From 15cd914b303a99270f9ff01852e66ed3701f52af Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Fri, 4 Sep 2026 13:44:36 +0200 Subject: [PATCH] =?UTF-8?q?docs:=20TODO.publish-client=2001=20closed=20?= =?UTF-8?q?=E2=80=94=20G2b=20verdict=20recorded?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.publish-client/01-g2b-label-scale.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TODO.publish-client/01-g2b-label-scale.md b/TODO.publish-client/01-g2b-label-scale.md index 130c758..fea1d53 100644 --- a/TODO.publish-client/01-g2b-label-scale.md +++ b/TODO.publish-client/01-g2b-label-scale.md @@ -2,4 +2,4 @@ The domain-residual causal test: build units from the cleaned tashkeela-full corpus, relabel with the r7 teacher, distill at the G2a recipe. Gate: >=0.5pp full-set gain attributes to domain. [models+papers] -Status: pending (2026-09-02). +Status: DONE (2026-09-04). Full-set verdict 4.8231 [2.194,2.554] vs G2a 4.5701 [1.91,2.35] — flat-negative, gate FAILED. With E6's swap-negative, the domain-coverage attribution is rejected in both directions; residual reframes as corpus-unreachable (on-policy GKD is the live lever). RESULTS.md PR #170, Paper B cell PR #171.