From fe7b1f24fbef858db40018ba087c619f78e11928 Mon Sep 17 00:00:00 2001
From: Ronald Tse <ronald.tse@gmail.com>
Date: Tue, 1 Sep 2026 08:23:26 +0200
Subject: [PATCH] =?UTF-8?q?feat(e6):=20Tashkeela-mix=20rung=20=E2=80=94=20?=
 =?UTF-8?q?registered=20+=20spec=20+=20data=20on=20the=20volume?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Constant-budget register diversification: 8,000 classical Tashkeela
units replace 8,000 news units (30,000 total; steps, schedule,
optimizer, teacher identical to run-006/E5's control). Data: 120,000
decontaminated units at r5-units/tashkeela-scale.txt — stripped-
normalized exact + 80-char-prefix decontamination against
SadeedDiac-25 (48 dropped), deduped internally (17,036) and against
the existing mix, seeded shuffle. Gate: adopt at <=4.5218; prediction
4.45-4.75. Launch gated on a free A100 slot.
---
 docs/EXPERIMENTS.md        | 22 ++++++++++++++++++++++
 src/gpu/distill_specs.yaml | 23 +++++++++++++++++++++++
 2 files changed, 45 insertions(+)

diff --git a/docs/EXPERIMENTS.md b/docs/EXPERIMENTS.md
index 8ce60c0..b0096ea 100644
--- a/docs/EXPERIMENTS.md
+++ b/docs/EXPERIMENTS.md
@@ -225,6 +225,28 @@ All rows passed the CER parity gate at release. Readings:
   the tail, but the factorial attributes most of the remaining gap to
   domain coverage, so the effect should be second-order.
 
+## E6 — label-scale/domain-mix rung (run-008-tashkeela-mix)
+
+- **Status:** REGISTERED 2026-09-01; data built (120,000 decontaminated
+  Tashkeela units on rababa-datasets:r5-units/tashkeela-scale.txt);
+  launches when a GPU slot frees (E5 + owner runs hold the budget).
+- **Hypothesis:** the E2/E3 factorial attributes the client tier's
+  remaining ~2.0pp to domain coverage; the shipped mix is news-heavy.
+  Replacing 8,000 of the 24,000 news-domain units with classical-
+  register Tashkeela units AT CONSTANT TOTAL (30,000 units, identical
+  steps/schedule/optimizer/teacher) diversifies register coverage —
+  classical text is also the benchmark's own register.
+- **Data hygiene:** lines <=1,450B with Arabic; decontaminated against
+  SadeedDiac-25 by stripped-normalized exact match and 80-char prefix
+  match (48 units dropped); deduped internally (17,036) and against
+  the existing domain/replay units; seeded shuffle; 120k emitted, the
+  chain samples 8,000.
+- **Pre-agreed gate:** adopt if full-set windowed DER <= 4.5218
+  (>=0.3pp over 4.8218, the E3-style bar); honest report otherwise.
+- **Prediction (registered):** 4.45-4.75 — a register-diversification
+  effect at constant budget should be real but second-order; if the
+  domain attribution is right, this should beat E5's MTP-aux.
+
 ## Parked
 
 - **Speculative decoding** (LongCat converts sparsity→speed): revisit
diff --git a/src/gpu/distill_specs.yaml b/src/gpu/distill_specs.yaml
index 47c6e7c..67e0583 100644
--- a/src/gpu/distill_specs.yaml
+++ b/src/gpu/distill_specs.yaml
@@ -139,6 +139,29 @@ ara-diac-small-2-mtp:
     beta: 0.15
   mode: sequence
 
+ara-diac-small-2-scale:
+  # E6 (EXPERIMENTS.md): constant-budget register diversification —
+  # 8,000 Tashkeela classical units replace 8,000 news units (30k total,
+  # identical steps/schedule to run-006/E5's control)
+  teacher: rababa_arabic_byt5/run-007-news/best
+  teacher_volume: rababa
+  student_init: google/byt5-small
+  optimizer: muon
+  muon_lr: '0.01'
+  train: r5-units/domain.txt
+  train_extra:
+  - r5-units/tashkeela-scale.txt
+  - r5-units/replay.txt
+  unit_limits:
+  - 16000
+  - 8000
+  - 6000
+  max_len: 1450
+  label_beams: '1'
+  out: rababa_arabic_distill_small/run-008-tashkeela-mix
+  labels_file: teacher_labels_r7.jsonl
+  mode: sequence
+
 ara-diac-small-2:
   teacher: rababa_arabic_byt5/run-007-news/best
   teacher_volume: rababa