From cb5e49e38e2aeba4664f94cd1b41744ff53d15ea Mon Sep 17 00:00:00 2001
From: Ronald Tse <ronald.tse@gmail.com>
Date: Wed, 7 Oct 2026 18:09:58 +0800
Subject: [PATCH] =?UTF-8?q?docs:=20TODO.sota=2017-19=20=E2=80=94=20registe?=
 =?UTF-8?q?r=20closure=20arms,=20parity=20runs,=20lexicon=20loader?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 TODO.sota/00-README.md                  |  3 +++
 TODO.sota/17-arabic-register-closure.md | 21 +++++++++++++++++++++
 TODO.sota/18-parity-runs.md             |  7 +++++++
 TODO.sota/19-lexicon-loader.md          |  6 ++++++
 4 files changed, 37 insertions(+)
 create mode 100644 TODO.sota/17-arabic-register-closure.md
 create mode 100644 TODO.sota/18-parity-runs.md
 create mode 100644 TODO.sota/19-lexicon-loader.md

diff --git a/TODO.sota/00-README.md b/TODO.sota/00-README.md
index 9cf8a9a..8519322 100644
--- a/TODO.sota/00-README.md
+++ b/TODO.sota/00-README.md
@@ -54,6 +54,9 @@ hf jobs run --flavor a100-large -d \
 | 14 | 14-ts-plane-port.md | P2 | — | DONE (ts#99/#100/#101, CI leg models#265) |
 | 15 | 15-api-edge.md | P2 | 14 | DONE (api#29: edge-first /v1/infer, kind in index) |
 | 16 | 16-parking-lot.md | — | — | parked (not now) |
+| 17 | 17-arabic-register-closure.md | P1 | 11 | r8a RUNNING; r8b launching |
+| 18 | 18-parity-runs.md | P1 | 10 | pending |
+| 19 | 19-lexicon-loader.md | P2 | 13 | pending |
 
 ## Schedule
 
diff --git a/TODO.sota/17-arabic-register-closure.md b/TODO.sota/17-arabic-register-closure.md
new file mode 100644
index 0000000..55d86e6
--- /dev/null
+++ b/TODO.sota/17-arabic-register-closure.md
@@ -0,0 +1,21 @@
+# WO17 — Arabic register closure (the WikiNews gap attack)
+
+**Why:** WO01 proved the 17.38-vs-2.70 WikiNews-2024 gap is register,
+not vocabulary (OOV 1.61%). QCRI's own README confirms their news model
+is trained on 5M words of Wikipedia SILVER labels from their BiLSTM —
+the same noisy-teacher technique as our r8. Two arms decide which
+teacher's silver closes our gap:
+
+- **r8a (self)**: r7 pseudo-labels arwiki (RUNNING, label-v4; keep_frac
+  filter at train time). Our teacher: 2.2864 on the harder
+  expert-reviewed SadeedDiac-25.
+- **r8b (QCRI silver)**: their published Wikipedia_20240420.diac.jsonl
+  (95MB, ~5M words, machine-labeled by their BiLSTM ~3% WER). Published
+  dataset — same standing as Tashkeela/Nakdimon in our stack; NOT an LLM
+  (no-LLM rule intact). Provenance + license recorded in the dataset
+  repo.
+
+**Dual-surface gate (both arms, in-job):** SadeedDiac-25 DER (hold the
+2.2864 line) AND WikiNews-2024 multiref WER/DER (move 17.38/11.83).
+A model that wins OOD but collapses ID is a regression, not a win.
+Ship the arm that dominates both; if neither does, negative verdict.
diff --git a/TODO.sota/18-parity-runs.md b/TODO.sota/18-parity-runs.md
new file mode 100644
index 0000000..835caf5
--- /dev/null
+++ b/TODO.sota/18-parity-runs.md
@@ -0,0 +1,7 @@
+# WO18 — parity runs for the shipped artifacts
+
+1. Attach golden/heb-diac-plane-2.0.jsonl to the golden-v2 release.
+2. Dispatch neural-parity (3 legs: ruby, python, ts) for
+   heb-diac-plane-2.0 — first model validated across ALL runtimes at
+   ship time.
+3. Green runs recorded in RESULTS.md.
diff --git a/TODO.sota/19-lexicon-loader.md b/TODO.sota/19-lexicon-loader.md
new file mode 100644
index 0000000..12336a2
--- /dev/null
+++ b/TODO.sota/19-lexicon-loader.md
@@ -0,0 +1,6 @@
+# WO19 — lexicon fetch helper (client convenience)
+
+`thai_hybrid.fetch_lexicon()`: resolve the canonical release asset
+(tha-lexicon-kaikki-1.0) with sha256 verification into a Lexicon —
+same trust contract as model artifacts (no unverified data paths).
+TDD; py first (ts/ruby follow on demand).