Claim 1 — narrow recorded measurement
The retained CPU, 50-sample Gaussian empirical-NTK output reports condition numbers of 40.29 for ReLU and 3.68 for ReGLU. This supports only that recorded direction in the pinned source script’s setting.
Local release candidate. This package must receive independent review before any Space is created or uploaded.
GLU/NTK reproduction evidence
Evidence-bounded material for OpenReview w0JhOFWPJl, The Devil is in the Condition Numbers: Why is GLU Better than non-GLU Structure?.
Official-claim status
Claim 2 — unverified
No retained evidence measures optimisation, stochastic error, or generalisation. This claim is therefore not reproduced.
Verification and pins
Use the offline release verifier. It checks the card tags, page evidence, hash manifest, numerical boundary, and package boundaries. The active judge-input merge is recorded in judge evidence; immutable source and paper pins are in pins.json.