Published human signal

Human preference results

A diagnostic view of what reviewers preferred, published as an immutable aggregate snapshot—never a live look at assignments.

Loading the published snapshot

Gathering aggregate scores and coverage…

Immutable snapshot

Accepted votes—Validated human choices

Reviewers—Contributors represented

Compared pairs—Aggregate matchups

Contestants covered—Models plus the human reference

Reliability checks

Bounded summary statistics only—no reviewer histories or hidden source identities.

A-position win share—95% Wilson interval

Panel agreement—Exact agreement and kappa

Human ↔ automated excerpt panel—The same anonymous passages and review question

Matched excerpt-panel ranks—Spearman rank correlation on matched resolved panels

Draw share—Both-good and both-bad retained

Trust-weight sensitivity—Maximum rating delta

Bootstrap success—Reviewer-resampled

Balanced cells—Equal task/category mass

Diagnostic ranking

Published human preference point-estimate order with 1,000-centered Davidson diagnostic ratings, confidence intervals, votes, and coverage
Strict rank	Contestant	Diagnostic rating	95% interval	Votes	Coverage