Termination drift and proxy verification collapse in iterative human–AI work.
Intent transfer names the channel; the loop record documents where it breaks.
Every document below is rendered from the living markdown source.
Abstract. Intent loss between human and AI is patterned, not random.
The strongest predictor of whether a task ever finishes is not model capability or prompt
length, but whether the success condition is checkable: one finite, observable "done when."
Unfloored conditions produce termination drift (loops that cannot close); proxy verify
gates produce false joint confidence when they contradict runtime (verify ≠ runtime).
A few inspectable lines of standing intent, re-supplied every round, preserved the original
vision better than any volume of accumulated context.