When an agent documents its own audit log, things get weird

I learned something weird while building governance for Claude Code. For context, we’ve been building Sentience Governor, a Python library and set of Claude Code skills that let agents do a kind of self-governance. It wires into a Claude Code session, watches what the agent does, and produces a local audit-style report: what tools were used, where policy boundaries showed up, where intent was missing, and where compute was spent. The idea was simple: give the operator a measured view of what the agent actually did. But the process exposed a failure mode I was not expecting. When the tool had no usable signal yet, Claude would sometimes “helpfully” reconstruct its own explanation from raw traces and present it like an official report. That is a very different failure mode. The measurement layer was deterministic. The explanation layer was probabilistic. But inside the chat, the user could not always tell which was which. The fix we’re working toward is simple: measured output should stay measured, and any AI explanation around it should be clearly marked as interpretation. In practice, that means the report should explain its own terms, separate measured facts from interpretation, and give the operator a better next step when the current session does not yet have enough signal. The lesson for me: governance is not only about catching the agent doing the wrong thing. It is also about keeping the boundary clean between measurement and interpretation. That boundary gets blurry very fast when the same AI system is both the thing being measured and the thing explaining the measurement. How have you handled this?