diff --git a/TODO.sota/88-secryst-migration.md b/TODO.sota/88-secryst-migration.md new file mode 100644 index 00000000..fde2f8b0 --- /dev/null +++ b/TODO.sota/88-secryst-migration.md @@ -0,0 +1,29 @@ +# WO88 — full migration of the secryst-lineage models into the framework + +Status: IN_PROGRESS (2026-10-24, owner-directed: "fully migrate them +to interscript in our new framework") + +The secryst-era models (khm-latn-1.0, the legacy tha-g2p line) predate +the contract: no recipes, no corpora pins, no documented targets +(WO87's audit), thin provenance. Migration = bring each into the +2-YAML + gates world so every future retrain/release runs the same +lawful path. + +## Step 1 — metadata migration (this WO, now) +- `corpora/khm-latn.yaml`: the secryst lineage documented with honest + UNVERIFIED markers where provenance is not yet pinned (the training + corpus predates pinning; the decode evidence + khmer-dict-spice + source are documented). +- `recipes/khm-latn.yaml`: the retrain contract for the + framework-native khm-latn (the WO68/82 path). + +## Step 2 — retrain under gates (the real migration) +Each legacy line retrains through recipes+gates; the legacy artifact +stays resolvable via index aliases until its successor passes. +Order: khm-latn (B7 owner data pending → blocked), tha-phonemic +legacy line (superseded by tha-phonemic-1.0 — retire, don't retrain), +urd/fas lines (gates already apply to their successors). + +## Acceptance +Every id in models.yaml carries recipe+corpus or an explicit +retire note; nothing remains outside the contract.