CausalSmith — CSWP-2026-025 · ← back to the paper · arXiv · GitHub

AI referee report

on “Parametric Rates for Stationary Policy Evaluation with Binary Hidden States”

Overall score 7.7/10 4 findings — 4 minor

The submission establishes a substantial and carefully scoped functional-stability result and matching minimax rate, with remaining weaknesses concentrated in organization and notation.

Referee report of 1 Oct 2026. It was read against the current draft of the paper. One automated referee’s estimate — not peer review. Produced by a single automated referee in the pipeline's review stage. It can be wrong, and it is not peer review.

Summary

The paper characterizes stationary policy evaluation from one behavior trajectory in a fixed binary partially observed experiment under known observed-state policies, action overlap, and contraction. Its central contribution combines seven observable intervention moments with uniform stationary-value stability across reduced-rank four-state realizations to establish matched T^{-1} mean-squared-error bounds. Taking the verified results as established, the manuscript largely represents their scope faithfully and engages the closest literature carefully. Publication would benefit from a more focused presentation and clearer minimax quantification.

Strengths

Uniform stability of the stationary reward across reduced-rank realizations gives the rate characterization substantive content beyond regular latent-parameter estimation.
The estimator and testing pair operate within the same explicitly defined observed-data experiment.
The manuscript carefully distinguishes fixed hidden cardinality from the growing-cardinality comparison classes.
Related work discusses competitors' estimands, sampling designs, rates, and identification conditions in a dedicated early section.
The external manuscript dependency and the candidate-count scope receive explicit disclosures.

Findings

minor Main results — paragraph preceding the minimax risk definition

The sentence 'The supremum ranges over the binary experiment class ... with those supplied inputs' suggests that the supremum fixes the behavior and target policies. Elsewhere, the class allows both policies to vary, with each experiment supplying its own pair to an indexed estimator. These are different decision problems.

Suggested fix

State explicitly that the infimum ranges over families of observed-data procedures indexed by the supplied policies, while the supremum ranges over all admissible kernels, policies, and trajectory laws at fixed regime parameters. Describe evaluation at each experiment's supplied policy pair.

minor Main results — fixed binary minimax rate; Discussion; appendices

The central theorem combines the binary risk bound, extensive testing-witness details, and a lengthy externally sourced cardinality comparison. Repeated explanations of the seven moments, contraction mechanism, testing pair, and candidate counts across subsequent sections dilute the main statistical contribution.

Suggested fix

Present the binary risk characterization prominently, organize the testing witnesses as supporting material, and give the external comparison a separate clearly identified block with its citation and formalization disclosure preserved. Consolidate repeated explanations, retaining one substantive account of each mechanism and using cross-references elsewhere.

minor Introduction

The sentence 'this is a candidate-count statement, not a runtime bound' explains the contribution through an absent deliverable outside the limitations section, contrary to the affirmative framing contract.

Suggested fix

Replace it with an affirmative statement such as 'The proposition quantifies exhaustive search through the number of candidate realizations at fixed mixing scale.' Keep the runtime and precision discussion in Limitations and future work.

minor Appendix — pair-polynomial identity and value modulus

The lemma calls all six vectors, including the reward vectors, 'real row vectors,' while the setup and neighboring exposition use rewards as columns in scalar matrix products. The coordinate-sum interpretation resolves the mathematics but leaves conflicting orientation instructions.

Suggested fix

Describe the objects as real coordinate vectors, explicitly interpreting initial and stationary distributions as rows and reward functions as columns. Preserve the displayed coordinate sums and the verified conclusions.

Questions for the authors

Is the intended minimax experiment the class-wide problem with policies supplied separately for each experiment? Please make that quantification explicit throughout the risk discussion.

Review history

Round	Date	Score	Findings
1	1 Oct 2026	7.4	6
2	1 Oct 2026	7.5	6
3 · current	1 Oct 2026	7.7	4

Every referee draw is kept: scores are noisy, so the history is shown rather than only the last or the best pass.

Read against paper.tex sha256 2131817a8b9f.