+
+The owner is waiting for an Urdu corpus from CLE (Center for Language
+Engineering, UET Lahore). Owner-side steps enumerated for the record:
+
+1. Written license terms from CLE covering (a) research vs commercial use,
+   (b) redistribution of DERIVED WEIGHTS (the clause our gate blocks on),
+   (c) required attribution wording, (d) corpus version and checksum.
+2. The license_review record filed by the owner (law: UNVERIFIED/LicenseRef-*
+   blocks publish; waivers are license_review records). Until then urd-g2p
+   stays GATED-1 ready-pending-license (weights done, 0.16 CER, parity
+   cleared, fixtures built — 53-ship-set).
+3. Data channel decision: GitHub Releases re-release if the license permits
+   corpus redistribution; gated store + weights-only otherwise.
+4. Fallback if CLE stalls: ship urd-g2p research-only, or retrain on an open
+   Urdu dictionary (pre-recorded ship-set options).
+5. On license-green: retrain urd-g2p on the licensed corpus, re-run WO51
+   gates, flip card/index from research-only, ship. Separately: whether the
+   CLE corpus also feeds the urd-diac 3.57 -> < 3 push is its own WO.

tools/certify.py

@@ -0,0 +1,77 @@
+"""A5 — the self-certification kit: L0–L3 for any model id.
+
+L0 parse       the id resolves through the index (sha256 verified)
+L1 fixtures    the model ships frozen conformance vectors
+L2 byte-exact  the reference loader replays the fixtures exactly
+L3 parity      torch vs shipped bytes agree (>=500 rows, shipped mode)
+
+Usage:
+    python tools/certify.py <model-id> [--index <index-url>]
+
+Exit code 0 = certified at the highest level reached; the JSON report
+states the level. A third party self-certifies with this alone.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+REPO = Path(__file__).parent.parent
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("model_id")
+    ap.add_argument("--index", default=None)
+    args = ap.parse_args()
+
+    report = {"model": args.model_id, "levels": {}}
+    sys.path.insert(0, str(REPO / "runtime" / "src"))
+    from interscript_ml import Model  # noqa: E402
+
+    # L0 — resolution + integrity
+    try:
+        m = Model.load(args.model_id, index_url=args.index)
+        report["levels"]["L0"] = {"pass": True, "note": "resolved + sha256 verified"}
+    except Exception as e:
+        report["levels"]["L0"] = {"pass": False, "error": repr(e)[:200]}
+        print(json.dumps(report, indent=2))
+        sys.exit(1)
+
+    # L1 — frozen fixtures present
+    vecs = sorted((REPO / "conformance").glob(
+        f"{args.model_id.split('-')[0]}*{args.model_id.split('-')[1]}*-vectors.json"))
+    if not vecs:
+        report["levels"]["L1"] = {"pass": False,
+                                  "note": "no frozen vectors in conformance/"}
+        print(json.dumps(report, indent=2))
+        sys.exit(2)
+    vectors = json.loads(vecs[0].read_text(encoding="utf-8"))
+    report["levels"]["L1"] = {"pass": True, "vectors": len(vectors["vectors"]),
+                              "file": vecs[0].name}
+
+    # L2 — byte-exact replay through the reference loader
+    bad = 0
+    for row in vectors["vectors"]:
+        if m.translate(row["src"]) != row["out"]:
+            bad += 1
+    report["levels"]["L2"] = {"pass": bad == 0, "mismatches": bad,
+                              "note": "reference-loader byte-exact replay"}
+
+    # L3 — parity: the caller runs the torch-vs-artifact harness
+    # (parity_urd_imf_hf.py lineage, >= 500 rows) and appends the verdict;
+    # a certified-at-L2 model without an L3 record states so.
+    report["levels"]["L3"] = {"pass": None,
+                              "note": "run tools/… parity harness >=500 rows "
+                                      "and append {pass: bool}"}
+
+    print(json.dumps(report, indent=2))
+    ok = report["levels"]["L2"]["pass"]
+    sys.exit(0 if ok else 3)
+
+
+if __name__ == "__main__":
+    main()