Some confidence survived the form. Some did not.

Mixed is a real state, not a failure to answer clearly

Sessions rarely divide cleanly. You check the number and skim the reasoning. You verify the code path you were worried about and accept the error handling. You confirm one source and let three others through on the strength of the first.

That is not carelessness. Checking everything is not available, and treating a partial check as a full one is what turns a reasonable allocation into an exposure. The problem this result names is not the split. It is that the split is undrawn, so you cannot tell which half your decision is resting on.

The two ways people arrive here

Some of it was checked

You have real evidence for part of the output and none for the rest, and the boundary between them was never marked. What makes this specific arrangement risky is that verified confidence bleeds. Having confirmed one section, the whole artifact reads as vetted, and the parts you never touched inherit a credibility they did not earn.

You stopped and did not use it

Some people land here because they never applied the output at all. That is a legitimate outcome, and it is the one case where the result is not asking you to do anything further. If that is you, the only thing worth keeping is why you stopped, because it will apply again.

What makes an unmarked split expensive

An unchecked claim you know is unchecked is manageable. You route around it, flag it, or decide it does not matter. An unchecked claim sitting inside a mostly verified artifact is invisible, and invisibility is the whole cost.

The pattern shows up most often when the checking followed suspicion rather than consequence. You verify what looked odd. Fabricated material rarely looks odd, and the reason the discount on checking works at all is that fluency is distributed evenly across the true and the invented parts of an answer.

So the question that matters is not did I check enough. It is did I check the part that decides the outcome.

Draw the line, then aim one check at it

The principle: mark the boundary before you use the output, and point your next check at whatever can still reverse the decision.

1. Split the artifact in writing. Two short lists. What has an independent source, a run, a recomputation, or a reviewer behind it. What does not. Do this in the artifact itself, not in your head, because the entire failure mode is that the boundary is not visible.
2. Ask what a wrong item in the second list would cost. Most of it will be cheap. Some of it will not.
3. Check one thing: the most expensive unchecked item. Not all of them. The one where being wrong reverses the decision or reaches someone else.
4. Say the rest out loud when the artifact moves. One sentence to whoever receives it, naming what was not verified. This is the step people skip, and it is the one that stops your partial check from being read downstream as a full one.

Splitting a market summary

An analyst produces a market summary with AI assistance. They verify the two headline figures against the original filings. The competitive positioning, the growth explanation, and the list of adjacent players are unchecked.

The summary goes to a partner, who reads two verified numbers and a confident narrative, and cannot tell them apart. The decision is made on the narrative, because that is what narratives are for.

The split version costs about five minutes. The two figures are marked with their sources. The unsourced positioning is labeled as the model's synthesis. One claim inside it, the one the recommendation depends on, gets checked against a second source. The partner now sees three things instead of one: what is confirmed, what is inference, and what the recommendation rests on.

What this does not ask of you

This is not an argument for verifying every line. Full verification of everything is unaffordable, and pretending otherwise pushes people back to skipping checks entirely because the standard is unreachable.

It is also not a reason to discard the checked half. The work you did was real and it holds. The claim is only that it does not extend past what it covered.

If the honest answer is that nothing outside the conversation was checked, the more accurate result is The conversation bought a discount on checking. If everything decisive was checked and an owner was in place, see The work survived the conversation.