## MESSIER v1.0.2
This release corrects the treatment of non-scorable attempts in Harbor evaluations.
- **Corpus.** 904,051 records from 31 benchmarks, 734 agents, 11,999 tasks, and 72,000 verifiers.
- **Data.** The release contains task definitions, results, verifier definitions, and 119,482 available trajectories.
- **Harbor attempts.** Attempts without verifier rewards are retained with `result=None` instead of being omitted.
- **Failure handling.** Provider, environment, and evaluator failures are marked as non-scorable when they prevent evaluation, while agent timeouts and complete graded outputs remain scored.
- **Analysis.** The HarveyAI-Lab counterfactual analysis and derived corpus statistics are updated accordingly.
