Uniqueness 7/10: the synthesis is novel, the imputation primitive is not

Asked:

“rate uniqueness of this idea” — followed by the full Plan A, a verification-budgeted RLVR proposal: learn a proxy for an expensive verifier, actively buy a subset of true reward labels per rollout group, impute the rest, and train with group-normalized advantages under a fixed verification budget.

Assessed as of 2026-09-05 against 10 curated closest prior works (2018–2026) and 6 estimator-theory sources. The bare core idea rates about 4/10; the specific synthesis rates about 7.5/10; the overall research package lands at 7/10 — promising and differentiated, but not a new primitive.

Uniqueness scorecard — 0 to 10, hover a bar for the reasoning

How close is the prior art

Overlap with Plan A, per prior work — longer and darker means closer. Blue works share the reward-verification axis; grey works allocate rollout generation, a related but orthogonal budget.

Findings

The closest single work is iopscience.iop.org — actively learning costly reward functions during RL — which already pairs a learned proxy with active labeling; Plan A adds the propensity correction and agentic-RLVR study it lacks.

Reward imputation itself is established: huggingface.co imputes rewards from limited labels (offline RL), and proceedings.mlr.press learns fast reward approximations for expensive systems — so imputation cannot be the novelty claim.

The RLVR budget neighbors — SARA, TRACE, VIGOR, HORA, rollout pruning, Spec-RL — allocate rollout generation, not reward verification: related, orthogonal, useful positioning contrast (e.g. pith.science, huggingface.co).

Fatal as written: deterministic entropy top-k gives some rollouts zero inclusion probability, violating positivity — randomized auditing is required (arxiv.org); and unbiased pseudo-rewards do not survive nonlinear group normalization, so “gradient unbiased in expectation” is too strong without a new proof (arxiv.org).

Arithmetic problems: group size 8 with at least one verification per group imposes a 12.5% floor, conflicting with the planned 10% and 5% budget points; and the stated 60 runs omit the listed baselines and ablations, so compute must be recalculated.

Fatal as written, fixable in revision

Estimator and theory sources behind the technical verdict

Source	Result	Caveat	Implication for Plan A	Link

The revision path: replace deterministic top-k with stochastic acquisition and logged propensities plus an ε-uniform audit floor; cross-fit or lag the imputer; derive the correction at the gradient/advantage level; compare fixed or lagged scaling and leave-one-out centering; report effective sample size and weight tails; add calibration (Brier, ECE) and worst-slice audits beyond AUROC; and separate generation, environment-execution, and judge cost in the budget accounting. At 5% verification, inverse-propensity variance may be severe.

Prior-art matrix

Work	Date	Overlap	Novelty implication	Link

Verdict

Worth pursuing. Position it as “verification-budgeted agentic RLVR with randomized auditing and estimator-aware group advantages,” make the main claim empirical unless a rigorous group-normalized policy-gradient theorem is proved, and let the novelty rest on the synthesis and the careful estimator/audit study — never on imputation itself.

Method: uniqueness review of the Plan A verification-budgeted RLVR proposal using live web research through 2026-09-05. Evidence: 10 curated closest prior works (method, overlap, novelty implication, date, source) and 6 estimator/theory sources (result, caveat, implication). Ratings are reviewer judgments on a 0–10 scale, not measurements. The scan cannot guarantee exhaustive or legal novelty like a patent search. Full method texts of prior works were trimmed for space; sources are linked per row.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.