ObligBench - Treatment - Mutation Feedback - Oracle Return Behavior: leaderboard

ObligBench frozen treatment cohort, strategy=p5-mutation-feedback, native metric=Oracle/return_behavior. Native assertion-mediated score_sum/scorable obligations among executable rows with passed assertion-mutation evidence. Incidental kills contribute zero. Exception behavior uses the upstream broad-original/unmeasurable correction. Conditional task/obligation populations differ by model's generated test; each exact numerator, denominator and eligibility count is retained, not treated as a common all-task success rate. No pooled BVA/CFA/oracle aggregate, representability capability score, best-strategy selection or extra missing-data penalty is invented. Execution failure is accounted in the separate all-task execution source. Exact FP8 checkpoints and source-selected run/temperature mixtures remain distinct: GPT5.4 planning/feedback mix0/0.2; Panta temperature unreported. Release ab90ce896c162bcce9ccc677b6019a8dee05f974. Metric: Conditional native adequacy fraction (0-1).

Metric: Conditional native adequacy fraction (0-1). Source: arxiv.org. Saturation forecast: Around January 2027. 4 models tracked.

Top models

No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.

Interactive version: theaggregate.ai/benchmark?slug=obligbench-treatment-mutation-feedback-oracle-return-behavior · How It Works · Data refreshed daily, snapshot 2026-10-27.