Built for agents,from the start.

Four things happen on every Xaish call, enforced at the response boundary so the model can't opt out:

1. Schema-vs-prompt alignment. Every Pydantic Literal value the schema accepts shows up verbatim in the system prompt. A 125-module audit script blocks the PR if they drift.
2. Mandatory caveats. Each tool has a 3-theme caveat set with phrase-coverage checks. If the model skips a theme, the server adds it back before the response leaves.
3. Caps on thin input and critical findings. Confident dispositions like allow, healthy, or grade A collapse to safer values when the input is sparse or a critical finding is present.
4. Credential redaction. Inputs and outputs both pass through a redactor that strips API keys, bearer tokens, full card numbers, bank accounts, and crypto secrets. Last four digits only, always.