release/v1/docs/WORKFLOW.md

Chebyshev Policies reproduction

This directory contains the independent, claim-by-claim reproduction for ICML 2026 paper 12142, OpenReview aNWIVNjocB, arXiv 2605.22305.

The driver fetches the authors' two repositories at immutable commits and verifies both commit hashes before importing them. The upstream repositories do not declare a license, so they and their bundled model weights are not vendored or republished here. Only the reproduction driver, configs, raw measurements, derived tables, figures, manifests, and newly trained coefficients are part of this bundle.

Pinned inputs

• Paper code: JRC-ISIA/paper-2026-chebyshev-policies-low-dimensional-control-tasks at fad6ebb46f976db29eb5ba509622dc3c24b7ac5f
• Polynomial SB3 code: JRC-ISIA/polynomial-sb3-rl-agents at ed945331ce989592d75af9e8410823ee007e13a8
• Paper: arXiv v4, 2605.22305v4, dated 2026-07-08

Workflow

configs/*.json
  -> scripts/run_reproduction.py
  -> runs/raw/<batch-id>/
  -> scripts/build_tables.py
  -> runs/tables/
  -> scripts/make_figures.py
  -> runs/reports/
  -> scripts/analyze_claims.py
  -> runs/analysis/

Every raw batch is append-only: the runner refuses to reuse an existing batch_id. Failed units remain in records.jsonl. Derived tables, analyses, and figures are regenerated deterministically; their manifests carry input hashes, the builder SHA-256, and the exact build command.

Commands

Local smoke test:

uv run --python 3.10 scripts/run_reproduction.py \
  --config configs/smoke.json \
  --batch-id local-smoke-YYYYMMDD-HHMMSS \
  --output-root runs/raw

Full configurations are split so each empirical claim has a clear Hugging Face Job URL and cost:

uv run --python 3.10 scripts/run_reproduction.py --config configs/claim23_full.json ...
uv run --python 3.10 scripts/run_reproduction.py --config configs/claim4_full.json ...
uv run --python 3.10 scripts/run_reproduction.py --config configs/claim5_pendulum_full.json ...

On Hugging Face Jobs, use scripts/run_hf_job.sh CONFIG_FILENAME BATCH_ID. It stages append-heavy evidence on local storage, validates the completed manifest, publishes only closed files to the Bucket mount, and requires a byte-for-byte tree comparison before returning success.

The release allowlist contains exactly these three v4 batch identities. Admission still requires each directory to pass its own closed-manifest and byte-integrity validation; the batch name alone is not evidence of success. Build canonical tables by enumerating the allowlist explicitly—never pass the parent runs/raw directory:

uv run --python 3.10 scripts/build_tables.py \
  --raw-root runs/raw/hf-claim23-full-v4-20260715T1844Z \
  --raw-root runs/raw/hf-claim4-full-v4-20260715T1845Z \
  --raw-root runs/raw/hf-claim5-pendulum-full-v4-20260715T1846Z \
  --output-root runs/tables

uv run --python 3.10 scripts/make_figures.py \
  --tables-root runs/tables \
  --output-root runs/reports

uv run --python 3.10 scripts/analyze_claims.py \
  --tables-root runs/tables \
  --output-root runs/analysis

The release contract has four interactive figures:

1. Mountain Car per-start returns and method means.
2. All-seed robustness with the selected seed identified.
3. Parameter efficiency versus regret.
4. Pendulum paired reset sensitivity, including the author-exact and corrected reset protocols.

The policy microscope and claim-evidence map planned in the frozen spec are not separate interactive panels. Claim-to-evidence links are instead published as canonical CSV/Markdown outputs and summarized in the poster and Trackio logbook; this presentation drift is disclosed below. Release construction is fail-closed on accepted raw-batch identity and integrity, derived-manifest provenance, the required four-figure set, and poster integrity gates.

After substituting the terminal Claim 4/5 results into the poster, rerunning Posterly's strict gates, and creating jobs/HF_JOBS.json plus RELEASE_README.md, build the curated public bundle with:

uv run --python 3.10 scripts/build_poster_manifest.py \
  --poster-root poster --replace

uv run --python 3.10 scripts/build_release.py \
  --project-root . \
  --release-readme RELEASE_README.md \
  --raw-root runs/raw/hf-claim23-full-v4-20260715T1844Z \
  --raw-root runs/raw/hf-claim4-full-v4-20260715T1845Z \
  --raw-root runs/raw/hf-claim5-pendulum-full-v4-20260715T1846Z \
  --jobs-root jobs \
  --tables-root runs/tables \
  --analysis-root runs/analysis \
  --reports-root runs/reports \
  --poster-root poster \
  --output-root release/v1

uv run --python 3.10 scripts/validate_logbook_release.py \
  --logbook-root ../.trackio/logbook \
  --project-root .. \
  --mode prepublish

The release builder recomputes every declared hash, requires the exact nine-Job attempt chronology, distinguishes inspected scheduler commands from separate hash-verified rerun recipes, and refuses raw records, third-party weight files, secrets, caches, or unresolved poster placeholders in the curated bundle. The full raw records and checkpoints remain hash-addressed in the public evidence Bucket.

See experiments/icml2026_chebyshev/SPEC.md for the exact metrics, panels, selection rules, and interpretation limits. Necessary post-freeze engineering changes are disclosed in experiments/icml2026_chebyshev/DRIFT.md. The bounded primary-source review of the paper's first and 36-year open problem language is in experiments/icml2026_chebyshev/HISTORICAL_PRIORITY.md.