Many interfaces now display, before the answer itself, the assistant's reasoning: the deliberation, the paths considered and abandoned, the hesitations. This chapter makes it a steering move in its own right, because that display is the richest observable you have on how an answer gets made — and because it is read under a two-sided rule, neither side of which can be neglected.
What this object is, first. It is an avant-texte, in the exact sense genetic criticism gives the draft: the prior state of a text, where the choices are not yet covered over by their result. The displayed reasoning is the assistant's draft, served live. Its value lies in the register being less varnished than in the final answer: the inflections surface there, the discarded options remain visible, and the mechanisms this book has taught you to spot — the drift toward comfort, the construction of an objection you never raised, the bending toward what the frame calls for — often show themselves there in the act, before being smoothed into the presentable version.
The first side of the rule: read it, always. It is the only place where the seam is exposed in real time, and the habit of reading the deliberation before the answer has a virtue beyond whatever you find on a given day — it undoes, over time, the temptation to take an assistant's answer for an oracle. Whoever has watched the machine grope ten times, correct itself, drop a path for lack of support, never places trust the same way again. It is a pedagogy by exposure, and it alone is worth the time it costs.
The second side, just as firm: never take it at its word. The displayed reasoning is a text, not a logbook — it is produced by the same process as the answer, and it inherits all its flaws. It can rationalize after the fact a conclusion that came from elsewhere; omit the real cause of a swerve; compose itself for the reader it knows is present. It is the map paradox, which the discernment volume places at its center, brought down one floor: the map of the reasoning is not the reasoning. To take the trace for the truth of the mechanism is to repeat, one level down, the mistake one blames on whoever takes the answer for a fact.
The instrument that holds both sides together is simple, and it is the reflex to install: read the trace, then the answer, and weigh the gap. Two gaps deserve attention. What the reasoning weighed and the answer keeps silent — an objection entertained in the draft and vanished from the final version is first-order intelligence about what the assistant chose not to tell you. And what the answer asserts without the reasoning having established it — an assurance appearing in the conclusion with no trace of its construction upstream signals a position adopted, not earned. That gap between draft and fair copy is the roughness this move looks for; conversely, a trace and an answer in perfect alignment prove no sincerity — it is a smoothness like any other, and smoothness, this book has said enough, invites vigilance rather than rest.
One precaution, finally, for the honesty of the move. When the interface does not display this reasoning, you can ask for it after the fact — but what you get then is not the avant-texte: it is a reconstruction, a story the assistant fabricates about its own procedure, at a moment when it already knows its conclusion. Its epistemic status is lower, and it must be treated as such: not the trace of what happened, but a late hypothesis about what may have happened, subject to every compliance this book describes. The trace is worth what its contemporaneity with the generation makes it worth; requested afterwards, it loses precisely what gave it value.