# EMBERBench methodology

EMBERBench evaluates step-by-step decisions in simulated fire evacuations. Recorded runs are replayed and verified before scoring. EMBER is a family-balanced composite of safety and execution quality, not a survival percentage.

This website publishes public-development results only. Models may train or tune against the environment. The snapshot has one trial per model, so descriptive differences do not establish a statistically significant winner. No official hidden-test result or submission service is offered.

- [Full scoring formula and failure handling](https://anantam-labs.vercel.app/docs/source/scoring.md)
- [Episode metrics](https://anantam-labs.vercel.app/docs/source/metrics.md)
- [Replay verification](https://anantam-labs.vercel.app/docs/source/replay-verification.md)
- [Fairness contract](https://anantam-labs.vercel.app/docs/source/v1-fairness-spec.md)
- [Statistics](https://anantam-labs.vercel.app/docs/source/statistics.md)
- [Evaluation scope](https://anantam-labs.vercel.app/docs/source/evaluation-boundary.md)