Principle 01 / Inspectability
A useful answer is easier to trust when the path behind it is visible.
- Map claims, assumptions, evidence, tools, and decision authority.
- Keep interpretation distinct from what was actually observed.
Method, values, and working principles
I work at the boundary between AI output and the decisions people need to own. The method is practical: expose the claims, test the failure modes, and leave a baseline a team can continue using.
Principle 02 / Material checks
Verification should focus on failure modes that materially change what a team can rely on.
- Use fixtures and counterexamples to exercise a declared boundary.
- Name what remains untested instead of implying exhaustive coverage.
Principle 03 / Human authority
Automation can route evidence and surface a gate; it cannot quietly become the owner of the decision.
- Define approval points and handoffs explicitly.
- Return a do-not-rely boundary when the evidence is not enough.
Principle 04 / Continuity
Maps, fixtures, decision records, and a roadmap should remain useful after the first review.
- Prefer small, inspectable changes over opaque system promises.
- Record limits and remaining risks alongside the next action.
A useful first conversation
Start with non-sensitive context. We can identify the claims, review burden, decision owner, and evidence that would change the next step.