ProCreations/icml-2026-mds-falsification-trace

Codex trace: Minimum Distance Summaries falsification audit

This public trace documents the focused claim-complete upgrade of Minimum Distance Summaries for Robust Neural Posterior Estimation. It covers primary-source inspection, pinned author-code review, real 32×32 HSP90 cryo-EM simulation, a 71,077-parameter Gaussian NPE, 480 adaptations, the OC-SVM counterexample, two byte-identical executions, static-bundle verification, and exact-SHA judge discovery.

Falsification result

The registered compound claim says MDS outperforms NPE, NNPE, and NPE-OR across the contamination range. The independent rerun finds that MDS beats plain NPE and NNPE at 10–40% contamination, but the pinned OC-SVM NPE-OR baseline has lower posterior-mean RMSE at 20%, 30%, and 40%. The corrected claim is therefore narrower: MDS improves robustness over unfiltered NPE and NNPE, but does not dominate NPE-OR across the full tested range.

Provenance and privacy review

• —Native source: a Codex JSONL session, focused UTC interval 2026-07-20T09:29:04Z–2026-07-20T10:23:07Z, plus the single later exact-verdict observation at 2026-07-20T12:21:14Z.
• —Scientific result SHA-256: 019d789e8139e8044eac43f09c0d7662dff550e30ae8d1fa2eb37ed9d2d6f67e.
• —Published Space SHA: d378aa9071e7fa13e0dc0afcaf876dfebde36c79.
• —Privacy edits: encrypted/private reasoning, internal policy and token telemetry, and unrelated persistent-goal boilerplate were removed. Absolute local paths were replaced with <WORKSPACE> and <USER_HOME>. User and assistant task messages, tool inputs, tool outputs, scientific results, and timestamps were otherwise retained. The later multi-Space judge-status result was reduced to its MDS row only.
• —Automated scans reject Hugging Face, GitHub, AWS, and bearer-token patterns before the trace is emitted.

The public file mds-falsification-codex.jsonl opens in Hugging Face's native agent-trace viewer.