docs: G2b verdict — 4.8231; register-add neutralized the epoch gain

docs/EXPERIMENTS.md

@@ -297,6 +297,27 @@ All rows passed the CER parity gate at release. Readings:
   shows a convergence-speed effect distinguishable from noise on the
   surviving curves; the DER gates remain the only verdicts.
 
+## G2b — label-scale ADD rung (run-011-tashkeela-6ep) — verdict recorded 2026-09-04
+
+- **Design** (ml #135): 48,000 total units — the control's news-mix
+  plus 18,000 additional classical Tashkeela units — trained 6 epochs
+  (G2a's epoch count). Spec `ara-diac-small-2-6ep-tashkeela`.
+- **Status:** COMPLETE — **4.8231** full-set windowed DER (n=1200;
+  teacher reproduces 2.289; paired bootstrap student−teacher +2.3717,
+  CI [2.194, 2.554]). From final_eval.json on the volume; run
+  completed and evaluated by its owners; recorded here per
+  TODO.substantiate item 04. Labels sha256 b59e2f5677693eab8968208e0
+  0a70590d03ce743fbc43b2acdb0ac9bd508c8fb.
+- **Read:** the domain-coverage attribution is now negative in BOTH
+  directions at matched epochs. Swapping news out (E6, 3ep): 5.8057.
+  Adding classical units on top (G2b, 6ep): 4.8231 — flat vs the 3ep
+  30k control (4.8218) and 0.25pp WORSE than news-only G2a (4.5701):
+  the Tashkeela add neutralized the epoch gain entirely. The
+  post-E3 residual is not register-shaped; the only levers that have
+  moved this rung are optimizer (−2.96pp), fresher teacher labels
+  (−0.47pp), and epochs (−0.25pp). What remains is GKD (on-policy
+  exposure, run-012 in flight) — after which the ladder is closed.
+
 ## GKD — on-policy distillation rung (REGISTERED/ARMED 2026-09-03)
 
 - **Status:** REGISTERED, launch pending owner ordering (the last

docs/PUBLICATION-NOTES.md

@@ -77,11 +77,16 @@ confound (fresh aux head for the final 23% of steps); E6
 constant-budget register swap (2026-09-02): 5.8057 — replacing news
 units with classical Tashkeela at constant total HURT (−0.98pp vs
 control), the domain-shaped-residual hypothesis's causal test
-failing in the swap direction; the add direction (G2b) remains
-open.** The E5/E6 pair is the paper's data-vs-architecture exhibit:
+failing in the swap direction; G2b label-scale ADD (2026-09-04,
+verdict recorded per TODO.substantiate 04): 4.8231 — 48k total
+(+18k classical) at 6 epochs is FLAT vs the 3ep control and 0.25pp
+worse than news-only G2a (4.5701): the add neutralized the epoch
+gain, closing the domain hypothesis negative in BOTH directions.**
+The E5/E6/G2b set is the paper's data-vs-architecture exhibit:
 two levers from the frontier-LLM literature (MTP, register
-diversification at constant budget) both regressed on byte-level
-student distillation — the levers that moved the rung were optimizer
+diversification — swap or add) plus the label-scale hypothesis all
+regressed or flat on byte-level student distillation — the levers
+that moved the rung were optimizer
 (E3), fresher teacher labels (E4), and epochs (G2a).
 
 ### 9. Muon optimizer A/B on the memory student (E3) — LANDED 2026-08-28
@@ -226,14 +231,19 @@ The decomposition for paper B, every line full-set with brackets:
 | + Muon (E3) | 5.2945 | — |
 | + r7 teacher | 4.8218 | — |
 | + 6 epochs (G2a) | 4.5701 | delta 2.12 [1.91, 2.35] |
+| register swap (E6, 3ep) | 5.8057 | negative |
+| register add (G2b, 6ep) | 4.8231 | delta 2.37 [2.19, 2.55] |
+| MTP-aux (E5, 3ep) | 5.0853 | negative (confound disclosed) |
 | depth halved (lite, 6ep) | 5.784 | delta 3.25 [3.03, 3.49] |
 
