bundle_v1/README.md

Reproduction bundle v1 — Stable Deep RL via Isotropic Gaussian Representations (arXiv:2602.19373, OpenReview gc7Gg18ejz)

Clean successor to v0 (removed 52 HF download-cache metadata/lock entries under outputs/_dl/, incl. 25 zero-byte files; added integrity manifest, validator, and figure-provenance manifest).

Contents

• scripts/ — all reproduction scripts (Claims 1–4 numerical checks, the self-contained cifar_repro.py GPU grid, claim5_analyze.py, canonical-data builder)
• outputs/ — per-claim raw evidence: 12 GPU-run CSVs + 12 covariance-spectra NPZs (claim5_gpu/), local-check CSVs/JSON, figures (PNG + self-contained HTML)
• poster/ — gate-passing 60×36in poster: poster.html (source), poster_preview.pdf (print), poster_preview.png, poster_embed.html (logbook embed, 9 hotspots), GATE_REPORT.json (all 13 style rules enabled, zero warnings), style_check.json, FIGURE_PROVENANCE.json (every poster figure is REPRODUCTION-GENERATED — none are paper crops; each maps to its generating script, raw evidence, sha256, and natural px)
• FEASIBILITY_AND_EVIDENCE_PLAN.md, DEVIATIONS.md, POSTER_BUILD_NOTES.md

Integrity

• SHA256SUMS.json — deterministic manifest (sorted relpaths → sha256) with entry/byte counts and the credential-scan result
• MANIFEST_DIGEST.txt — sha256 of SHA256SUMS.json (the artifact-version digest)
• validate_bundle.py — re-hashes every entry, flags missing/mismatched/extra files: python3 validate_bundle.py (exit 0 iff clean)

Excluded replaceable inputs (regeneration commands)

• CIFAR-10 canonical 10k subset: hf download ProCreations/isogaussian-drl-repro cifar10k.npz --type dataset or rebuild from the md5-verified canonical tarball via scripts/build_canonical_npz.py
• Paper PDF: https://arxiv.org/pdf/2602.19373v3
• Authors' code: https://github.com/asahebpa/IsoGaussian-DRL

Provenance

• GPU run (12 configs × 100 epochs): https://huggingface.co/jobs/ProCreations/6a587cef85d9643ce16d5eb3 (t4-small, 1h24m, COMPLETED, ≈$0.56; total paid spend incl. cpu-basic dry-run ≈$0.57)
• Raw results mirror: https://huggingface.co/datasets/ProCreations/isogaussian-drl-repro (claim5_gpu/)
• Intermediates: https://huggingface.co/buckets/ProCreations/isogaussian-repro-work