README.md

metadata

title: Independent audit
sdk: static
emoji: 🔬
short_description: Executable checks of SGLRW claims
tags:
  - icml2026-repro
  - paper-mtRgrp0u7Z
colorFrom: indigo
colorTo: purple

Robust Stochastic Gradient Posterior Sampling with Lattice Based Discretisation

This logbook audits the six current anchored claims for paper ORID mtRgrp0u7Z from immutable arXiv version 2602.15925v1 and its released TeX/figure source.

The evidence pages contain one decisive VERIFIED or FALSIFIED decision for each anchored claim: VERIFIED for Claims 1, 2 and 3, FALSIFIED for Claims 4, 5 and 6.

Claims 2 and 3 are decided by executing the paper's own experiments rather than by argument:

• code/heavy1d.py reruns the Figure 2 heavy-tailed experiment with the paper's clipped SGLRW implementation (4,000 chains, 2,000 iterations, alpha-stable noise with alpha = 1.5, 5 seeds per cell, two schedules, three mixtures);
• code/blr.py and code/blr_driver.py run the Bayesian linear regression experiment of Table 1 (N=1000, d=20, 2,000 parallel particles, 10,000 iterations, 169 sampler runs across batch sizes, step sizes, designs, initialisations and schedules).

Raw per-run records are in outputs/. Every executed table states its seed count, swept range and error bars.