feat: ara-diac-small-2.0 retargeted to run-006 — E4 landed at 4.8218

The E4 arm completed as pre-registered (gate ≤6.26, prediction 4.3-5.0):
r7 canonical teacher labels + Muon on the vanilla ByT5-small scores
4.8218 full-set (teacher reproduces 2.289), superseding the run-005
rung (5.2945) this branch originally targeted — as its own metadata
anticipated. 42% error reduction vs the shipped 1.0 at identical
architecture and artifact size. Strict teacher+0.5pp gate still missed
(+2.53pp), disclosed.

Author: Ronald Tse

docs/EXPERIMENTS.md

@@ -152,7 +152,11 @@ All rows passed the CER parity gate at release. Readings:
 
 ## E4 — ara-diac-small-2.0 candidate (run-006-r7-muon)
 
-- **Status:** queued (registered 2026-08-28 before launch).
+- **Status:** COMPLETE (2026-08-29). **PASSED — 4.8218** full-set
+  windowed DER (gate ≤ 6.26; registered prediction 4.3–5.0; teacher r7
+  reproduces 2.289 vs documented 2.2864). −3.44pp / 42% relative vs the
+  shipped 1.0 at identical architecture and size; matches the PKM arm's
+  4.829 without memory layers. Released as ara-diac-small-2.0.
 - **Hypothesis:** the two measured wins compound — the r7 canonical
   teacher (better labels; ID 2.2864 vs 2.5793) plus the E3-adopted
   Muon optimizer (−2.727pp on r6 labels) — moving the client rung far

docs/RESULTS.md

@@ -315,6 +315,25 @@ gap) — roughly additive, slightly sub-additive on memory. Residual
 ~2.2pp is domain coverage. Optimization is the dominant recoverable
 term of the distillation gap at the ByT5-small rung.
 
+### run-006-r7-muon — E4: the 2.0 release candidate (2026-08-29)
+
+The two measured wins compounded on the vanilla architecture: r7
+canonical teacher (fresh greedy labels) + Muon optimizer, same
+corpus/limits/seed family. Pre-registered E4 gate ≤ 6.26 (prediction
+4.3–5.0):
+
+| Model | DER-CE (full 1,200) |
+|---|---|
+| Teacher (r7, in-run) | 2.2890% |
+| **ByT5-small, r7 labels + Muon (run-006)** | **4.8218%** |
+| ByT5-small, r6 labels + Muon (run-005) | 5.2945% |
+| ByT5-small, r6 labels + AdamW (run-002, shipped 1.0) | 8.2590% |
+
+**4.8218 — gate passed; −3.44pp / 42% relative vs the shipped 1.0** at
+identical architecture and artifact size. Matches the PKM arm's 4.829
+without the memory layers. Release: ara-diac-small-2.0 (run-006
+checkpoint; strict teacher+0.5pp still missed at +2.53pp, disclosed).
+
 Leaderboard context (SadeedDiac-25, Misraj evaluator, zero-skip,
 harakat-projected DER-CE): the teacher tier (r6, 580M) at 2.5793
 (reproduced at 2.5815, 2026-08-26) is the best dedicated model measured

models/ara-diac-small/ara-diac-small-2.0.README.md

@@ -1,17 +1,18 @@
 # ara-diac-small-2.0
 
 Arabic diacritization (haraqat restoration). Client-tier ByT5-small
-student, identical corpus/labels/epochs to ara-diac-small-1.0 with one
-variable changed: **the Muon optimizer** (E3 factorial). That single
-change closes 2.96pp of the 5.68pp teacher-student gap:
+student — the two measured wins of the campaign compounded on the same
+architecture and artifact size as 1.0: the **r7 canonical teacher**
+(2.2864; fresh greedy labels) and the **Muon optimizer** (E3-adopted).
 
