docs: matched-architecture arms (WO32)

TODO.sota/00-README.md

@@ -67,6 +67,7 @@ hf jobs run --flavor a100-large -d \
 | 28 | 28-oracle-complementarity.md | P2 | — | CLOSED — ceiling 96.78 real; naive voting catastrophic |
 | 30 | 30-heb-learned-ipa-v0.md | P1 | 09 | NEGATIVE — CER 1.18; ASR teacher is the bottleneck |
 | 31 | 31-heb-phoneme-asr-v1.md | P1 | 30 | SPEC'D — tune the phoneme CTC on Hebrew first (owner block) |
+| 32 | 32-matched-architecture-arms.md | P1 | 17 | RUNNING — run-033 plane-large + run-034 BiLSTM on their corpus |
 | 26 | 26-heb-noisy-student.md | P1 | 04 | NEGATIVE — 8.72 vs 8.18; all three text levers closed |
 
 ## Schedule

TODO.sota/32-matched-architecture-arms.md

@@ -0,0 +1,16 @@
+# WO32 — the matched-architecture program (same corpus MUST win)
+
+Audit: no contamination either side; seq2seq converged at 10.03 → gap
+is architecture + teacher. Arms:
+
+- run-033: byt5-large plane + full 900K silver (news-pure, K=3) —
+  their recipe class, stronger backbone. Gate: WN-2024 multiref < 10
+  (stretch: single digits).
+- run-034: literal BiLSTM tagger on the same corpus — reproduces their
+  class from scratch. Gate: WN-2024 < 15 validates the class at our
+  scale; comparison anchor.
+- SOTA-stack layer above both: K-pass conditioning, byte-pretrained
+  backbones, preserve mode (already in the plane family).
+
+Ship rule: whichever arm dominates WikiNews-2024 becomes the
+ara-diac-news successor; seq2seq line is closed (converged at 10.03).

docs/RESULTS.md

@@ -1397,3 +1397,25 @@ NOT triggered — scaling a weak teacher wastes it. The audio path
 remains open but requires: fine-tune the phoneme CTC on Hebrew audio
 (FLEURS he ~10h is exactly that training set), then relabel, then
 v1. Spec'd as WO31; needs owner green-light for the next compute block.
+
+## The "same corpus must win" program — convergence CLOSED, both matched arms in flight (2026-10-09)
+
+**Contamination audit: CLEAN** — 0/356 WikiNews-2024 test sentences
+share even one 8-gram with the QCRI 5M-word corpus; their 2.70 is
+honest generalization, ours is too. **run-029c (+2 epochs, seq2seq
+specialist): 10.03/8.95** (from 10.13) — the seq2seq line has
+CONVERGED; doubling training bought 0.10 WER. The remaining 7.3-point
+gap is therefore ARCHITECTURE (tagging vs generation) + teacher quality,
+and two arms now attack it directly on their exact recipe:
+
+- **run-033** (byt5-large PLANE = our tagging family + full 900K-unit
+  corpus, news-pure, 3ep, K=3, gradient checkpointing) — their class
+  with a far stronger backbone.
+- **run-034** (literal BiLSTM tagger, Fadel/QCRI-style: char-emb →
+  3×BiLSTM-512 → per-position haraqat head, from scratch, same corpus)
+  — their architecture, rebuilt and run at our data scale.
+
+Fixes en route (train#128-#131): haraqat split_planes mark-only crash
+(QCRI silver has pure-combining-mark windows — latent library bug, now
+tested); byt5-large OOM → PLANE_CHECKPOINT gradient checkpointing;
+LSTM OOM → BIL_BS; padded-label loss smearing → ignore_index=-100.