# Claim 1 — narrow NTK condition-number outcome

## Legacy judge claim

> GLU structure reshapes the NTK spectrum, leading to smaller condition number and more compact eigenvalue distribution.

## Release status

**Narrow recorded outcome only.** The retained result is from the authors’ pinned `scripts/ntk/ntk_eigen_mlp.py` source script, executed on CPU with its fixed seed and 50 standard-Gaussian inputs. It printed:

| Activation | Condition number |
| --- | ---: |
| ReLU | 40.29 |
| ReGLU | 3.68 |

The package retains the script, source-tree, environment, and result-output hashes in [`../evidence/run-evidence.json`](../evidence/run-evidence.json), not the source checkout or full log.

The active judge input is anchored and overrides this legacy claim set; see [`judge-evidence.md`](judge-evidence.md). This release status remains deliberately narrower than the active anchored theory claims.

## Boundary

This is a finite empirical-NTK measurement under one released-script configuration. It does not establish a general GLU result, prove the paper’s asymptotic analysis, quantify the full eigenvalue distribution’s compactness, or reproduce training behaviour.