ronaldtse · Oct 22, 2026 · Oct 22, 2026
diff --git a/TODO.sota/57-a1-authority-completeness.md b/TODO.sota/57-a1-authority-completeness.md
@@ -1,6 +1,6 @@
 # WO57 — A1: authority-completeness audit
 
-Status: PLANNED · Plane A
+Status: IN_PROGRESS (phase 1 DONE 2026-10-21: docs/coverage ledger + three authority diffs — alalc 30/47, iso, bgnpcgn 17 gaps; phase 2 = encode the gaps)
 Goal: enumerate the UNIVERSE of authority romanization/transliteration
 systems and diff it against the 293 encoded in `maps/`. Every gap is
 either encoded or cited out-of-scope with an authority reference.

diff --git a/TODO.sota/58-a2-invertibility.md b/TODO.sota/58-a2-invertibility.md
@@ -1,6 +1,6 @@
 # WO58 — A2: invertibility ledger + reverse maps
 
-Status: PLANNED · Plane A
+Status: IN_PROGRESS (phase 1 DONE 2026-10-21: docs/invertibility ledger — 10 pairs probed, 1 lossless, 39/278 collisions + probe tool; phase 2 = reverse-map generation)
 Goal: for every system, determine whether forward∘reverse = identity;
 publish the loss conditions (homophone mergers, digraph ambiguity,
 case conventions, vowel elision).

diff --git a/TODO.sota/59-a3-equivalence.md b/TODO.sota/59-a3-equivalence.md
@@ -1,6 +1,6 @@
 # WO59 — A3: system-equivalence checker + distinguishing corpora
 
-Status: PLANNED · Plane A
+Status: MOSTLY DONE (ruby shipped #816 + self-testing CI #819; ts/py port pending)
 Goal: mechanically decide/estimate whether two systems coincide, and
 produce the minimal test corpus that distinguishes them.
 Method:

diff --git a/TODO.sota/60-a4-is1-semantics.md b/TODO.sota/60-a4-is1-semantics.md
@@ -1,6 +1,6 @@
 # WO60 — A4: IS-1 v2 — denotational semantics
 
-Status: PLANNED · Plane A
+Status: IN_PROGRESS (9 rulings frozen #818/#820/#821; remaining: metadata/tests, item algebra, charset, funcall — see WO80 inventory)
 Goal: complete the ISC specification with a semantics section so every
 runtime behavior is derivable from the document; zero open rulings.
 Method:

diff --git a/TODO.sota/62-b1-dispatch-doctrine.md b/TODO.sota/62-b1-dispatch-doctrine.md
@@ -1,6 +1,6 @@
 # WO62 — B1: the dispatch doctrine as registry defaults
 
-Status: PLANNED · Plane B
+Status: IN_PROGRESS (doctrine written: WO56 structural rules + E2 paper; registry-as-config pending)
 Goal: encode the technique-dispatch table — (script class, ambiguity
 type, data regime) → recipe — as registry defaults so a new language
 gets a recommended recipe automatically.

diff --git a/TODO.sota/64-b3-arabic-sota.md b/TODO.sota/64-b3-arabic-sota.md
@@ -1,6 +1,6 @@
 # WO64 — B3: Arabic to external SOTA (DER < 2.0, WikiNews < 8)
 
-Status: PLANNED · launchable now
+Status: IN_PROGRESS (r8 stage 2 training 2026-10-22, stage-3 SadeedDiac eval auto-chained; VRR closure = WO81)
 Goal: SadeedDiac DER < 2.0 (canonical 2.29), WikiNews multiref WER < 8
 (news specialist 8.85). VRR reported.
 Method: r8 noisy-student chain — synthesize with the canonical r7

diff --git a/TODO.sota/65-b4-urdu-sota.md b/TODO.sota/65-b4-urdu-sota.md
@@ -1,6 +1,6 @@
 # WO65 — B4: Urdu — diac CER < 3; G2P license path
 
-Status: PLANNED · G2P half BLOCKED-OWNER (commercial-license data)
+Status: IN_PROGRESS · G2P half BLOCKED-OWNER (commercial-license data); diac push 3.57 → < 3 pending
 Goal: urd-diac 3.57 → CER < 3; urd-g2p research-only → commercial-
 usable data lineage.
 Method: diac — oracle relabel arms only if teachers decorrelate (banked

diff --git a/TODO.sota/67-b6-thai-lexicon.md b/TODO.sota/67-b6-thai-lexicon.md
@@ -1,6 +1,6 @@
 # WO67 — B6: Thai — lexicon-first, three layers (PER < 2)
 
