ProCreations/repro-revisiting-zeroth-order-hessian-approximation-a-single-step-policy-optimization-lens

Reproduction - Revisiting Zeroth-Order Hessian Approximation

Independent CPU audit of all six registered claims. It reconstructs the Gaussian smoothed-Hessian and optimal-baseline identities in 288 Monte Carlo cells and directly executes the pinned author code for Claims 5--6. The native recertification contains 480 claim-bearing histories and 8,400,480 objective values at the paper's exact 10,000-dimensional scale. Claims 1--3 verify; Claims 4--6 are literal, scope-preserving falsifications.

Primary source: https://arxiv.org/pdf/2605.30960v1

OpenReview: https://openreview.net/forum?id=nEQYu4ndGA

Author code: https://github.com/Qjbtiger/ZoVH/tree/0d22908da1bdb1e0c531b917f73eeb8c488b3e5a

No competitor artifact, output, or result was used.

Judge-facing entry point: pages/00-judge-evidence-scorecard/page.md. Structured evidence: EVIDENCE_MATRIX.json. Native protocol and artifacts: native_protocol/README.md and native_outputs/native_recertification.json.

python3 -m pip install -r requirements.txt
PYTHONDONTWRITEBYTECODE=1 PYTHONWARNINGS=error python3 reproduce.py --output-dir /tmp/zovh-replay
PYTHONDONTWRITEBYTECODE=1 python3 validate_evidence.py
PYTHONDONTWRITEBYTECODE=1 python3 verify_manifest.py