diff options
Diffstat (limited to 'RESULTS.md')
| -rw-r--r-- | RESULTS.md | 14 |
1 files changed, 14 insertions, 0 deletions
@@ -1147,6 +1147,20 @@ 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. +The final bounded no-KP follow-up, Oral-A-v6, tests whether strict stagewise +timing and locally whitened fits repair V5 without increasing its 760-query +budget. The dense readout fit works: feedback/forward cosine reaches +`0.81319` and last-hidden teaching alignment reaches `0.87219`. The +convolutional diagonal fit does not. Correlated child contexts make the +diagonal inverse unstable: finite feedback norm ratios reach `49.08`, parent +field MSE grows from `0.0258` at edge 18 to `6.84e30` at edge 9, and edge 9 +through the input become nonfinite. All forward and BatchNorm state remains +bitwise fixed, and no validation or test endpoint is evaluated. This closes +the present task-loss-only no-KP causal-vectorizer route. A full covariance +solver or local response measurement would be system identification; the +latter is already represented by the response-mirror baseline, while KP is +the substrate behind the positive standard-depth scaling result. + ## 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 <ds> "<depths>" <res> <act> "<seeds>" <ep> <pfx>`. |