diff --git a/TODO.sota/87-target-accurate-naming.md b/TODO.sota/87-target-accurate-naming.md
new file mode 100644
index 00000000..79872724
--- /dev/null
+++ b/TODO.sota/87-target-accurate-naming.md
@@ -0,0 +1,28 @@
+# WO87 — target-accurate model naming: document, then rename
+
+Status: PLANNED · Phase 1 DONE (tha-phonemic-1.0, models#357); Phase 2
+= the legacy 33
+
+Owner decision (2026-10-24): model names target the output convention.
+Phase 1 renamed the release candidate (tha-g2p-lex-1.0 ->
+tha-phonemic-1.0) — its target is exhaustively verified.
+
+Phase 2 CANNOT blanket-rename today: the 2026-10-24 audit found ZERO
+entries documenting their output target convention (source/target
+scripts empty zoo-wide). Renaming undocumented models fabricates.
+
+Method (per model, mechanical):
+1. Decode 3 representative rows through the reference loader
+   (one batch job over all 33).
+2. Record the target convention in models.yaml
+   (target_script/target_convention fields — a contracts.py extension).
+3. Rename only on evidence: {lang}-{target}[-{tier}]-{version};
+   old ids become aliases in the next index tag (zero breakage —
+   index-v12 consumers keep resolving forever).
+4. Quantization/architecture/register tiers (int8/int4/fp16,
+   layerdrop/plane/news) are accurate orthogonal dimensions — they
+   ride along unchanged.
+
+Acceptance: every models.yaml entry carries a documented target
+convention; every rename is evidence-backed; index-vN aliases keep
+every historical id resolvable; docs + catalogue regenerate.