From 774ab32132fe10456191cd839150dd8df506361e Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Mon, 20 Jul 2026 09:28:14 -0500 Subject: =?UTF-8?q?RESULT=2055:=20=E5=8D=95=E4=BE=A7=E5=81=8F=E7=BD=AE?= =?UTF-8?q?=E9=A6=96=E6=AC=A1=E7=9B=B4=E6=8E=A5=E5=AE=9A=E9=87=8F(?= =?UTF-8?q?=E7=BA=BF=E6=80=A7=E2=88=9D=CE=B2,0.3-1%@3e-3,2.7=CF=83;BBP?= =?UTF-8?q?=E7=BA=BF=E7=9C=9F=E7=A9=BA=E7=82=B9=E8=A1=A5=E4=B8=8A);=20?= =?UTF-8?q?=E9=80=90=E6=A0=B7=E6=9C=AC=E7=AC=A6=E5=8F=B7+batch=E5=B9=B3?= =?UTF-8?q?=E5=9D=87=3D=E6=97=A0=E5=81=8F(=E7=94=A8=E6=88=B7=E5=91=BD?= =?UTF-8?q?=E9=A2=98=E5=9B=9B=E5=B1=82=E4=B8=A4=CE=B2=E5=85=A8=E9=AA=8C?= =?UTF-8?q?=E8=AF=81,=E5=81=8F=E7=BD=AE=E6=B6=88=E5=88=B0=E5=99=AA?= =?UTF-8?q?=E5=A3=B0=E5=9C=B0=E6=9D=BF,=E9=9B=B6=E4=BB=A3=E4=BB=B7);=20?= =?UTF-8?q?=E9=A2=84=E7=99=BB=E8=AE=B0=E6=8E=A8=E8=AE=BA:=E8=8B=A5bsign?= =?UTF-8?q?=E6=8C=A1=E4=B8=8D=E4=BD=8F=E4=B8=8B=E6=B2=89=E5=88=99cent?= =?UTF-8?q?=E6=8A=A4=E7=94=B2=E9=9D=9E=E6=B6=88=E5=81=8F?= 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 | 17 +++++++++++++++++ 1 file changed, 17 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 6dd93e5..4cd5f3a 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -568,6 +568,23 @@ direction), not training-under-fault; wave-2 = co-training with faults injected +### 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). +probe_bias @plain2_s35000, 16 batches, beta {3e-3,1e-2}, selfcheck 3e-9: plain rel_bias +|mean(gEP-gBP)|/|mean(gBP)|: w2_b11 0.0035->0.0114 (x3.26 ~ beta ratio 3.33, 2.7sigma over sem +0.0013->0.0042); qkv_b0 0.0106->0.0325; all four layers >sem with ~x3.1-3.4 scaling => the +Laborieux O(beta) term measured directly on a 72M transformer. PSIGN (per-sample random nudge +sign, read re-flipped): bias -> noise floor in EVERY row (w2_b11: 0.0014 vs sem 0.0014), both +betas; sem unchanged vs plain => debiasing is FREE (no extra phase, no variance penalty at B=8; +improves with B — the scale-native estimator, supersedes per-batch bsign in theory). +Batch size alone does NOT debias (bias survives expectation); it only debiases AFTER the +per-sample sign flip makes the O(beta) term batch-antisymmetric — user's "sign+大batch=无偏" +is the correct composition. PRE-REGISTERED corollary: bias is only 0.3-1%, so if the RUNNING +bsign crown fails to stop the ceiling sink (user predicts it will fail), cent's armor is NOT +debiasing — it must be a variance/geometry property of the symmetric read; new puzzle. +In flight: fw72m_bsign (GPU1/3, per-batch sign), fw72m_ratchet (GPU0 queued, halve-only). + ### 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. -- cgit v1.2.3