From 7794ea33c2330beb49543b192aa6a3287255f27b Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Mon, 19 Oct 2026 06:14:23 +0800 Subject: [PATCH] =?UTF-8?q?docs(wO47):=20FINAL=20=E2=80=94=20oracle=20dist?= =?UTF-8?q?illation=20misses=20(correlated=20arms);=20run-077/best=20ships?= =?UTF-8?q?=20as=20fas-diac-base-1.0?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.sota/47-fas-diac-release.md | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/TODO.sota/47-fas-diac-release.md b/TODO.sota/47-fas-diac-release.md index 5f29105f..548fd26d 100644 --- a/TODO.sota/47-fas-diac-release.md +++ b/TODO.sota/47-fas-diac-release.md @@ -130,3 +130,19 @@ remains owner-gated: accept 0.603% row-mean (0.777% corpus) as the canonical fas-diac number and re-anchor the recipe gate, or commission the corpus-rebuild arm (recover the ~18% of rows the derivation dropped vs the raw 528,837) before shipping. + +## FINAL (2026-10-19): run-077/best is fas-diac-base-1.0 + +The improvement arm (run-081-fas-stage2: RC continue-trained on the +run-080 oracle targets) MISSED — ep0 0.649% / ep1 0.631% vs RC 0.603% +row-mean. Root cause of non-transfer: the fas arms are too correlated +(10.0% disagreement, p077 picks 97.3%) — the oracle targets were ~90% +the RC's own outputs, so the +16% oracle gain had no new signal to +distill. Contrast ara: 31.4% disagreement made its oracle worth +training on. + +Five levers now tested and closed with data: epochs, backbone, +optimizer recipe, metric convention, oracle distillation. Remaining +untried (recorded, not scheduled): corpus rebuild to the raw 528,837 +rows. **Ship: run-077/best — 0.603% row-mean / 0.777% corpus / +82.33% word-acc on the 9,288 docs split, CC0 data.**