# Diffusion differentiable resampling — evidence logbook

**Paper.** [Diffusion differentiable resampling](https://openreview.net/forum?id=M0e5XORjAW), Jennifer R. Andersson and Zheng Zhao, [arXiv:2512.10401](https://arxiv.org/abs/2512.10401).

**Claim source.** ICML 2026 challenge claims for `M0e5XORjAW`.

**Evidence type.** Preserved historical numerical output plus post-hoc record validation. No numerical evaluation was rerun while assembling this Space.

## Scope

This logbook presents one stored Gaussian-mixture comparison from the released
[`zgbkdlm/diffres`](https://github.com/zgbkdlm/diffres) implementation at
commit `767effe3e755067eb8a04422597fbf37eb8ab754`. It contains Monte Carlo IDs
`0`–`99`, 10,000 particles, 1,000 sliced-Wasserstein projections,
probability-flow ODE `T=3`, 128 steps, and the Jentzen--Kloeden integrator. The
comparison is Diffusion resampling versus the multinomial baseline only.

## Evidence pages

- [Claim 1: benchmark outperformance](claim-1-gms-comparison/page.md)
- [Claim 2: consistency distribution and proof](claim-2-consistency-theory/page.md)
- [Conclusion and evidence boundary](conclusion/page.md)

## Inspectable artefacts

- [Stored 100-ID record](../evidence/gms-selected-ids-0-99.json)
- [Historic validation record](../evidence/gms-selected-ids-0-99-validation.json)
- [Self-contained record verifier](../evidence/verify_stored_gms_evidence.py)
- [Evidence guide](../evidence/README.md)

The pages state the available evidence and its limits directly. Linked files
allow independent inspection but do not substitute for a numerical rerun,
complete benchmark comparison, significance study, or proof audit.