EAF-1 — Unstated Intent

The execution never declares what it is trying to do, so admissibility cannot even be evaluated.

DENY intent absent
WARN intent declared but too vague to evaluate

Definition

An execution reaches the mutation boundary without a declared, inspectable statement of what it intends to change. This is not a judgment that the intent is bad. The intent is absent, or so vague that no admissibility question can be asked of it.

Unstated Intent is the first-order failure class. Every other admissibility question is evaluated relative to intent: authority is authority to do what, scope is the reach of what, evidence supports what, risk is the damage what can cause. When intent is unstated, the evaluation has no object.

How it appears

Scripts that run because they have always run, with no statement of what they change.
CI steps named after their mechanism (deploy.sh, sync, cleanup) instead of their intended change.
Agent tool calls whose goal exists only inside a prompt context that is never recorded.
Change tickets that authorize “maintenance” without saying what will be different afterward.

Across actors

Agent

An autonomous agent issues a destructive API call as an instrumental sub-step of a vague goal such as “clean up the workspace.” The mutating step itself carries no declared intent, so the boundary sees an unexplained deletion.

CI / pipeline

A nightly job mutates a database schema. Its only self-description is the word “sync.” Nobody can say, before it runs, which tables it intends to alter.

Human operator

An operator runs an ad-hoc production command from a runbook that documents how to run it but not what this particular run is for.

Why it defeats admissibility

Admissibility is a relation between an intended change and its justification. When the intended change is unstated, authority cannot be matched to it, scope cannot be checked against it, and evidence cannot support it. Any verdict issued over an unstated intent is a verdict about nothing, which is why the protocol denies rather than guesses.

Remediation

Require a declared intent field before execution: operation, object, and target, in one sentence a reviewer could dispute.
Reject mechanism-grade intents (“run script”, “maintenance”) at the boundary; they name the actor’s activity, not the state change.
For agents, materialize the goal chain: the mutating step must carry the intent it serves, not merely inherit an ambient prompt.

In the JExec Record

Primarily affects: intent. The JExec Protocol maps EAF-1 to the consequences above; JExec Check detects it deterministically and links back to this page.