release: urd-diac-base-1.0

cards/urd-diac-base-1.0.md

@@ -1,9 +1,28 @@
 # urd-diac-base-1.0
 
 Urdu diacritization, base tier (run-066): byt5-base seq2seq on the
-full 573,130-pair urdu-diacrit corpus. IMF v1 fp32.
+full 573,130-pair urdu-diacrit corpus. IMF v1 **fp16**.
 
-## Metrics (full 11,940-row test split, raw char CER)
+## Metrics (full 11,940-row test split, raw char CER, row-mean)
 
 CER **3.57** (urd-diac-1.0 small tier: 3.77 baseline) — gate PASSED.
 Word accuracy 69.8%.
+
+## Artifact parity (WO51 contract)
+
+Single-row greedy decode, the runtime's actual decode mode: ONNX
+artifact CER **3.64** — identical to torch on **500/500 rows**
+(agreement 1.0; three-way bf16/fp32/ORT agreement 1.0). Batched
+decode without an attention mask is outside the artifact contract
+(CONFORMANCE.md).
+
+## Conformance
+
+The shipped runtime reproduces all 12 frozen vectors byte-exactly
+(`conformance/urd-diac-base-vectors.json`).
+
+## Data / license
+
+Weights depend on the original training-data license (the urdu-diacrit
+corpus; provenance in `corpora/PROVENANCE.md`, owner-approved ship
+decision 2026-10-18). Code and weights: BSD-3-Clause.

docs/RESULTS.md

@@ -2348,3 +2348,12 @@ succeeds, then the p047 remainder (~4h) + p029 + consolidation run
 death-proof. The mount (/ckpt = the bucket) may itself be durable
 for properly-closed files — untested; the explicit sync removes the
 doubt.
+
+## run-066 urd-diac-base gate pass and artifact parity (2026-10-19)
+
+Gate: full 11,940-row test split, raw char CER, row-mean — **3.57**
+(beat urd-diac-1.0 small tier 3.77), word accuracy 69.8%. Artifact
+parity under the WO51 single-row contract: ONNX fp16 CER 3.64,
+500/500 rows identical to torch, three-way bf16/fp32/ORT agreement
+1.0. Shipped runtime reproduces 12/12 frozen conformance vectors
+byte-exactly. Data: 573,130 pairs (urd-dia registry, sha256-pinned).

models.yaml

@@ -472,3 +472,15 @@ models:
       - {name: wer_wikinews2024_multiref, value: 8.85, protocol: "greedy decode, first-alternate protocol, WikiNews-2024 multiref (WO35)", source: interscript/interscript-models docs/RESULTS.md}
       - {name: der_total_greedy, value: 2.84, protocol: "greedy decode, first-alternate protocol, WikiNews-2024 multiref (WO35)", source: interscript/interscript-models docs/RESULTS.md}
     license: BSD-3-Clause
+  urd-diac-base-1.0:
+    task: diacritization
+    scripts: [Arab, Arab]
+    precision: fp16
+    filename: urd-diac-base-1.0-fp16.zip
+    url: https://github.com/interscript/interscript-models/releases/download/urd-diac-base-1.0/urd-diac-base-1.0-fp16.zip
+    sha256: 8058220d72df76b31260358070499c2cdb504f030fa10259b48aa05c7caf5711
+    size: 1347487633
+    metrics:
+      - {name: cer, value: 3.57, source: interscript/interscript-models docs/RESULTS.md#run-066-urd-diac-base-gate-pass-and-artifact-parity-2026-10-19}
+    parity: {samples: 500, cer_delta: 0.07}
+    license: BSD-3-Clause

models/urd-diac-base/urd-diac-base-1.0.metadata.yaml

@@ -0,0 +1,23 @@
+format: imf-v1
+id: urd-diac-base-1.0
+task: diacritization
+source_script: Arab
+target: Arab
+tokenizer: bytes
+opset: 14
+decoder: plain
+precision: fp16
+license: BSD-3-Clause
+trained_from: byt5-base, 573130 urdu-diacrit pairs, run-066 (TODO.sota WO41)
+metrics:
+- name: cer
+  value: 3.57
+  source: interscript/interscript-models docs/RESULTS.md#run-066-urd-diac-base-gate-pass-and-artifact-parity-2026-10-19
+  protocol: greedy decode, full 11940-row test split, row-mean CER (WO51 single-row
+    artifact contract)
+sha256:
+  encoder.onnx: ac2ccda95bdfeea5536a81099a3f70241b44b05a8d6bfb05f59ec21c086ff88e
+  decoder.onnx: 13a140f8a85794fc93537b88c77493483d414aad1c5d53316a5e4c86dd930c10
+parity:
+  samples: 500
+  cer_delta: 0.07

tests/test_pipeline_plan_gate.py

@@ -91,8 +91,8 @@ def test_parity_gate_runtime_mode() -> None:
 
 def test_license_gate_blocks_publish(tmp_path: Path) -> None:
     # the shipped urd-dia is owner-waived (2026-10-18) and must PASS now
-    waived = evaluate(recipe := load_recipe(RECIPES / "urd-diac.yaml"),
-                      CORPORA, metrics={"cer": 0.0352}, parity=None)
+    recipe = load_recipe(RECIPES / "urd-diac.yaml")
+    waived = evaluate(recipe, CORPORA, metrics={"cer": 0.0352}, parity=None)
     assert [g for g in waived if g.name == "license"][0].passed
     # an UNWAIVED restricted source still blocks — synthetic registry
     import yaml as _yaml