-Status: IN_PROGRESS (lexicon phase landed 2026-10-21)
+Status: IN_PROGRESS (RTGS encodings landed maps#204; lexicon corpus staged; training arm + IPA layers = WO82 pending)
 Lexicon phase DONE: corpora/tha-phonemic.yaml -- 28,644 train / 3,225
 test (word -> phonemic-with-tones pairs, ORST vocabulary x pythainlp
 royin engine); 6,291 engine-corrupt rows filtered and documented;

diff --git a/TODO.sota/68-b7-khmer-lexicon.md b/TODO.sota/68-b7-khmer-lexicon.md
@@ -1,6 +1,6 @@
 # WO68 — B7: Khmer — Chuon Nath lexicon conditioning (CER < 15)
 
-Status: IN_PROGRESS (phase 1 landed 2026-10-21; phase 2 = training arm)
+Status: IN_PROGRESS (phase 1 landed 2026-10-21; phase 2 = training arm launch)
 Phase 1 DONE: khm-dict corpus built + pinned (corpora/khm-dict.yaml) —
 3,261 gold Chuon-Nath (word→IPA) pairs, 326-row holdout, 72,048 SBBIC
 vocab rows; staged at bucket Interscript/isx-training/run-080-khm-lexicon

diff --git a/TODO.sota/69-b8-community-intake.md b/TODO.sota/69-b8-community-intake.md
@@ -1,6 +1,6 @@
 # WO69 — B8: prove the contribution story (3 community languages)
 
-Status: PLANNED
+Status: PLANNED (intake doc = WO86)
 Goal: three external-contributor languages go end-to-end through the
 2-YAML path (corpora + recipe), landing as shipped models without
 framework forks.

diff --git a/TODO.sota/70-b9-external-benchmark-cert.md b/TODO.sota/70-b9-external-benchmark-cert.md
@@ -1,6 +1,6 @@
 # WO70 — B9: external-benchmark certification
 
-Status: PLANNED
+Status: IN_PROGRESS (SadeedDiac protocol shipped as interscript-ml-tools; protocol-package consolidation pending)
 Goal: no self-graded SOTA claims — every headline reconciles to an
 external leaderboard/protocol package.
 Method: per language, pin the external protocol (Arabic: SadeedDiac +

diff --git a/TODO.sota/71-c1-conformance-matrix.md b/TODO.sota/71-c1-conformance-matrix.md
@@ -1,6 +1,6 @@
 # WO71 — C1: model × runtime conformance matrix in CI
 
-Status: PLANNED
+Status: DONE (all gates green, sustained: 3-compiler CI 16,482+8,238/0, six legs; model x runtime byte-exact, CI-enforced)
 Goal: every shipped model byte-exact in ruby/ts/py, CI-enforced, forever.
 Method: fixtures already frozen in conformance/ (12 rows each; sidecar
 on the release). Wire each runtime's CI to download the fixture set

diff --git a/TODO.sota/72-c2-metric-convention-gate.md b/TODO.sota/72-c2-metric-convention-gate.md
@@ -1,6 +1,6 @@
 # WO72 — C2: gates reject undeclared metric conventions
 
-Status: PLANNED
+Status: DONE (anchor law enforced in src/pipeline/contracts.py; undeclared conventions rejected)
 Goal: quality gates refuse a metric that doesn't declare its convention
 (row-mean vs corpus-level) and anchor.
 Method: extend the gate registry (WO55 pattern): MetricDeclaration

diff --git a/TODO.sota/73-c3-perf-budgets.md b/TODO.sota/73-c3-perf-budgets.md
@@ -1,6 +1,6 @@
 # WO73 — C3: performance budgets in CI
 
-Status: PLANNED
+Status: IN_PROGRESS (BudgetGate size check DONE; decode-latency budget = remainder)
 Goal: decode latency + artifact size budgets enforced so regressions
 block like test failures.
 Method: measure current shipped baselines (single-row greedy, p50/p95

diff --git a/TODO.sota/76-e1-paper-formalism.md b/TODO.sota/76-e1-paper-formalism.md
@@ -1,6 +1,6 @@
 # WO76 — E1: paper — the ISC formalism, corpus, and interoperability
 
