# Paper and proof audit

The direct paper source used here is the author-released arXiv v2 TeX package, pinned by archive and `main.tex` hashes in `evidence/pins.json`. Its title and author list match the current official OpenReview PDF result for `l35QweVxgn`; direct OpenReview downloads were challenge-gated from this machine. The TeX describes one-hidden-layer quadratic neural networks trained with gradient descent on sequential XOR-cluster tasks with Gaussian noise.

The static source check found explicit statements corresponding to the active claims: a train-time-forgetting theorem under stated `n`, `m`, and `eta T` scalings; a uniformly small train-error theorem; delayed-generalisation-gap results; and a self-bounded-loss refinement. The source also explains test-time forgetting as the sum of train-time forgetting and delayed generalisation gap.

This is a textual cross-reference, not a fresh proof verification. In particular, constants, hidden logarithmic factors, probability conditions, theorem assumptions, and the extension from stylised XOR data to broader settings require an independent mathematical review before any claim can be upgraded.