A finding is a claim with sources: the stated ARR includes a cancellable pilot; see contract 14, clause 3, and the ARR schedule, row 22. Models are good at producing these now. The question is what happens next, and most tools answer it with a checkbox that nobody clicks.
Pending is a real state
In Navent every finding arrives pending. It is visible, it is cited, and it counts for nothing. It does not appear in the memo. It does not complete the step. It sits in the review queue until a person gives it a verdict.
There are three verdicts. Accept means you stand behind it. Challenge sends it back with a note, and the note is in scope when the step re-runs, so the model is answering your objection and not starting over. Reject means it was wrong, and a rejected finding stays on the record with that disposition. Nothing disappears.
Enforced downstream, not just displayed
The rule has teeth in two places. The memo assembles only accepted findings. Everything else, pending, challenged or rejected, is listed under not in the memo with its disposition, so an omission is a decision on the record rather than an accident of what the model happened to draft.
And a run does not complete a step. It moves the step to needs review. The step advances when a person does. The process map shows what the model produced and what a person decided, separately, so "done" always means a person said so.
Why this is not friction
The objection is obvious: thirty-one findings to rule on is thirty-one clicks. Yes. It is also thirty-one things a person actually looked at, each with its passage a click away, instead of a twelve-page summary nobody read past page four. The review takes an afternoon. The alternative is a memo whose author cannot tell you which sentences they believe.
Trust, in the end, is not a property of the model. It is a record of who looked, and what they decided.