From 98d5ea09d1498f802f2ad06f5f67b385e228f063 Mon Sep 17 00:00:00 2001 From: Yuren Hao Date: Tue, 28 Jul 2026 13:20:34 -0500 Subject: =?UTF-8?q?RESULT=2066:=20head=E8=A2=AB=E8=8A=82=E6=B5=81=E5=88=B0?= =?UTF-8?q?60%(=E4=BD=8D=E7=A7=BB=E5=BE=AE=E6=8E=A2=E9=92=88:=E5=85=A8?= =?UTF-8?q?=E6=97=8F1.00-1.02=E5=94=AFW=5Fout=200.60)=20=E2=80=94=20?= =?UTF-8?q?=E7=94=A8=E6=88=B7=E7=AD=89=E4=BB=B7LR=E5=81=87=E8=AF=B4?= =?UTF-8?q?=E5=91=BD=E4=B8=AD=E5=B9=B6=E5=AE=9A=E4=BD=8D=E5=88=B0=E5=8D=95?= =?UTF-8?q?=E7=9F=A9=E9=98=B5;=20Adam=E5=B2=9Bm/=E2=88=9Av=E8=8A=82?= =?UTF-8?q?=E6=B5=81,=E5=AF=B9centered/fp32/K8=E5=85=A8=E4=B8=8D=E6=95=8F?= =?UTF-8?q?=E6=84=9F=3DEP=E5=A4=B4=E8=AF=BB=E5=85=B1=E4=BA=AB=E7=BB=93?= =?UTF-8?q?=E6=9E=84;=20--head=5Flr=5Fmult=E5=AE=9E=E8=A3=85;=20=E7=94=B5?= =?UTF-8?q?=E6=B1=A02(headmix=E8=AF=8A=E6=96=AD+hlr1.67/2.5=E8=A1=A5?= =?UTF-8?q?=E5=81=BF)=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 | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) (limited to 'docs') diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index 277ebf1..5c4b25c 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 66 (2026-07-28): THE HEAD IS THROTTLED TO 60% — user's equivalent-LR hypothesis +CONFIRMED and localized to ONE matrix; symmetrization/precision/depth all irrelevant to it. +Displacement micro-probe (600 steps from s250000, ||dW||/||W|| per family): every family +IDENTICAL EP-vs-BP (ratio 1.00-1.02) EXCEPT W_out: EP 0.0068 vs BP 0.0114 = 60%. Muon-managed +blocks are magnitude-immune (msign); W_out lives on the Adam island where step ~ m/sqrt(v) — +the EP head-read's per-coordinate temporal coherence is lower, v inflates ~2.8x, Adam throttles +the loss-facing matrix by 40%. Explains R44 (gap lives head-side), the anneal freeze (head +can't track features), and width-scaling (louder states -> stronger throttle). +BATTERY 1 (all closed ~0% of the 0.0685 leak): centered 0%, fp32 +3%, K8 (arm died, rerun +pending) — the throttle source is INVARIANT to estimator symmetry, precision, and relax depth: +it is the shared structure of the EP W_out read, not removable noise. (cos flat 0.986-0.988 +through the losing segment; EP gn 0.15, clip never fires — user's checklist all measured.) +FA-argument (user): static direction quality exonerated a priori — FA learns at cos 0.3. +BATTERY 2 IN FLIGHT: leak_headmix (--bpmix head = the head's share of the leak, diagnostic +ceiling), leak_hlr17 / leak_hlr25 (--head_lr_mult, the zero-cost compensation candidate). +If hlr claws back what headmix shows, the width fix is ONE hyperparameter. + ### RESULT 65 (2026-07-28): THE LEAK INSTRUMENT WORKS — first reading +0.0685/6k at the SAME state (20x the trajectory-averaged rate: front-loaded/self-adapting), screening battery 1 in flight. Paired 6k segments from fw135m_bsign_s250000: BP tail val_mean 3.4982 vs EP-plain -- cgit v1.2.3