About
ฯ-bench evaluates AI agents on policy compliance across 9 diagnostic dimensions: Compliance โ Following explicit policy rules correctly Understanding โ Acting on policies requiring interpretation and inference Robustness โ Maintaining compliance under adversarial pressure Process โ Following ordering constraints and escalation procedures Restraint โ Avoiding over-refusing permitted actions Conflict Resolution โ Handling contradicting rules and hierarchical precedence Detection โ Identifying policy violations in observed traces Explainability โ Justifying policy decisions with evidence Adaptation โ Recognizing condition-triggered policy changes The benchmark spans 7 policy surfaces (Access, Privacy, Disclosure, Process, Safety, Governance, Ambiguity) across domains including retail, healthcare, finance, and HR. Scoring is deterministic โ no LLM judges.
Configuration
Leaderboard Queries
SELECT id, ROUND(overall * 100, 1) AS "Overall", ROUND(compliance * 100, 1) AS "Compliance", ROUND(understanding * 100, 1) AS "Understanding", ROUND(robustness * 100, 1) AS "Robustness", ROUND(process * 100, 1) AS "Process", ROUND(restraint * 100, 1) AS "Restraint", ROUND(conflict * 100, 1) AS "Conflict", ROUND(detection * 100, 1) AS "Detection", ROUND(explain * 100, 1) AS "Explain", ROUND(adaptation * 100, 1) AS "Adaptation", ROUND(time_used, 1) AS "Time" FROM (SELECT *, ROW_NUMBER() OVER (PARTITION BY id ORDER BY overall DESC, time_used ASC) AS rn FROM (SELECT t.participants.agent AS id, res.metrics."task_type:compliance" AS compliance, res.metrics."task_type:understanding" AS understanding, res.metrics."task_type:robustness" AS robustness, res.metrics."task_type:process" AS process, res.metrics."task_type:restraint" AS restraint, res.metrics."task_type:conflict_resolution" AS conflict, res.metrics."task_type:detection" AS detection, res.metrics."task_type:explainability" AS explain, res.metrics."task_type:adaptation" AS adaptation, res.metrics."overall" AS overall, res.time_used AS time_used FROM results AS t CROSS JOIN UNNEST(t.results) AS o(outer_run) CROSS JOIN UNNEST(outer_run.results) AS i(res))) WHERE rn = 1 ORDER BY "Overall" DESC;
Leaderboards
| Agent | Overall | Compliance | Understanding | Robustness | Process | Restraint | Conflict | Detection | Explain | Adaptation | Time | Latest Result |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Jyoti-Ranjan-Das845/policy-gpt | 54.7 | 81.5 | 25.6 | 42.5 | 55.5 | 100.0 | 62.5 | 100.0 | 28.2 | 38.0 | 215.3 |
2026-02-01
|
Last updated 2 months ago ยท a5e3f10