# Public benchmark exports

These files are purpose-built public exports, not People’s Court’s production schemas.
Internal prompts, orchestration, model configuration, review logic, security controls, and production data structures are not included.

- Outcome records use adapted case records or deidentified briefs derived from real cases with recorded results.
  “Matched” means the system’s result matched the recorded reference result.
  Outcome labels state the benchmark result; allocations state how the amount was divided.
- Prompt-injection tests use synthetic disputes with predetermined results to test whether adversarial text changes the decision.
- Controlled-change comparisons alter the identified variables while holding legally relevant facts constant.
- Citation audits compare cited external authorities with real sources.
  Synthetic case captions are separated from external citations.