What’s shipped, what we’re building, and what we’re exploring.

0.3.2.1 closes the gap between what sentience profile validate accepts and what the runtime can use. A profile that parsed but held the wrong kind of value where the runtime reads one, such as a number in the high_consequence.tools list or a word in dir_change_depth, could pass validation in 0.3.2 and then raise inside the governed run; in Claude Code the tool call went unrecorded. Validation now checks every field the runtime consumes, an invalid profile is never bound (a bound file degrades to the machine default, an invalid default leaves the session on no profile, and each case is one warning and a truthful registration), and the runtime substitutes safe defaults for anything that still reaches it. Valid profiles keep their fingerprints and their behaviour, event for event. Nothing is blocked; the record now says what actually governed the session.

Pydantic AI support ships as its own distribution. pydantic-ai-governor 0.1.0 (source) adds a Pydantic AI capability: attach it to an agent and each run opens its own Sentience Governor session, recording what the agent dispatched against the declaration state recorded before the run, with measured per-turn token usage. It installs separately, versions separately, and Sentience Governor itself takes on no Pydantic AI dependency. It pins core below 0.3.2; a 0.1.1 companion release adopting per-agent policy resolution follows separately, on core 0.3.2.1.