From 1c89d1b5bee0349aa6a2f39f44ac68e751562a94 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Sat, 25 Jul 2026 13:57:16 -0500 Subject: =?UTF-8?q?RESULT=2061:=20EP=E8=AF=8A=E6=96=AD=E6=9A=82=E5=81=9C,B?= =?UTF-8?q?P=E9=94=9A=E7=82=B9=E4=BC=98=E5=85=88(=E7=94=A8=E6=88=B7?= =?UTF-8?q?=E4=BB=A4)=20=E2=80=94=20b15=E5=B0=B8=E6=A3=80:=E5=90=88?= =?UTF-8?q?=E8=A7=841.5e-3=E8=A1=80=E7=BB=9F=E5=A4=A9=E8=8A=B1=E6=9D=BF?= =?UTF-8?q?=E7=85=A7=E6=B2=89=E5=88=B01.5e-3(=E5=93=A8=E6=88=925=E8=BF=9ED?= =?UTF-8?q?IVERGE),C768=E4=B8=8B=E6=B2=89=E8=BF=BD=CE=B2;=20BP=E7=94=B5?= =?UTF-8?q?=E6=B1=A0=E5=85=A8=E4=B8=89=E5=8D=A1:135M=E5=AD=AA=E7=94=9Fs1/s?= =?UTF-8?q?2(41h)+72M=20s2/s3=E9=93=BE(11h=C3=972)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn --- docs/campaign/CASCADE_ABLATION_PLAN.md | 13 +++++++++++++ 1 file changed, 13 insertions(+) (limited to 'docs/campaign/CASCADE_ABLATION_PLAN.md') diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index 68d7bc2..7c55675 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -568,6 +568,19 @@ direction), not training-under-fault; wave-2 = co-training with faults injected +### RESULT 61 (2026-07-25): EP-SIDE DIAGNOSTICS HALTED FOR THE BP ANCHOR (user call: +"这种情况下应该先跑bp baseline"). b15 postmortem seals the C768 picture: the COMPLIANT +beta=1.5e-3 lineage ALSO sank its ceiling to 1.5e-3 (sentinel: operating row DIVERGE for 5 +consecutive ckpt probes; r18316 backstop halvings; best 3.4643 @387.8k, still short of the +3.3839 first run). The C768 sinking chases beta down — every EP arm so far (3e-3 fixed, +7.5e-4 endgame, 1.5e-3 compliant) fails differently, and NONE of it is interpretable without +knowing what BP does at this width. ATTRIBUTION FIRST. +IN FLIGHT (all 3 GPUs, BP battery): fw135m_bp s1 (GPU0) + s2 (GPU1) — exact EP twins, +440k/2.7B tok, ~41h each; fw72m_bp_s2 -> s3 chained (GPU3, ~11h each) — closes the 72M +honest-gap denominator (currently n=1). Readout: (a) 135M BP endpoint vs 72M BP 3.2884 = +does the RECIPE scale at all (if BP also disappoints, the width problem is shared, not EP's); +(b) honest 135M gap = EP best vs BP twin mean; (c) 72M gap ledger goes n=3-vs-n=2. + ### RESULT 60b (2026-07-24): THE RES METER IS WIDTH-BLIND UNDER AMP — in-band ratcheting is impossible at C768; the instrument is the offline fp32 probe. Ratchet arm killed at launch (6 knocks in the first step, beta -> 4.7e-5 death spiral: gate 1e-3 sat BELOW the bf16 floor). -- cgit v1.2.3