docs: TODO.sota 25-26 — the SOTA push arms (AR specialist full-scale, HE noisy student) (#280)

Co-authored-by: Ronald Tse <ronald.tse@gmail.com>

Author: ronaldtse

TODO.sota/00-README.md

@@ -62,6 +62,8 @@ hf jobs run --flavor a100-large -d \
 | 22 | 22-runtime-releases.md | P1 | 14 | PREPARED — owner-gated versions |
 | 23 | 23-ort-alignment.md | P1 | 18 | CLOSED — characterization complete; 3-leg parity GREEN at documented bound |
 | 24 | 24-verdict-wave.md | P1 | 17 | r8a/b/c + run-026 + thai CLOSED; r8d (knee) RUNNING |
+| 25 | 25-ar-news-specialist.md | P1 | 17 | RUNNING (run-029, register-pure full scale) |
+| 26 | 26-heb-noisy-student.md | P1 | 04 | stage 1 RUNNING (labeling); stage 2 staged |
 
 ## Schedule
 

TODO.sota/25-ar-news-specialist.md

@@ -0,0 +1,11 @@
+# WO25 — Arabic news specialist at full scale (the OOD SOTA play)
+
+RUNNING (run-029, a100-large): register-PURE — NO classical tashkeela —
+QCRI silver at full scale (900K units ≈ 4.5M words), 2 epochs, r7-init.
+r8b proved 1M words of silver moves OOD 6.35 WER; 4.5x the data and
+zero register dilution is the honest shot at their 2.70 (in-domain for
+them, trained on the same corpus family).
+
+- Gate: WikiNews-2024 multiref <= 5.0 (stretch <= 3.0); ID reported
+  as-is (specialist doctrine — two families, never averaged).
+- Ship: ara-diac-news-1.0 via export_imf_hf (one command).

TODO.sota/26-heb-noisy-student.md

@@ -0,0 +1,9 @@
+# WO26 — Hebrew noisy student (defend and extend the crown)
+
+8.18 is text-only SOTA; byt5-large showed capacity is NOT the lever at
+50K gold units — DATA is. Stage 1 (RUNNING): plane-2.0 pseudo-labels
+40K hewiki windows with per-position margins (NO LLM). Stage 2:
+gold v4 (50K) + pseudo (cap 40K, keep_frac >= 0.9) on byt5-base.
+
+- Gate: artifact-level DER < 8.18 else negative.
+- Lesson applied: capacity arms are closed; only data moves Hebrew.