Useful intelligence must remain governable.

Models may reason, generate, summarize, and propose. Deterministic systems control permissions, quotas, billing, external writes, source status, confirmations, execution, and verification.

Current facts, research, legal explanations, test results, and consequential claims require evidence appropriate to the task. Sources, authority, dates, limitations, and unresolved conflicts should remain visible.

Generated code and tool actions should run only inside bounded environments. External actions require authorization, and consequential changes require explicit confirmation.

Private work remains private unless deliberately published. Context is scoped to the relevant product, project, conversation, or user-approved memory link.

Agatha products are designed with explicit boundaries. Agatha Law explains public law rather than giving individualized legal advice. Agatha Notice may suggest or prepare but does not silently book, purchase, send, publish, or deploy. Agatha Coder may propose code changes, but tests and execution evidence determine whether the work succeeded.

Safety is continuous. Agatha systems should be tested against prompt injection, unsupported claims, wrong-source retrieval, permission failures, secret leakage, destructive actions, cross-context contamination, and false verification.

Safety claims must reflect implemented and verified controls. Planned architecture, schemas, fixtures, and design documents must not be presented as deployed safeguards.