The bounded AI workflow

Place model output inside a visible contract, review point, and reversible action path.

DIRECT ANSWER

A safe first architecture treats model output as an untrusted proposal. Validate the input, require structured output, record evidence, route exceptions, and let an authorized person or deterministic rule approve any consequential action.

SYSTEM FLOW

Make every boundary visible.

01request
02validate
03model
04parse
05review
06act
07audit

DESIGN DECISIONS

Three decisions carry the architecture.

Validate before generation

Reject missing, oversized, prohibited, or unauthorized input before it reaches the model boundary.

Parse into a contract

Accept only the required schema and treat free text, missing fields, and unknown values as failed output.

Separate proposal from action

Queue the result for review or deterministic policy instead of granting the model direct side effects.

FAILURE MODES

What breaks the boundary.

A fluent response bypasses schema validation
Retry creates the same external action twice
Review exists in the diagram but has no owner or evidence

VERIFICATION

Evidence before release.

Invalid-input test
Schema-failure test
Duplicate-action test
Manual fallback rehearsal