Merge pull request #157 from interscript/ara-21-readme

docs: ara-diac-small-2.1 README + registry pointer

Author: ronaldtse

models/ara-diac-small/ara-diac-small-2.1.README.md

@@ -0,0 +1,29 @@
+# ara-diac-small-2.1
+
+Arabic diacritization (haraqat restoration). Client-tier ByT5-small
+student — the 2.0 recipe (r7 canonical teacher, Muon) held for
+**double the epochs** (6 instead of 3, 39,018 steps): the epochs lever
+moves the rung 4.82 → **4.5701** full-set windowed DER-CE.
+
+The paired-bootstrap delta vs the teacher (same harness, seed 42,
+1,000 resamples) is [1.91, 2.35] — disjoint from the 2.0 rung's
+[2.36, 2.82]: the improvement is statistically real, and the frontier
+separations hold end to end (1.0 8.26 → lite 5.78 → 2.0 4.82 → 2.1
+4.57 → teacher 2.29).
+
+The epochs result reframes the residual: doubling training buys
+−0.25pp, so undertraining is not the dominant term — the E2/E3 domain
+attribution substantially stands. The strict teacher+0.5pp gate
+remains missed (disclosed); the causal follow-ups (E6 register swap —
+negative; G2b 5× classical corpus — the add-direction test) are in
+docs/RESULTS.md.
+
+```python
+from interscript_ml import Model
+model = Model.load("ara-diac-small-2.1")
+model.translate("قوله فحكمها في الوفاة")
+```
+
+Same IMF v1 contract as 2.0: dynamic fetch, sha256-verified, KV
+decode, margin report alongside the parity block. Checkpoint:
+`rababa_arabic_distill_small/run-007-r7-muon-6ep/best`.

src/gpu/modal_export.py

@@ -127,7 +127,7 @@
         "volume": "/volumes/rababa-checkpoints",
         "checkpoint": "rababa_arabic_distill_small/run-007-r7-muon-6ep/best",
         "metadata": "models/ara-diac-small/ara-diac-small-2.1.metadata.yaml",
-        "readme": "models/ara-diac-small/ara-diac-small-2.0.README.md",
+        "readme": "models/ara-diac-small/ara-diac-small-2.1.README.md",
         "test_volume": "/datasets/rababa",
         "test_data": "arabic-sadeed-imf/test.jsonl",
         "probe": "قوله",