From 6db80927045e6c8ca91df9dd505f07463cc3d73d Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Mon, 20 Jul 2026 06:48:29 -0500 Subject: =?UTF-8?q?RESULT=2054:=20=E4=B8=8A=E7=95=8C=E2=88=9D1/=CF=83?= =?UTF-8?q?=C2=B2=E4=B8=8B=E6=B2=89(=CF=83242=E2=86=92473=E7=BF=BB?= =?UTF-8?q?=E5=80=8D,=CE=B2*=E4=BB=8E0.12=E8=B7=8C=E5=88=B0<0.03),?= =?UTF-8?q?=E5=9B=BA=E5=AE=9A=CE=B2=3D=E5=81=9C=E6=BB=9E=E9=9D=9E=E5=AE=89?= =?UTF-8?q?=E5=85=A8(plain2=E6=94=B6=E5=AE=9889%skip,best=E5=85=B6?= =?UTF-8?q?=E5=AE=9E=E6=98=AF200k=E6=88=90=E7=BB=A9);=20K=3D30=E5=AF=86?= =?UTF-8?q?=E6=8E=A2=E9=92=88=E6=8E=A8=E7=BF=BBCodex=E7=9A=84=CE=B2*=3D0.7?= =?UTF-8?q?(K=3D8=E8=AF=AF=E5=88=A4=E6=85=A2=E5=8F=91=E6=95=A3=E4=B8=BA?= =?UTF-8?q?=E6=94=B6=E6=95=9B);=20=CF=83-scaling=E5=8E=9F=E8=AE=BE?= =?UTF-8?q?=E8=AE=A1=E5=AF=B9=E4=BD=86=E8=A2=ABfloor=E9=92=89=E6=AD=BB+?= =?UTF-8?q?=E6=97=A9=E6=9C=9F=CF=83=E6=9A=B4=E6=B6=A8=E5=BC=A0=E5=8A=9B;?= =?UTF-8?q?=20--beta=5Fcos=5Fmin=E5=AE=9E=E8=A3=85(=E6=8C=89=E8=BF=9B?= =?UTF-8?q?=E7=A8=8Bcosine=E9=99=8D=CE=B2=E8=B7=9F=E8=B8=AA=E4=B8=8B?= =?UTF-8?q?=E6=B2=89=E4=B8=8A=E7=95=8C);=20betacos=E7=8E=8B=E5=86=A02e-2?= =?UTF-8?q?=E2=86=928e-4=E5=9C=A8=E9=A3=9E?= 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 | 24 ++++++++++++++++++++++++ 1 file changed, 24 insertions(+) (limited to 'docs/campaign') diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index f0a82be..6dd93e5 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -568,6 +568,30 @@ direction), not training-under-fault; wave-2 = co-training with faults injected +### RESULT 54 (2026-07-20): THE CEILING SINKS ∝1/σ² AS TRAINING SHARPENS — no fixed β survives; +"fixed β never explodes" was FALSE (it STALLS: endgame skip-storm = frozen training). β must DESCEND. +User caught it: "固定β永不炸个屁！你一直skip还训不训了？" — dead right. +EVIDENCE: plain2 fixed-3e-3 full run — best 3.3437 hit at step 199800, then skip rate SURGES: +84%/step @224k, 89%/step @231k. Endgame (200-234k) = 34k steps of ~zero effective updates; +the "234k crown 3.3437" is really a 200k result. skip ≠ safe, skip = STALL. +DENSE ρ-PROBE (probe_cx3_rhogrid, K=30 asymptotic — corrects Codex's K=8 β*=0.7 which mistook +slow divergence for convergence): true ceiling β* (max β whose resK reaches the fp floor) SINKS: + s35000 β*≈0.12 (σ242) | s95000 β*≈0.12 | s150000 β*≈0.08 (σ443) | s230000 β*<0.03 (σ473). +MECHANISM: training sharpens the model (σ 242→473, ~2×); a sharper block Jacobian makes the nudge +relaxation harder to converge → ceiling drops. It's ∝1/σ^~2. The sunk endgame ceiling (<0.03, +approaching the 3e-3 training β) is WHY fixed 3e-3 skip-storms late. +CODE: σ-scaling EXISTS (line 473: β=β0·SIG0²/σ²) — the ORIGINAL wall-1 design already tracks σ — +but line 477 floor=3e-3 PINS it (plain2 wanted β↓1.94e-3 @σ473, floor forced 3e-3). My beta_simple +"full ownership" that DELETED σ-scaling was exactly backwards. But naive un-flooring fails too: +early σ 3.9→242 (62×, network growing structure NOT sharpening) would crush β∝1/σ² to 8e-7 and +starve training — which is WHY the floor existed. The real tension: early σ-growth (structure, +don't cut β) vs late σ-growth (sharpening, DO cut β); σ magnitude can't separate them. +FIX (least-assumption): SCHEDULE β descent by progress like LR — --beta_cos_min added (cosine +β from --beta to min, bypasses σ-scaling+floor). Calibrated to the measured ceiling: early 2e-2 +(<0.12), endgame 8e-4 (