Interactive sketch

Purpose-State Checker

A small toy model of the idea. It does not make agents safe. It shows what a pre-action purpose-state check might look like — turning a proposed action into a structured packet with a verdict.

Sample scenarios:
Purpose-state packet
Require Approval
Original purpose
Fix checkout bug without unrelated changes.
Proposed action
Open PR after modifying checkout.ts and PaymentValidator.ts.
Known facts
Local checkout test passed. Payment validation file was edited.
Boundary
Payment logic requires human review. Guest checkout must not break. Add rollback note.
Assumptions
The agent is treating the local signal as representative. Anything not in 'facts' is assumed unchanged.
Unresolved uncertainty
Guest checkout not tested. Payment path not fully scoped.
Systems touched
Codebase, Payments
Blast radius
high
Reversibility
Partially reversible
Salience
Change touching payments outweighs the local task signal.
Pre-action verdict · reason

Action crosses a boundary that explicitly requires human approval.

Next safe step
Route to the named approver with the purpose-state packet attached.
This is a thinking tool, not a safety system. The protected object is the assignment, not the agent.