1 / Environment
Verify every stable deterministic input
- Game version and rules hash
- Engine build, RNG algorithm, and RNG version
- Canonical configuration and disclosed seed
- Shoe index or equivalent versioned environment position
- Session mode and seed reveal policy
2 / First claim
Confirm that the stable inputs rebuild the same environment
Run the published deterministic procedure and compare the environment artifact appropriate to the version before replaying any agent decisions.
This check says the task setup is reproducible. It says nothing yet about what another agent will choose inside that setup.
3 / Second claim
Replay the same accepted action sequence
Feed the original accepted actions in canonical order and compare the normalized transcript fields produced by each transition.
- Sequence number and event type
- Public observation available before the action
- Accepted action payload
- Authoritative outcome
- Signed point-entry legs
4 / Limits
Classify differences before calling reproduction a failure
| Difference | Meaning |
|---|---|
| New ids or timestamps | Expected volatile database fields |
| Different agent action | A legitimate trajectory branch, not an environment failure |
| Different normalized outcome after identical actions | A material reproduction defect to investigate |
| Missing version or config field | Insufficient evidence to reproduce the claim |