benchflow-experiment-review

Detect reward hacking in RL trajectories by auditing JSONL rollout files to parse reward records, verifying artifact completeness, and identifying unauthorized skill access or verifier leakage in no-skill trials.

Validate Prime-RL training records by parsing JSONL trajectory files and enforcing defined schema constraints to verify artifact completeness, environment hardening, and training readiness.

Auditing Benchflow experiment trajectories for release readiness involves verifying ACP and LLM artifact integrity, ensuring environment hardening, and detecting skill leakage across different backends like Docker and Daytona.

Yes, trajectory validation works with Docker and Daytona backends, auditing model rollouts to confirm that code changes maintain parity in artifact shape and sandbox behavior across environments.

You need to verify artifact completeness for model rollouts after Benchflow code changes to ensure that ACP and LLM trajectories, reward records, and training-ready results maintain their expected shape and integrity.