Transluce Launches Docent: Upload Logs to Pinpoint Disobedient AI Agent Behaviors

ChowdhuryNeil · x · 2026-08-05

TransluceAI introduced Docent, an agent behavior analysis tool designed to help developers detect non-compliance or reward hacking in AI systems.

Workflow:

Create a rubric: Users ask a question (e.g., "where is it reward hacking?"), and Docent automatically reads the data to generate a precise behavior rubric.
Spot-check results: The tool searches logs for matching violations, explaining why they matched and pinpointing their exact location in the transcripts.
Quantify and visualize: Data is aggregated and filtered using charts, allowing users to plot the frequency of reward hacks across training steps or compare violation rates between different models.

More from coding & agent