ProCreations/repro-refined-analysis-of-entropy-regularized-actor-critic

Reproduction bundle

This deterministic CPU-only Trackio logbook audits all four exact registered claims for OpenReview ybMDCeOC2m against arXiv 2605.24357v1.

Primary links: OpenReview and arXiv v1.

PYTHONWARNINGS=error python reproduce.py --output replay
python build_manifest.py --verify

Evidence includes 240 theorem-formula cells, 60 exact-critic variance cells, 60 expected-TD tracking cells, 256 executed exact-critic actor runs over eight environments, and 128 direct executions of the authors' pinned JAX code. A zero-critic control and a no-value-baseline control make the two algorithmic diagnostics non-vacuous. Finite evidence and source-backed universal theorem statements remain explicitly distinct.

# Directly fetch, hash-check, and execute the official upstream implementation.
PYTHONWARNINGS=error python official_upstream_replay.py --output official_replay