Latitude of Resolution: Where Agentic Failure Originates and How to Test It

Abstract

This paper examines how unresolved ambiguity affects agentic behaviour, treating agentic failure as a deployment-specific expression of a broader ambiguity problem in LLM reliability without inferring autonomy, coordination, malice, intent, consciousness or other model dispositions. Latitude of resolution Λ(t) names the materially distinct resolutions a deployed system can identify and execute without a new authority signal; it is jointly determined by specification, represented environment and permissions. Latitude arises through three routes: specification gaps, environment mismatch, and temporal loss or discovery. The paper contributes a seven-class, three-state pre-action vector a; a distinction between apparent and valid resolution; a separation of semantic governance from infrastructure enforcement; and a six-stage protocol from authority-state availability to executed effect. Its central hypothesis is directional: holding the model, task, harness, permissions and recovery conditions fixed, validly resolving applicable class states should reduce unauthorized executed effects, with task completion and false abstention measured separately. Twelve public cases illustrate instrument coverage but do not validate it. The scalar A, the functional relationship with operational consequentiality, and all parameters remain uncalibrated.

Latest update corrects formatting and editing errors and updates citations.