From 68ba9c7e91a352c8d81b40f5e80067b52cce4dd6 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Wed, 29 Jul 2026 20:44:09 -0500 Subject: =?UTF-8?q?RESULT=2073:=20=E6=B2=BB=E7=90=86=E6=9D=A1=E4=BB=A4(?= =?UTF-8?q?=E7=94=A8=E6=88=B7=E8=A3=81=E5=86=B3)=20=E2=80=94=20=E6=AD=A3?= =?UTF-8?q?=E5=90=91scaling=20claim=E6=9A=82=E5=81=9C;=E4=B8=89=E7=B1=BB?= =?UTF-8?q?=E4=B8=A5=E6=A0=BC=E5=88=86=E7=B1=BB(=E5=86=BB=E7=BB=93?= =?UTF-8?q?=E9=85=8D=E6=96=B9=E6=9B=B2=E7=BA=BF/=E4=BA=8B=E5=90=8Eoracle?= =?UTF-8?q?=E5=8C=85=E7=BB=9C/=E8=BF=81=E7=A7=BB=E6=9B=B2=E7=BA=BF);dg128?= =?UTF-8?q?=3D=E5=8F=A6=E4=B8=80=E6=9B=B4=E6=96=B0=E5=9C=BA+exploratory?= =?UTF-8?q?=E9=99=8D=E6=A0=BC;=E6=9C=BA=E5=88=B6=E8=AF=86=E5=88=AB?= =?UTF-8?q?=E5=8D=8F=E8=AE=AE(=E7=BA=BF=E6=80=A7=E5=8C=96=E7=94=B5?= =?UTF-8?q?=E6=B1=A0+=E6=97=A0=E9=87=8F=E7=BA=B2=E5=BE=8B+C1152=E7=9B=B2?= =?UTF-8?q?=E6=B5=8B)=E9=A2=84=E6=B3=A8=E5=86=8C=E5=86=BB=E7=BB=93?= 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 | 32 ++++++++++++++++++++++++++++++++ 1 file changed, 32 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 8d0ed5c..c29d476 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -578,6 +578,38 @@ power law (dose ~ width^k -> soft wall, budget k for 270M) vs width THRESHOLD ne 0.076 gap hides a small closable leak (lk512 arms answer directly). Chained behind fw135m_dg128 on all three GPUs; ~12h of ladder after the validation completes. +### RESULT 73 (2026-07-29): GOVERNING DOCTRINE (user ruling) — NO TRANSFERABLE SCALING +KNOWLEDGE EXISTS YET; positive scaling claims SUSPENDED; results reclassified into three +strict classes; next phase = mechanism identification with a pre-registered transfer law and +a BLIND holdout. All prior "progress" on the C768 leak is symptom-and-intervention, not root +cause: amp (demoted), head throttle (retracted), softmax saturation (refuted) — a chain of +per-scale patches with no unifying mechanism = an invalid scaling experiment by definition. +THE THREE CLASSES (every past and future number must declare its class): +1. FROZEN-RECIPE CURVE (valid, negative): the fixed recipe fails C512->C768 (gap 0.047 -> + 0.294); small-width frozen gaps already trend with width. This is the one established fact. +2. POST-HOC ORACLE ENVELOPE (repairability only, ZERO extrapolation power): per-width tuned + results incl. every leak-battery arm and dgain dosing. dg128 at cos 0.25-0.28 is NOT the + original estimator made precise — it is a DIFFERENT update field; if it works it is a new + algorithm candidate and may not silently rejoin the original method's scaling curve. + fw135m_dg128 (running, left to complete per user) is EXPLORATORY regardless of outcome — + also a 250k-resume, so at best it proves rescue-ability of one state. +3. TRANSFER CURVE (the only thing called scaling): a control law with no explicit size + dependence, fitted at small widths, hitting an UNSEEN width with ZERO tuning. +THE LAW MUST SIMULTANEOUSLY EXPLAIN: top-half localization; the 4-6x C768/C512 leak-rate +factor; leak ~ 1/beta; the x128 displacement requirement; immunity to bsign/centered/fp32/ +K8/head-LR; complete closure under BP substitution. +PROTOCOL (frozen before any validation run): C256/C384/C512/C768 for mechanism ID only — +per-layer normalized displacement RMS(d_l)/RMS(z_l), Jacobian/curvature at each nonlinearity +family, true optimizer-step differences; LOCAL LINEARIZATION battery per nonlinearity class +(softmax vs SwiGLU vs RMSNorm — substitute frozen local linearization in the read, see which +substitution kills the leak = which family carries the width-growing term); fit the +dimensionless law (e.g. gain from RMS(d)/RMS(z) x local curvature scale — never a hand-written +per-width constant); C1152/270M = blind holdout, formula/thresholds/cost/error-tolerance +frozen BEFORE launch, from-scratch, n>=2 seeds; any unpredicted phenomenon at the holdout = +verdict "does not scale", no patch-and-reconnect. +Project page reworded same day: 135M row states the frozen-recipe failure plainly; scaling +section = negative result + protocol; "97% treatment" claim removed from the abstract. + ### RESULT 72 (2026-07-29): THE CURVE GOES ALL THE WAY — dg128 closes 97% (tail within 0.0015 of the all-BP bar); the physics leak is FULLY REDEEMABLE by displacement amplification. FULL-DOSE VALIDATION LAUNCHED. -- cgit v1.2.3