A system must be able to halt without producing a confident output.

Most software is judged by what it produces. A legal system must also be judged by what it declines to produce. The Zeroth Law says a refusal is always preferable to a fabrication, and that principle has to be built into the pipeline rather than left to the model, because a language model asked a question will almost always answer it.

Halting is an architectural capability with concrete parts. The system needs defined conditions under which it stops: a validity gate that quarantines a brief resting on a dead or misapplied citation before anything is computed from it, a calibration check that fails when confidence no longer matches accuracy, an abstention rule for when signal does not clear noise. It needs a visible state that means "stopped, and here is why," distinct from both success and error. And it needs somewhere for the work to go: a human reviewer, a logged question, an element marked unresolved.

The blind eDiscovery review we published shows what this looks like in practice. The engine's calibration gate failed, so automatic acceptance was locked and all 2,038 responsive determinations were routed to human review. The elusion gate failed too, and the report said so: the evidence supported automated classification with mandatory human review of the machine-produced pile, not fully automated production at the target elusion standard. The system did less than it could have, because it could not defend more.

A tool that always returns a confident answer has not been shown to be reliable. It has only been shown to be unable to say when it is not. The capacity to halt is what gives its other outputs meaning.