-- 1.0 (AdamW): 8.26 full-set windowed DER-CE
-- **2.0 (Muon): 5.29** (teacher r6: 2.58; in-run reproduction 2.60)
+- 1.0 (r6 labels, AdamW): 8.26 full-set windowed DER-CE
+- 2.0 (r7 labels, Muon): **4.82** (teacher r7 in-run: 2.289)
 
-The strict teacher+0.5pp gate is still missed; the residual decomposes
-as ~0.70pp capacity + ~2.25pp domain coverage (E2/E3 factorial), and an
-r7-teacher re-distillation is in flight. Identical IMF v1 contract:
-dynamic fetch, sha256-verified, KV decode, margins JSON alongside.
+A 42% error reduction, pre-registered as E4 (gate ≤ 6.26; prediction
+4.3–5.0 — landed at 4.82). The strict teacher+0.5pp gate is still
+missed (+2.53pp, disclosed); the E2/E3 factorial attributes the
+residual to domain coverage. Identical IMF v1 contract: dynamic fetch,
+sha256-verified, KV decode, margins JSON alongside.
 
 ```python
 from interscript_ml import Model

models/ara-diac-small/ara-diac-small-2.0.metadata.yaml

@@ -8,25 +8,26 @@ opset: 14
 decoder: kv
 precision: fp32
 license: BSD-3-Clause
-trained_from: 'sequence-level KD from the r6 teacher (rababa_arabic_byt5/run-006-morph/best,
-  2.5793 windowed DER-CE full protocol): the same 29,322 greedy teacher labels as
-  ara-diac-small-1.0, with the Muon optimizer (E3 factorial, vanilla student arm).
-  Checkpoint rababa-checkpoints:/rababa_arabic_distill_small/run-005-muon/best. The
-  optimizer alone closes 2.96pp of the 5.68pp teacher-student gap measured for the
-  1.0 (AdamW) release: 8.259 -> 5.2945 full-set windowed DER-CE (teacher reproduces
-  2.5997 in-run). Still misses the strict teacher+0.5pp gate — the residual decomposes
-  as ~0.70pp capacity + ~2.25pp domain coverage per the E2/E3 factorial; miss disclosed.
-  An r7-teacher re-distillation is in flight and expected to supersede as a later rung.'
+trained_from: 'sequence-level KD from the r7 canonical teacher (rababa_arabic_byt5/run-007-news/best,
+  2.2864 windowed DER-CE full protocol): fresh greedy r7 labels on the same r5-units
+  corpus/limits as ara-diac-small-1.0, Muon optimizer (E3-adopted), vanilla ByT5-small
+  (E4, pre-registered gate <= 6.26). Checkpoint
+  rababa-checkpoints:/rababa_arabic_distill_small/run-006-r7-muon/best. The two
+  measured wins compound: 8.259 -> 4.8218 full-set windowed DER-CE (teacher reproduces
+  2.289 in-run vs documented 2.2864) — a 42% error reduction on the 1.0 release at the
+  same architecture and artifact size. Still misses the strict teacher+0.5pp gate
+  (+2.53pp; miss disclosed); the E2/E3 factorial attributes the residual to domain
+  coverage.'
 metrics:
 - name: der_teacher_fullset
-  value: 2.5997
+  value: 2.289
   protocol: windowed DER-CE (1400-byte windows, word-boundary split, greedy,
     haraqat-projected, Misraj evaluator); full 1,200-paragraph SadeedDiac-25;
-    in-run reproduction of the documented 2.5815/2.5793
-  source: interscript/interscript-ml docs/RESULTS.md#run-005-muon
+    in-run reproduction of the documented 2.2864 (r7 canonical teacher)
+  source: interscript/interscript-ml docs/RESULTS.md#run-006-r7-muon
 - name: der_student_fullset
-  value: 5.2945
-  protocol: same full-set harness; +2.69pp over the in-run teacher — the Muon
-    arm of the E3 2x2 factorial (vanilla student, optimizer as the only variable
-    vs the 8.259 AdamW release)
-  source: interscript/interscript-ml docs/RESULTS.md#run-005-muon
+  value: 4.8218
+  protocol: same full-set harness; E4 (r7 teacher labels + Muon, vanilla
+    ByT5-small) vs the 8.259 AdamW/r6-labels 1.0 release — a 42% error
+    reduction at identical architecture and artifact size
+  source: interscript/interscript-ml docs/RESULTS.md#run-006-r7-muon

src/gpu/modal_export.py

@@ -116,7 +116,7 @@
     },
     "ara-diac-small-2": {
         "volume": "/volumes/rababa-checkpoints",
-        "checkpoint": "rababa_arabic_distill_small/run-005-muon/best",
+        "checkpoint": "rababa_arabic_distill_small/run-006-r7-muon/best",
         "metadata": "models/ara-diac-small/ara-diac-small-2.0.metadata.yaml",
         "readme": "models/ara-diac-small/ara-diac-small-2.0.README.md",
         "test_volume": "/datasets/rababa",