The interrogation test

Most records are written for a friendly reader. You can tell because they read so easily. The author imagines someone skimming, nodding, looking for confirmation that the right boxes were checked, and writes to satisfy that reader — a reader who wants to be reassured and has no intention of pushing. A document built for that reader is optimized for the wrong thing. It is optimized to look complete. And looking complete is exactly the property that a hostile reader, the one who actually matters, will ignore.

The friendly reader asks: does this account hang together? The hostile reader asks: where does it break? These are not the same question, and a record tuned to pass the first will routinely fail the second. Completeness is a property of appearance. It can be manufactured by anyone with a template and an afternoon. What cannot be manufactured after the fact is the ability to withstand questions the author did not see coming — because those questions probe whatever the author was not thinking about, which is precisely where the manufactured completeness has nothing underneath it.

So here is the test I want to put to any record that claims to account for a consequential decision. Forget how it reads. Imagine it under interrogation by someone competent and unsympathetic, someone who did not write it and does not want it to be true, and who is allowed to ask anything. The record is good to exactly the degree that it can keep answering — not the questions it was built to answer, but the ones it was not.

Anticipated questions are the easy ones

Every record answers the questions its author anticipated. That is what authoring is: you decide what matters, and you address it. The trouble is that the questions an author anticipates are a tiny and self-flattering subset of the questions that can be asked. They are the questions where the author already knows the answer is fine. Nobody volunteers the line of inquiry that would sink them. So a record that handles all its anticipated questions beautifully has demonstrated only that the author had good judgment about which questions to invite — which tells you nothing about the decision and everything about the author's instinct for self-defense.

The unanticipated question is the whole game. Why did you not consult the source that would have contradicted you? What did the rule say before it was convenient to read it this way? If I change this one input, does the outcome move the way you claim it would? These questions cannot be pre-answered with prose, because prose is the author's medium and the author is the interested party. They can only be answered by structure — by a record that carries the underlying material in a form the interrogator can pull on directly, without going back through the author at all.

You cannot anticipate every question. You can only build a record that does not depend on having anticipated them.

This is the difference between a record that explains and a record that exposes. An explanation is the author's answer to the questions the author chose to hear. An exposed record — the actual evidence in the order it was consulted, the rules as they stood at the moment of decision, enough state to replay the thing and watch where it lands — is a surface the interrogator can question without the author's cooperation. The first is a performance of completeness. The second is a structure that survives hostility because it was never relying on the hostility staying away.

Build for the reader who does not want it to be true

The practical instruction that falls out of this is uncomfortable, because it inverts the instinct of nearly everyone who writes a record. Do not write for the person you hope will read it. Write for the person you fear will read it. The friendly reader needs no help; they were going to accept the account anyway. It is the adversary who reveals whether the account was ever more than an account — and if you build for the adversary, the friendly reader is served as a side effect, because everything that satisfies hostile scrutiny also satisfies sympathetic scrutiny. The reverse is not true. Nothing built for the friendly reader survives the hostile one.

This is why a Decision Receipt is a different object from a report, even when they describe the same decision. A report is addressed to whoever will be reassured by it. A receipt is addressed to whoever will try to break it, and is built to give that person every input they would need to try — not as a concession, but as the source of its standing. A receipt that withheld the material an interrogator would want would be admitting, structurally, that it expected not to survive the asking. The willingness to be interrogated is not a feature bolted onto the record. It is the thing that makes the record evidence rather than testimony about evidence.

And notice what this does to the burden of trust. A record that passes the friendly reading asks you to trust the author — you accept the account because it reads as competent and complete, which is a judgment about the author, not the decision. A record that passes the interrogation asks you to trust nothing, because you can check it. The point of building for hostile interrogation is not to win the argument. It is to make the argument resolvable by something other than whose word you take. A record you have to take on faith has failed the test before anyone asks it a single question, because the first hostile question is always the same one: why should I believe you? — and the only good answer is to hand over what would let them stop having to.

So when you are handed an account of a decision and asked to accept it, do not read it for whether it hangs together. Read it for whether it could withstand someone who wanted it not to. Ask it the question it was not built for, and see if it still has an answer. If it does, you are holding something that earned its place. If it goes quiet — if the answer turns into the author's reassurance that of course it was all considered — then the completeness was always a surface, and you have just found its edge.

— Dispatches · Summit Cognitive