MarxistLeninist/repro-markov-decision-contests-bundle

Reproduction bundle — Reinforcement Learning with Pairwise Preferences in Long-Term Decision Problems

Complete artifacts for an independent, CPU-only numerical audit of the Markov decision contest framework (ICML 2026, OpenReview E9QkGczNUw, arXiv 2606.00367).

Logbook: https://huggingface.co/spaces/MarxistLeninist/repro-markov-decision-contests

Result: 6 of 6 audited claims supported.

Experiment	Paper result	Verdict
E1	Lemma 4.3, Thm 5.6(1) — stationary policies suffice	supported
E2	Thm 5.6(2) — optimal iff solves the game	supported
E3	Thm 6.1 — exact solution in P	supported
E4	Example 5.3 — MDPs embed; MDCs strictly more general	supported
E5	Thm 7.5 — HPI converges at O(1/sqrt(K))	supported
E6	Lemmas 7.3/7.4 — marginal values and the PDL	supported

Contents

protocol.md              pre-registered plan (committed before any experiment)
src/                     MDC machinery, the six experiments, figures, poster, Job builder
tests/                   15 unit tests
configs/                 smoke and full configurations
outputs/full/            raw_E*.json (per instance) + summary.json (aggregates)
figures/                 PNGs, each regenerable from outputs/
logs/                    Job stdout, Job metadata, run logs
experiments/             the self-contained script that ran on the Hugging Face Job
poster/                  poster HTML, print PDF, PNG, gate report
MANIFEST.json            environment, Job evidence, SHA-256 for every file
SHA256SUMS.txt           plain checksum list

Rerun

python -m pytest tests -q
python src/run_audit.py --config configs/full.json --out outputs/full
python src/make_figures.py --results outputs/full --out figures

Requires numpy and scipy (plus matplotlib for figures, pytest for tests). Everything is seeded, so re-running reproduces the stored numbers.

Hugging Face Job: https://huggingface.co/jobs/MarxistLeninist/6a69ea7b4497041dbfc386de (cpu-basic, 504 s, ~$0.00168).

Scope

Sections 4–7 and Appendix C are audited. The paper's Section 8 comparison of HPI/HPI-Clip against SPPO on 13 MuJoCo tasks needs GPUs and is out of scope — neither supported nor contradicted here.

A numerical audit verifies stated equalities and inequalities to double precision on a seeded family of instances, with controls that break them when hypotheses are relaxed. It is evidence for the theorems, not a replacement for their proofs.