The translator in the middle

A machine produces a number. Somewhere downstream, a person lives with a consequence — a loan denied, a claim flagged, a name moved up or down a list that determines who gets seen and who waits. Between those two events there is a gap that almost nobody names, because it does not look like a step. It looks like nothing at all. The number simply arrives, and the consequence simply follows, and the gap between them — the place where the number would have to be turned into something a person could read and answer — is left empty. We have built systems that are very good at the first half and the third half and have quietly skipped the middle.

The middle is translation. Not translation between languages, but between kinds of statement: from a thing the machine can say to a thing a human can act on. A score of 0.83 is not a reason. It is the residue of a reason, compressed past the point where a person could recover what it means. To make it usable, someone has to render it back into a sentence: this application was declined because the model weighed these factors against this threshold, under this policy, and landed here. That sentence is the decision made legible. Without it, there is no decision a person can engage — only an outcome that happened to them.

The trouble is that translation is work, and work that nobody is assigned tends not to get done. The model output is cheap; it falls out of the system for free. The translation is expensive; it requires someone to hold the number, the inputs, the rule, and the person all in mind at once and produce an account that fairly represents what occurred. So the path of least resistance is to skip it — to take the number and treat it as if it were already the sentence, as if a score could explain itself by being shown. It cannot. A score shown is not a score understood.

A number passed through untranslated is not an explanation that happens to be terse. It is the absence of an explanation, formatted to look like one.

Why the number does not speak for itself

There is a seductive idea that a sufficiently precise output needs no interpreter — that the figure is its own justification, and that asking for more is asking for decoration. This gets the situation exactly backward. Precision is not legibility. A figure can be perfectly precise and entirely opaque to the person it governs, because precision tells you how confident the machine is and legibility tells you what the machine did and why it bears on you. Those are different questions, and the second is the one a person actually needs answered. The number answers the first and is silent on the second, and that silence is filled, by default, with the assumption that the number must mean something good enough to act on.

What gets lost in the untranslated pass-through is everything that would let a person check the result against their own situation. The inputs the system actually used — not the ones it was supposed to use, the ones it did. The rule it applied them under. The threshold it compared them to. The order in which evidence accumulated, which can change what the same facts add up to. A translation worth the name carries those things forward; it turns the score into an account that names its own basis. A pass-through carries none of them, because a pass-through is, definitionally, the decision to carry nothing — to let the number stand alone and hope it lands.

The translation is the record

This is where translation stops being a courtesy and becomes the act that makes accountability possible at all. A Decision Receipt is, at bottom, that translation made durable. It is the work of turning what the machine produced into an account a person can read, written down at the moment it happened and held in a form that does not decay. The receipt is not a transcript of the model and it is not a press release about the outcome. It is the middle term — the sentence that says, in language a person can engage, what the number meant and what it did.

Notice what this implies about who the translation is for. It is not for the engineer, who can already read the number, and it is not for the system, which has no use for sentences. It is for the affected party and for anyone who later stands in for them — the reviewer, the auditor, the second reader who arrives after the fact and needs to know not just what was decided but what the decision rested on. The translation is the only artifact that speaks to all of them in a language they share. Skip it, and you have left every one of those people holding a figure they cannot interrogate, in a moment when interrogation is exactly what fairness requires.

The cost of skipping the middle is rarely visible at the moment it is skipped, because nothing breaks. The number flows, the outcome lands, the system reports success. The cost shows up later, when someone wants to understand or contest what happened and discovers there is nothing to understand or contest — only a residue and a result, with the reasoning that once connected them never having been written down. By then the translation cannot be reconstructed honestly, because the only honest time to write it was when the decision was live and the basis was still present. Translation is not something you can add after the fact without it becoming a story instead of a record.

So the question to ask of any system that turns a machine's output into a human consequence is simply: who, here, is doing the translation? If the answer is no one — if the number is passed straight through and the affected party is handed a figure and a verdict and left to infer the rest — then the system has not made a legible decision. It has made an outcome and declined to account for it. The number was never the hard part. The sentence in the middle, the one almost nobody writes, was always the whole job.

— Dispatches · Summit Cognitive