From bc0da2be225e9555f2bc781e1e640bceb14d5e14 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Sat, 5 Sep 2026 06:06:41 +0200 Subject: [PATCH] =?UTF-8?q?docs:=20item=2004=20closed=20=E2=80=94=20Hebrew?= =?UTF-8?q?=20layerdrop=20negative?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TODO.publish-client/04-hebrew-layerdrop-replication.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TODO.publish-client/04-hebrew-layerdrop-replication.md b/TODO.publish-client/04-hebrew-layerdrop-replication.md index fc830fd..fbb9ecf 100644 --- a/TODO.publish-client/04-hebrew-layerdrop-replication.md +++ b/TODO.publish-client/04-hebrew-layerdrop-replication.md @@ -2,4 +2,4 @@ The capacity law cross-lingually: layerdrop on the Hebrew teacher, same protocol. [papers] -Status: pending (2026-09-02). +Status: DONE (2026-09-05). NEGATIVE: the depth-cut collapsed on Hebrew — 77.48 DER vs full-depth 30.38 (+53.77pp [51.64,55.92]), teacher reproduced 23.72. Confound noted (Arabic = sequence-KD+Muon+6ep; Hebrew = logit-KD+3ep): the counterexample may be depth-cut-under-logit-KD, but the cross-lingual universal claim is closed. RESULTS.md PR #179; Paper B scoping PR #180.