 Paper-B framing: levers compose roughly additively (optimizer >>
-teacher > epochs), and the E6 causal test came back NEGATIVE —
+teacher > epochs), and BOTH register tests came back NEGATIVE —
 swapping 8k news units for classical-register Tashkeela at constant
-30k total scored 5.8057 (−0.98pp vs control), so the residual is
-not fixed by register mix at constant budget; the add direction
-(G2b, 48k total) is the live test, and capacity appears ONLY as
+30k total scored 5.8057 (−0.98pp vs control), and adding 18k
+classical units on top at 6 epochs scored 4.8231 (flat vs the 3ep
+control, 0.25pp worse than news-only G2a — the add neutralized the
+epoch gain): the residual is not register-shaped in either
+direction, and capacity appears ONLY as
 depth — width is load-bearing (both stitch ratios collapsed) while
 depth trades 1.21pp for 37% of the artifact. The subset-overstatement
 phenomenon (five instances, up to 3.2x inflation) is the

pyproject.toml

@@ -68,6 +68,9 @@ markers = [
 [tool.ruff]
 line-length = 100
 target-version = "py310"
+# ADR-0001: the vendored Misraj evaluator is a provenance artifact —
+# excluded, never rewritten or reformatted (same policy as rababa's copy)
+extend-exclude = ["src/sadeedbench/vendored_sadeed_evaluator.py"]
 
 [tool.ruff.lint]
 select = ["E", "F", "I", "B", "UP", "SIM", "C4"]

scripts/publish_model.py

@@ -18,8 +18,8 @@
 import hashlib
 import os
 import re
-import subprocess
 import shutil
+import subprocess
 import sys
 import tempfile
 import zipfile

src/gpu/modal_export.py

@@ -191,7 +191,7 @@
 app = modal.App("interscript-ml-export", image=IMAGE)
 
 
-def normalize_precisions(precisions: "str | list[str]") -> list[str]:
+def normalize_precisions(precisions: str | list[str]) -> list[str]:
     """Accept both invocation forms: the parity/margins entrypoints pass
     a pre-split list, direct ::parity_model-style CLI calls pass a
     comma string."""
@@ -220,7 +220,7 @@ def _zip_has_parity(zip_path: Path) -> bool:
 
 
 def pending_precisions(
-    out_dir: Path, mid: str, precisions: "str | list[str]"
+    out_dir: Path, mid: str, precisions: str | list[str]
 ) -> list[str]:
     """Precision stages still to run. A stage is durably complete only
     when BOTH its margin report exists and its zip carries the parity

src/sadeedbench/scoring.py

@@ -9,8 +9,6 @@
 
 from __future__ import annotations
 
-from typing import Iterable
-
 
 def score_predictions(preds: list[str], gts: list[str]) -> dict:
     """Aggregate DER-CE over aligned (pred, gt) paragraph pairs."""

tests/test_parity_stage_skip.py

@@ -17,6 +17,7 @@
 
 def _write_zip(path: Path, parity: bool) -> None:
     import zipfile
+
     import yaml
 
     meta = {"id": "m", "precision": "int4"}

tests/test_sadeedbench_cli.py

@@ -56,7 +56,8 @@ def test_score_plain_text_lines(tmp_path: Path, capsys) -> None:
 def test_bootstrap_vs_reference(tmp_path: Path, capsys) -> None:
     data = _parquet(tmp_path)
     cand = _preds(tmp_path, "c.jsonl", [{"idx": i, "student": s} for i, s in enumerate(GT)])
-    stripped = [g.replace("َ", "").replace("ُ", "").replace("ْ", "").replace("ً", "").replace("ّ", "") for g in GT]
+    marks = "ًَُّْ"
+    stripped = ["".join(c for c in g if c not in marks) for g in GT]
     ref = _preds(tmp_path, "r.jsonl", [{"idx": i, "student": s} for i, s in enumerate(stripped)])
     rc = main(["score", "--preds", str(cand), "--data", str(data), "--key", "student",
                "--vs", str(ref), "--vs-key", "student"])