From ffaa3695da6f00a32d7b5636cff2e5faac41e6a8 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Thu, 6 Aug 2026 14:34:10 -0500 Subject: results: close no-KP causal bootstrap screen --- RESULTS.md | 15 +++++++++++++++ 1 file changed, 15 insertions(+) (limited to 'RESULTS.md') diff --git a/RESULTS.md b/RESULTS.md index 6da95a0..9bf803e 100644 --- a/RESULTS.md +++ b/RESULTS.md @@ -1132,6 +1132,21 @@ its survival under the four-times-RMS mixed-traffic intervention. `results/figs/figure6_standard_depth_scaling.{pdf,png}` is regenerated from all 60 records and verifies their hashes plus the historical training source. +The separate Oral-A-v5 feasibility branch asks whether scalable feedback can +be learned without KP or forward-weight mirroring. It replaces V4's one +batch-level scalar for all 267,904 feedback parameters with 128 independent +per-example node-perturbation observations for one hierarchical edge at a +time. At a matched 760-query frozen-forward budget, the run is finite, +preserves every forward/BatchNorm/optimizer tensor bitwise, and keeps feedback +norm ratios in `[0.8778,1.0036]`. All-layer alignment improves from fixed +HFA's `0.011279` to `0.086687`, with late-layer alignments as high as `0.3837`. +The preregistered gate nevertheless fails: early-third alignment is only +`0.001147` versus the required `0.10`, and mean feedback/forward cosine is +`0.013162`. No accuracy run is opened. This is evidence that the causal +observations are informative near the output, not evidence that the no-KP +vectorizer scales; the existing positive ResNet scaling result still belongs +to the inherited KP substrate. + ## How to run `experiments/run.py --mode {bp,fa,dfa,sdil} --dataset {mnist,fmnist,cifar10} --depth D --residual {0,1} --act {tanh,gelu,silu,relu}` Batteries: `experiments/run_v2.sh "" "" `. -- cgit v1.2.3