---
title: "DiffRes: traceable 100-ID Gaussian-mixture study"
emoji: "🔬"
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
tags:
  - icml2026-repro
  - paper-M0e5XORjAW
  - reproducibility
  - diffusion
---

# DiffRes: traceable 100-ID Gaussian-mixture study

This is a static, local publication package for **Diffusion differentiable
resampling** by Jennifer Rosina Andersson and Zheng Zhao
([arXiv:2512.10401](https://arxiv.org/abs/2512.10401),
[OpenReview:M0e5XORjAW](https://openreview.net/forum?id=M0e5XORjAW)).

**Evidence statement.** A local study wrapper using released implementation,
stored numerical random keys and pinned source config recorded a traceable 100-ID
Gaussian-mixture result; a separate validator recomputed raw-record aggregates,
paired intervals, and source hashes.

The files here are **historic stored outputs**, transferred without rerunning
the numerical evaluation. This static site has no execution, upload, or model
hosting component.

The judge-readable, claim-level record is in the
[evidence logbook](pages/index.md). It restates the concrete conditions,
numbers, and limits below in Markdown because the challenge judge reads those
pages rather than the static landing page or linked JSON files.

## What the evidence covers

- Released implementation: [`zgbkdlm/diffres`](https://github.com/zgbkdlm/diffres)
  at commit `767effe3e755067eb8a04422597fbf37eb8ab754` (MPL-2.0).
- One paper-selected Gaussian-mixture configuration, with stored 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.
- A paired comparison between diffusion resampling and the multinomial
  baseline only.
- A post-hoc, standard-Python validator that checks strict JSON, the 100 raw
  records, reported aggregates, paired intervals, the recorded source pin, and
  hashes of the released source files.

| Historical stored aggregate (lower is better) | Diffusion resampling | Multinomial baseline |
| --- | ---: | ---: |
| Sliced Wasserstein L1, mean ± population SD | 0.0807686 ± 0.0212899 | 0.0824334 ± 0.0250323 |
| Squared posterior-mean residual L2, mean ± population SD | 0.0373908 ± 0.0299358 | 0.0378330 ± 0.0442877 |

The stored paired 95% t intervals cross zero for both metrics. They are not
evidence of a statistically significant improvement.

## Evidence files

- [Raw 100-ID result](evidence/gms-selected-ids-0-99.json) — preserved
  per-ID values, source provenance, configuration, and recorded aggregates.
- [Post-hoc validation record](evidence/gms-selected-ids-0-99-validation.json)
  — the historic validator output.
- [Validator source](evidence/validate_gms_selected_result.py) — copied
  byte-for-byte as an audit artefact; it does not rerun the JAX evaluation.
- [Self-contained stored-evidence verifier](evidence/verify_stored_gms_evidence.py)
  — a standard-library-only check of the JSON structure and recorded
  arithmetic; run it with `python3 evidence/verify_stored_gms_evidence.py`.
- [Evidence guide](evidence/README.md) — lineage, portability, and limits.
- [Source and artefact manifest](MANIFEST.md) and
  [package SHA-256 manifest](MANIFEST.sha256).

## Limits

This is **not** an exact reproduction, an independent replication, a complete
Table 1 or Table 2 reproduction, a publication score, or a broad
outperformance claim. It does not evaluate the OT, Gumbel-Softmax, or
soft-resampling baselines, and it does not rerun the numerical study in this
package. The included validator recomputes historic raw-record statistics and
checks the recorded source identity; it is not a second numerical run. The
self-contained verifier recomputes the stored record's configuration, ordered
IDs, aggregates, and lower-counts without a source checkout or SciPy. It does
not validate source hashes, execution provenance, paper-level results, or any
comparative claim.

## Safety and selection note

The selected material is a benign numerical-methods study package: static
HTML/Markdown, two JSON evidence records, a non-networked validator, and
licensing/provenance files. The transfer excludes the upstream source tree,
virtual environments, caches, runners, model artefacts, partial studies, and
all LGSSM material. A pre-transfer content scan found no private local paths,
credential-shaped values, upload clients, or source files from excluded study
areas. The validator's only generic token-pattern match is a JSON-parser parameter,
not credential handling.

## Licence and attribution

The package-authored wrapper material is under the [MIT Licence](LICENSE).
The upstream DiffRes implementation is not redistributed here; it remains
available under MPL-2.0, with attribution in [NOTICE](NOTICE) and a copy of
the upstream licence at [UPSTREAM-MPL-2.0.txt](UPSTREAM-MPL-2.0.txt).