Research

Two preprints

One audits early-warning signals for grokking. One audits trust signals for neural PDE surrogates against an input-only null model. Both pages carry the figures and the result files the numbers come from.

Preprint2026

A Lead Time Is Not a Detection

Auditing Grokking Predictor Signals

Under a firing rule fixed before looking, the Fourier signals warned before grokking on every seed, weight norm never fired, and effective embedding rank arrived after the event on five seeds out of eight.

8/8Seeds where the Fourier signal family led the test-accuracy jump
2,550 to 2,650Median steps of lead for the three Fourier signals
0/8Seeds where weight norm fired at all

Read the page

Preprint2026

Surrogate Trust Audit

When Can We Trust a Neural PDE Surrogate?

All three trust indicators lost to an input-only null model on the six most shifted cells, with margins of -0.286, -0.694 and -0.077 and every interval below zero.

-0.286PDE residual against the null, 95% interval -0.333 to -0.235
-0.694Seed-ensemble disagreement against the null
-0.077Frozen-feature embedding probe against the null

Read the page