-Status: PLANNED · Depends WO57/WO60
+Status: IN_PROGRESS (draft merged #333; camera-ready needs A4 completion)
 Goal: a complete draft (LREC/COLING-class) presenting ISC as an
 executable formalism for authority script conversion; the 293-system
 corpus; the differential-testing interoperability methodology; the

diff --git a/TODO.sota/77-e2-paper-dispatch.md b/TODO.sota/77-e2-paper-dispatch.md
@@ -1,6 +1,6 @@
 # WO77 — E2: paper — technique dispatch for underdetermined mappings
 
-Status: PLANNED · Depends WO62 + plane B results
+Status: IN_PROGRESS (draft merged #334; results matrix needs plane B final numbers)
 Goal: a complete draft presenting the dispatch doctrine: ambiguity
 class → technique, with the negative-results registry (RL, CTC,
 LLM-teachers, correlated distillation) as a first-class contribution

diff --git a/TODO.sota/78-e3-benchmark-artifacts.md b/TODO.sota/78-e3-benchmark-artifacts.md
@@ -1,6 +1,6 @@
 # WO78 — E3: frozen benchmarks as citable artifacts
 
-Status: PLANNED
+Status: IN_PROGRESS (frontier-predictions-v1 shipped; conformance-suite bundle pending)
 Goal: eval protocols + conformance suites published as citable,
 versioned artifacts.
 Method: protocol packages (WO70) + conformance vectors bundled as

