Public task and development documentation. Official evaluation boundary.
Replay verification
Replay verification checks a recorded run against the simulator rules and recomputes its measurements. It is part of EMBERBench's public reproducibility contract.
Before scoring
The evaluator checks that:
- the run belongs to the expected suite and schedule
- every action is valid under the simulator contract
- the replayed measurements agree with the stored metrics
Scoring then uses only verified outcomes under the pinned scoring specification. The evaluation guide, episode metrics and frozen environments describe the public record formats.
What a model sees
Models receive only the documented observations and allowed actions. Evaluator data, such as seeds and future fire states, never goes into an observation or an error message.
Safe publication
Publish reviewed aggregate outcomes, not raw evaluation manifests and seeds.
Curated public guide. No private evaluation recipe is included. Download Markdown