From c43512cdcbc4c7667038089138615f1d3c17cd68 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Tue, 21 Jul 2026 03:05:08 -0500 Subject: =?UTF-8?q?RESULT=2056:=20bsign=E7=8E=8B=E5=86=A0=E5=B0=81?= =?UTF-8?q?=E5=8D=B03.3349=20=E2=80=94=201.0=C3=97=E7=8E=8B=E5=86=A0?= =?UTF-8?q?=E5=AD=98=E5=9C=A8;=E5=85=A8=E7=A8=8B=E9=9B=B6guard=E4=BA=8B?= =?UTF-8?q?=E4=BB=B6(vs=20plain2=2012840=E6=AC=A1skip=E9=A3=8E=E6=9A=B4),?= =?UTF-8?q?=E4=B8=8Ecent=203.3318=E5=B9=B3=E6=89=8B(0.003=E5=86=85)?= =?UTF-8?q?=E4=BD=86=E6=88=90=E6=9C=AC60%;=E4=B8=8B=E6=B2=89=3D=E5=81=8F?= =?UTF-8?q?=E7=BD=AE=E9=A9=B1=E5=8A=A8=E5=9B=A0=E6=9E=9C=E9=97=AD=E7=8E=AF?= =?UTF-8?q?(=E7=AC=A6=E5=8F=B7=E7=BF=BB=E8=BD=AC=E6=9C=9F=E6=9C=9B?= =?UTF-8?q?=E6=B6=88=E5=81=8F=E5=8D=B3=E6=A0=B9=E6=B2=BB);=E7=94=A8?= =?UTF-8?q?=E6=88=B7'=E6=B2=A1=E7=94=A8'=E8=B5=8C=E6=B3=A8=E8=B4=A5?= =?UTF-8?q?=E5=8C=97;s2=E5=A4=8D=E9=AA=8C=E5=B7=B2=E5=8F=91=E5=B0=84;psign?= =?UTF-8?q?=3D=E7=94=9F=E4=BA=A7=E5=BD=A2=E6=80=81=E5=80=99=E9=80=89?= 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 | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index 4cd5f3a..21661f7 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -568,6 +568,27 @@ direction), not training-under-fault; wave-2 = co-training with faults injected +### 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 +the FULL 234k schedule with ZERO guard events end to end (d-telemetry 0 at every step, under +the STRICTER res-alone gate in silent mode) — where plain2 (identical recipe minus the sign +flip) skip-stormed from 217k (84-89%/step, 12840 events, training frozen after 199.8k). +LADDER: cent 3.3318 (1.67x) | bsign 3.3349 (1.0x) | plain2 3.3437 (1.0x, stalled@200k) | +c190 hybrid 3.3732. bsign TIES the centered crown within 0.003 (noise) at 60% of its cost. +best@203.7k — the SAME dip step as cent's best; endgame keeps taking updates (val ~3.46 at +234k, healthy oscillation, final gate cos 0.9958). sigma trajectory identical to plain/plain2 +(475 endgame) => sigma exonerated again; the causal chain closes: single-sided systematic bias +(measured 0.3-1%@3e-3, RESULT 55) -> trajectory steering -> ceiling collapse; remove the bias +in EXPECTATION (per-step sign flip) and the collapse never comes. The entire controller saga's +answer was a coin flip on beta's sign. Note vs BP: bsign gap 0.047 vs cent's 0.043 — cent +remains the flagship gap number; bsign is the cost claim. +NEXT: fw72m_bsign_s2 LAUNCHED (seed discipline — three single-seed crowns have been retired +this campaign; no external claim before n>=2). PSIGN (per-sample sign, RESULT 55-verified +unbiased, batch-averaged) = the scale-native production form, candidate recipe for the 300M +stage. Ratchet arm continues as the no-debiasing control (its staircase should track the +bias-driven sink that bsign no longer has). + ### RESULT 55 (2026-07-20): SINGLE-SIDED BIAS DIRECTLY QUANTIFIED (the BBP-thread gap) — real, linear in beta, 0.3-1% of grad norm @3e-3; PER-SAMPLE SIGN + BATCH AVERAGE = UNBIASED (user's claim VERIFIED at zero measurable cost). -- cgit v1.2.3