diff --git a/TODO.sota/81-vrr-metric.md b/TODO.sota/81-vrr-metric.md
@@ -0,0 +1,29 @@
+# WO81 — VRR: valid-reading-rate, the ambiguity-aware metric
+
+Status: PLANNED · Plane B (B3 closure requirement, WO56 structural rule 1)
+
+Single-gold DER systematically undercounts legitimate ambiguity: when a
+skeleton has several dictionary-licensed readings, scoring against one
+gold penalizes a correct alternative. VRR scores a prediction as valid
+when it matches ANY licensed reading.
+
+Goal: VRR reported alongside DER for every Arabic (then Hebrew/Urdu)
+headline claim; the first number lands for `ara-diac-2.0` on
+SadeedDiac-25.
+
+Method:
+- Build the licensed-reading lexicon: skeleton (undiacritized form) →
+  set of fully-vocalized forms, extracted from the pinned vocalized
+  training corpora (provenance in `corpora/*.yaml`); dedupe by exact
+  diacritic pattern.
+- Implement `score_predictions_vrr` + CLI in interscript-ml-tools:
+  VRR = |{pred form ∈ licensed(skeleton)}| / |rows|, computed per row
+  and corpus-level, convention-declared (row-level match, exact
+  including diacritics).
+- Run against the frozen predictions in the
+  `frontier-predictions-v1` release — no model calls needed.
+- Record in `models.yaml` as a second metric block (protocol + anchor
+  per the C2 anchor law).
+
+Acceptance: VRR tool shipped; `ara-diac-2.0` VRR published next to
+DER 2.29; the metric convention declared and gated.
diff --git a/TODO.sota/82-thai-phoneme-layers.md b/TODO.sota/82-thai-phoneme-layers.md
@@ -0,0 +1,25 @@
+# WO82 — Thai IPA and IPA+tone layers (the TTS layer)
+
+Status: PLANNED · Plane B (B6 layers 2–3; layer 1 RTGS shipped in maps#204)
+
+WO56 requires three separately-tracked Thai target layers: RTGS
+(toneless, DONE — deterministic .isc encodings), IPA (toneless), and
+IPA+tone (TTS). The phoneme layers need a tonal data supply that is not
+yet identified.
+
+Goal: `tha-phonemic` eval reporting IPA layer v0 and IPA+tone v0 with
+tone accuracy; published PER < 2 on the RTGS layer stands from B6.
+
+Method:
+- Data supply (OWNER ASK): Royal Institute dictionary entries carry
+  pronunciation + tone; identify the machine-readable source and its
+  license (the 28,644-row staged lexicon covers orthographic→phonemic
+  without tone marks — audit whether tone is recoverable from it before
+  asking for more data).
+- Learned-IPA fallback for novel tokens: ByT5 arm conditioned per the
+  lexicon-first doctrine (hit-rate first-class).
+- Tone accuracy metric: consonant-class × syllable-type computed tone
+  vs lexicon tone, declared per C2.
+
+Acceptance: three layers tracked in models.yaml; IPA+tone v0 shipped
+with tone accuracy reported; RTGS layer unchanged.
diff --git a/TODO.sota/83-exporter-io-defaults.md b/TODO.sota/83-exporter-io-defaults.md
@@ -0,0 +1,24 @@
+# WO83 — Exporter defaults: fp32 graph IO always
+
+Status: PLANNED · Plane B/C (WO79 follow-through — the decision is
+recorded, the DEFAULT is not enforced)
+
+WO79's owner decision: exporter-side fp16→fp32 graph-IO normalization.
+The fix exists as a post-hoc tool (`tools/normalize_fp16_io.py`) and
+was applied to urd/fas by hand. Future exports can still ship raw-fp16
+boundaries unless the exporter itself defaults to normalized IO.
+
+Goal: no export path can produce an artifact whose graph boundary is
+fp16; the normalizer becomes the exporter's default, not a repair step.
+
+Method:
+- Fold the IO-normalization into `export_imf_hf.py` (interscript-train):
+  weights may stay fp16; graph outputs always Cast(to=FLOAT), fp16
+  decoder inputs narrowed via internal Cast, raw tensors internal.
+- Publish guard: reject (or auto-normalize + record) any artifact whose
+  graph IO carries fp16 — enforced in `src/pipeline/`, not by memory.
+- Verification: one fresh export decodes reference-identically (torch
+  vs artifact, ≥500 samples in shipped decode mode).
+
+Acceptance: exporter default normalized; publish guard blocks raw-fp16
+IO; a fresh export passes parity unchanged.
diff --git a/TODO.sota/84-train-repo-hygiene.md b/TODO.sota/84-train-repo-hygiene.md
@@ -0,0 +1,25 @@
+# WO84 — interscript-train: land the local uncommitted work
+
+Status: PLANNED · hygiene (discovered 2026-10-22; durability lives only
+on this laptop)
+
+The local checkout of `training/wo31-wo36-campaign-scripts` carries
+uncommitted improvements that production runs already depend on
+(r8_noisy_student_hf.py's self-contained decode + sync_to_bucket
+durability, C-Levenshtein eval path), plus untracked research docs and
+utilities. One `/tmp`-style loss and the running campaign's durability
+is unexplained by any commit.
+
+Goal: `git status` clean on the campaign branch lineage; every behavior
+a running job depends on exists in a pushed commit.
+
+Method:
+- Branch from `training/wo31-wo36-campaign-scripts`.
+- Commit the five modified scripts with per-change messages (explicit
+  paths only — never `git add -A`).
+- Triage untracked: `TODO.research/*` docs, `conformance_runtime_check.py`,
+  `classes.json` → commit if source; `dist/` → never (build output).
+- PR into the campaign lineage; merge after green.
+
+Acceptance: pushed branch + merged PR; running jobs' behavior matches
+a commit; worktree clean.
diff --git a/TODO.sota/85-ledger-status-sweep.md b/TODO.sota/85-ledger-status-sweep.md
@@ -0,0 +1,16 @@
+# WO85 — Ledger status sweep: headers must match verified reality
+
+Status: DONE (2026-10-22 sweep in the same PR that adds WO81–86)
+
+The `Status:` headers of WO57–78 lagged execution: C1/C2 recorded
+PLANNED while all gates were green; A3/E1/E2 recorded PLANNED while
+merged. A ledger that lies about state is worse than no ledger.
+
+Goal: every WO header reflects the verified state at sweep time;
+sweeps re-run whenever a plane milestone lands.
+
+Method: sweep headers against evidence (merged PRs, CI, gates) —
+never against intention. Dated.
+
+Acceptance: this file records the sweep; headers current as of
+2026-10-22.
diff --git a/TODO.sota/86-b8-intake-doc.md b/TODO.sota/86-b8-intake-doc.md
@@ -0,0 +1,25 @@
+# WO86 — B8 intake: the community language path, documented
+
+Status: PLANNED · Plane B
+
+B8's proof is three outside languages through the 2-YAML path — but
+there is no published onboarding document; the path exists only in
+maintainer heads and CLAUDE.md notes.
+
+Goal: a contributor can onboard a new language without reading the
+pipeline: what to bring (authority reference or corpus + license), the
+two YAML shapes, the eval-set expectation, the gates their PR must
+pass, and how a learned component slots in if the table
+underdetermines.
+
+Method:
+- `docs/contributing-language.md` in interscript-models: the 2-YAML
+  contract (`recipes/`, `corpora/`), provenance law, gate walkthrough
+  (quality/parity/license/conformance/regression), worked example
+  end-to-end (smallest shipped map as the template).
+- Mirror a short pointer from the maps repo README.
+- The path itself needs no code change — if it does, that is a
+  framework bug: record it, do not patch around it.
+
+Acceptance: doc merged; one dry-run onboard of a trivial language
+through the gates using only the doc.