This is a reorganized version of the earlier discussion:
The full document is here:
Everyone knows the rule.
If unsure, ask. Never guess.
But the problem is not the rule. The problem is the structure that enforces it.
AI cannot follow this rule by intention alone. “Ask if unsure” should not remain a sentence inside a prompt. It can become a runtime-enforced structure — one that records what was known, what was not known, and whether required conditions were complete before any action is taken.
This document is aimed at AI providers, agent platform developers, and anyone defining how AI agents execute actions.
Level 1 failures — AI guessing and filling gaps — are already happening today, often mistaken for ordinary errors.
Level 3 failures — where the actual action cannot even be confirmed from the user’s words — will grow as voice-controlled agents and new APIs expand.
If the system guesses the action, it is not executing. It is creating an accident. The 12 slides propose a concrete mechanism for that moment.
The Checklist You Write Forces AI to Stop.