diff --git a/TODO.sota/53-ship-set.md b/TODO.sota/53-ship-set.md new file mode 100644 index 00000000..22c169be --- /dev/null +++ b/TODO.sota/53-ship-set.md @@ -0,0 +1,55 @@ +# WO53 — The ship set (assembled 2026-10-18; publish owner-gated) + +Every lane, every gate, every artifact — the complete release picture. +The publish guard (`pipeline/publish.py`) makes an incomplete release +structurally impossible; this file is the human-readable counterpart. + +## Lane-by-lane + +| lane | candidate artifact | quality (verified) | parity | conformance | license | status | +|---|---|---|---|---|---|---| +| urd-diac | urd-diac-base-1.0-fp16.zip (1.35GB) | gate CER 0.0357 row-mean | ≡ torch, 300/300 rows (WO51) | runtime 12/12 byte-exact | **GATED-1** | ready pending license | +| urd-g2p | run-059/best export | 0.16 CER row-mean | cleared by WO51 contract | fixtures pending ship | **GATED-1** | ready pending license | +| ara-diac | r6/r7 canonical (shipped 1.1) | WER 2.58 / DER 1.53 (r6 verified) | int8 zipcheck GOOD | plane family | **GATED-2** | stage-2 may improve; relabel resume9 consolidating (C-lev fix) | +| fas-diac | run-077/best | 0.603% row-mean / 0.777% corpus, word-acc 82.33% | pending export | fixtures pending ship | CLEAR (CC0) | **GATED-4** (gate value) | +| he-* | — | — | — | — | no-license sources | **GATED-3** (data supply) | +| tha-* | shipped small tier | per WO08 contract | plane family | existing | clear | closed | + +## The owner-gated queue (exactly what blocks shipping) + +1. **urd-g2p weights** — training data verified **non-commercial** + (humair025/urdu-g2p-dictionary, HF card). Options: ship + research-only, or retrain on an open dictionary first. +2. **ara-diac (qcri-wiki)** — QCRI's Qatara repo carries **no license** + (default copyright). Options: contact QCRI for explicit terms, or + rebuild the ara lane on Tashkeela-derived data (GPLv2 — weights + redistribution implications then apply). +3. **urd-dia sources** — the paper's corpus mixes `mahwizzzz` + + `humairmunirawn`; **no licenses found on either**. Options: contact + the paper authors, or rebuild from licensed sources. +4. **fas-diac gate value** — the docs' 0.52% is **proven + unreproducible** (5-arm elimination; WO47). Accept the verified + ceiling (run-077/best, 0.603%/0.777%) and re-anchor the recipe gate, + or commission the corpus-rebuild arm (~18% of raw rows were dropped + by the derivation) before shipping. +5. **Versions + final go** — per standing law, version numbers are the + owner's decision. Name them (e.g. `urd-diac-base-1.0`, + `fas-diac-base-1.0`) when saying go; the publish guard enforces the + rest (explicit semver, green gates, conformance fixtures). +6. (HE supply asks — micsell no-license and QCRI-contact drafts sit in + TODO.sota/43/17, ready to send whenever you choose.) + +## Machine-gated (not owner) + +- **stage-2 (ara oracle distill)**: resume9 running — all three arms + complete; consolidation now uses the C levenshtein (the pure-python + one needed ~56h vs the 23h timeout — the loop that killed resume6/8). + `RELABEL_DONE` → stage-2 launch through the spine runner. + +## Non-blocking backlog (recorded, not holes) + +Housekeeper (ckpt pruning / dead-repo inventory) · exporter +attention-mask option for batched serving (WO51 remedy (b)) · +conformance coverage grows with each shipped artifact · +src/gpu/modal_* deprecation (owner-flagged; nothing deleted) · +runtime default index URL re-point at next runtime release.