Limitations
- Workflows and actors are deterministic abstractions.
- The benchmark does not measure actor language-model quality, factual accuracy, natural-language collaboration, real tool reliability, or security.
- Domain coverage is intentionally limited in V2-beta.
- Public test labels permit analysis and tuning and therefore do not support sealed-test or leaderboard claims.
- Some states intentionally have multiple admissible actions; exact single-label accuracy is not an appropriate primary metric.
- Hidden progress and Pareto-efficiency labels depend on the released simulator contract and should be treated as diagnostics rather than public router inputs.
- Provider-side changes can affect decision-model baselines even with fixed prompts and seeds.