diff --git a/TODO.sota/64-b3-arabic-sota.md b/TODO.sota/64-b3-arabic-sota.md
index 8a357e7..694aeed 100644
--- a/TODO.sota/64-b3-arabic-sota.md
+++ b/TODO.sota/64-b3-arabic-sota.md
@@ -26,3 +26,28 @@ the news specialist; decode policy unchanged (greedy, first-alternates).
 HF Jobs, --detach, --timeout 23h, step ckpts; ≤2 GPU budget shared
 with B7.
 Acceptance: gates green + external protocol reconciliation (B9).
+
+
+## Oracle-distill launch scoping (2026-10-29)
+
+The WO39 chain is script-complete and ready to run: stage 1
+`distill_oracle_hf.py` (300K QCRI rows, three arms, per-row oracle) and
+stage 2 `continue_oracle_train_hf.py` (continue-train run-047/best 2 epochs,
+SadeedDiac + WikiNews eval, gate first-alt WER < 3.5607, stretch <= 2.70).
+
+Sole launch prerequisite: staging the three WO35 arm checkpoints
+(run-047-large-strip/best — also the stage-2 init, run-040-strip-alt/best,
+run-029-news-specialist/best) and the QCRI train_data mount onto HF storage.
+Findings of the audit:
+
+- The legacy Modal volumes (secryst-checkpoints et al.) list only
+  thai/rababa-era artifacts — no run-047/040/029.
+- rhtse/jobs-artifacts holds 78 timestamped job-sync dirs (r8 era) —
+  candidate location, unverified.
+- Recovery options if the checkpoints are gone: (a) the news specialist arm
+  is SHIPPED (ara-diac-news-1.0/1.1) and the r7 canonical teacher is SHIPPED
+  (ara-diac-2.0 lineage) — two of three arms restageable from pinned
+  releases; (b) run-047-large-strip/best (the stage-2 init) may need a
+  short re-derivation from its recorded recipe, or the stage-2 init moves
+  to the closest recoverable ancestor — an owner-visible design decision
+  recorded before any run.
diff --git a/corpora/khm-latn.yaml b/corpora/khm-latn.yaml
index 43379ce..1d982e5 100644
--- a/corpora/khm-latn.yaml
+++ b/corpora/khm-latn.yaml
@@ -1,16 +1,21 @@
 id: khm-latn
 task: transliteration
 sources:
-  # WO88 migration: the secryst-era khm-latn training lineage, recorded
-  # with honest provenance markers — the original corpus predates the
-  # pinning contract. Decode evidence + the khmer-dict-spice source
-  # (Chuon Nath + SBBIC) are documented; the exact training tsv/tsv
-  # generation script lives in the secryst repo
-  # (scripts/export_onnx_byt5.py lineage). UNVERIFIED fields are the
-  # retrain blockers listed in WO68/88.
+  # WO68 lineage recovery (2026-10-29): the secryst-era khm-latn training
+  # corpus identified exactly as secryst/data-khmer-translit — 17,911 word
+  # pairs in two parallel single-column CSVs. Staged to the training bucket
+  # with sha256; the legacy raw.githubusercontent fetch is retired.
+  # Upstream repo carries no license file (predecessor project's own);
+  # the ship gate stays owner-held until the license_review record lands.
   - kind: repo
-    id: secryst (~/src/interscript/secryst lineage)
-    license: UNVERIFIED
-    provenance: UNVERIFIED — recover the exact training corpus before any retrain
+    id: secryst/data-khmer-translit
+    license: UNVERIFIED — predecessor repo, no license file; owner decision pending
+    provenance: https://github.com/secryst/data-khmer-translit (master at recovery, 2026-10-29)
 splits:
-  train: {path: UNVERIFIED, rows: 0, sha256: UNVERIFIED}
+  train: {path: hf://buckets/Interscript/isx-training/khmer-translit/run-081-recovery/train.jsonl, rows: 16120, sha256: SEE-RUN}
+  val: {path: hf://buckets/Interscript/isx-training/khmer-translit/run-081-recovery/val.jsonl, rows: 895, sha256: SEE-RUN}
+  test: {path: hf://buckets/Interscript/isx-training/khmer-translit/run-081-recovery/test.jsonl, rows: 895, sha256: SEE-RUN}
+raw:
+  data_kh.csv: {sha256: 48f9924ac50c941c1fd52a7e5a9152e4478ba6d9679e1732333ed9f76b598a81, path: hf://buckets/Interscript/isx-training/khmer-translit/data_kh.csv, rows: 17911}
+  data_rom.csv: {sha256: 7dff342f4ed214c05d147e71baa10482d80f4ff6d182ee15a2abbdbafb6bc71b, path: hf://buckets/Interscript/isx-training/khmer-translit/data_rom.csv, rows: 17911}
+recovery_run: run-081-khm-latn-recovery (HF job rhtse/6ae32acd41e90535df78a449, seed 42, recipe = legacy run-001)