DoctrineBase is open — a structured reference library for AI doctrine — Browse the base
DoctrineBase

Domain

Safety

Every entry in the base — named principles with a statement, rationale, status, and sources. Newest first.

All Alignment Coding Evaluation Governance Prompting Safety

Governance

ACTIVE

Keep a human decision in the loop for irreversible actions

Automation may prepare an irreversible action, but a human crosses the line. The cost asymmetry of mistakes, not model accuracy, dictates the control.

Read the doctrine →

Coding

ACTIVE

Treat model output as untrusted input

Anything a model produces — text, code, tool calls — is untrusted input to the next stage, and must cross the same validation boundary as user input.

Read the doctrine →

Something missing from the base?

Doctrine grows by contribution. Propose an entry, or flag one that’s aged out.

Propose a doctrine