From 014150378d8e8cd90c2b1a01d02a2f860379dbd5 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Tue, 21 Jul 2026 11:44:21 -0500 Subject: =?UTF-8?q?RESULT=2057:=20=E6=A3=98=E8=BD=AE=E5=88=A4=E5=86=B3=3D?= =?UTF-8?q?=E7=BA=AF=E9=80=80=E8=AE=A9=E8=B7=91=E4=B8=8D=E8=B5=A2=E5=81=8F?= =?UTF-8?q?=E7=BD=AE=E9=A9=B1=E5=8A=A8=E7=9A=84=E4=B8=8B=E6=B2=89(?= =?UTF-8?q?=E9=98=B6=E6=A2=AF131k=E8=B5=B7=E9=9B=AA=E5=B4=A94=E4=B8=AA?= =?UTF-8?q?=E9=87=8F=E7=BA=A7=E5=88=B0bf16=E5=9C=B0=E6=9D=BF,=E8=AE=AD?= =?UTF-8?q?=E7=BB=83=E6=AD=BB=E8=B5=B0,181k=E5=A4=84=E5=86=B3);=20?= =?UTF-8?q?=E5=90=8C=E4=B8=80=E9=97=A8=E4=B8=8Bbsign=E9=9B=B6=E4=BA=8B?= =?UTF-8?q?=E4=BB=B6=3D=E4=B8=8B=E6=B2=89=E6=98=AF=E5=81=8F=E7=BD=AE?= =?UTF-8?q?=E4=BC=B0=E8=AE=A1=E5=99=A8=E7=9A=84=E8=BD=A8=E8=BF=B9=E6=80=A7?= =?UTF-8?q?=E8=B4=A8=E8=80=8C=E9=9D=9E=CE=B2=E6=94=BF=E7=AD=96=E9=97=AE?= =?UTF-8?q?=E9=A2=98;=20=E9=98=B6=E6=A2=AF=3D=E9=A6=96=E4=B8=AA=E8=BF=9E?= =?UTF-8?q?=E7=BB=ADin-vivo=E5=A4=A9=E8=8A=B1=E6=9D=BF=E8=BD=A8=E8=BF=B9?= =?UTF-8?q?=E6=B5=8B=E9=87=8F;=20GPU0=E8=BD=ACBP=20s2=E8=A1=A5=E8=AF=9A?= =?UTF-8?q?=E5=AE=9Egap?= 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 | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) (limited to 'docs') diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index 21661f7..d36ed9c 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -568,6 +568,25 @@ direction), not training-under-fault; wave-2 = co-training with faults injected +### RESULT 57 (2026-07-21): RATCHET VERDICT — pure retreat cannot outrun the bias-driven +sink; the staircase IS the plain-lineage ceiling trajectory, and it chases beta to the bf16 +floor. The trio closes: debiasing is the only 1.0x recipe that holds. +fw72m_ratchet (user design: halve-on-crossing only, no climb, plain estimator, res-alone gate, +guards silent): mechanically flawless — zero explosions, knocks fired exactly as designed. +Staircase: 3e-3 held to 88.4k (first knock -> 1.5e-3), held to 131.2k, then CASCADE (5 knocks +in 3.2k steps -> 4.7e-5 by 134.4k), then 2.3e-5@140k, 5.9e-6@152k, 1.8e-7@154k = sub-bf16- +cliff, training dead-walks (val 3.42 best -> 4.28 noise-walk). KILLED at 181.4k (staircase +complete; remaining steps carried no information). GPU0 -> fw72m_bp_s2 (honest-gap seed). +READING: under the SAME res-alone gate, bsign saw ZERO events in 234k while ratchet's ceiling +collapsed through 4+ decades from 131k. Same estimator-family difference as plain2-vs-bsign, +now in the adaptive-beta version: the sink is a property of the BIASED estimator's trajectory, +not of the beta policy. No beta policy — fixed (plain2), descending schedule (would-be betacos), +or measured retreat (ratchet) — survives it; removing the bias (bsign) removes the sink. +Bonus instrument value: the staircase is the first continuous in-vivo ceiling-trajectory +measurement (a plain lineage's ceiling vs step curve, free of controller-climb artifacts). +LADDER STANDS: cent 3.3318 (1.67x) | bsign 3.3349 (1.0x, zero events) | plain2 3.3437 (stall) +| ratchet 3.4200 (stall@~131k). 135M gate now = bsign_s2 confirmation only (102k/234k, clean). + ### RESULT 56 (2026-07-21): BSIGN CROWN SEALED 3.3349 — THE 1.0x-COST CROWN EXISTS; the sink is BIAS-DRIVEN (user's "no effect" bet resolved AGAINST, happily). Sign randomization alone (Scellier-Bengio random nudge sign, one flag, zero extra phases) held the corridor for -- cgit v1.2.3