15.232 14.8.22 Leaderboard y prevención de gaming
| Elemento | Política |
|---|---|
officialTest |
hidden or access-controlled; labels unavailable to developers |
submissionLimit |
prespecified per time window and logged |
divisions |
ClosedComparable OpenResearch OracleDiagnostic EdgeDeployment |
ranking |
task/track vectors; no cross-version mixing and no mandatory universal composite |
qualityFloor |
profile-specific accuracy, calibration and safety gates before system-performance ranking |
manualIntervention |
prohibited unless a dedicated human-in-the-loop track declares timing and actions |
audit |
evaluation container, logs, hashes and sampled replay required |
withdrawal |
invalidated or privacy-affected runs remain tombstoned with reason |
El leaderboard técnico no incluye el estudio humano como una cifra combinada. La utilidad depende de tarea, interfaz, coach, tiempo y consecuencias de aceptación errónea.