A record you can read

Suppose an institution does everything we have been asking it to do. It keeps the inputs to a decision, preserves the model and policy in force at the time, signs the record so no one can quietly alter it, and stands ready to hand the whole thing over to the person the decision ruled against. The file is complete. Every field a careful auditor could want is present and accounted for. And the person it affects opens it, reads three lines, and understands nothing — not because they are slow, but because the record was written in a language meant for someone else. They are holding their own account of their own loss, and it might as well be in code.

This is the failure that survives all the other repairs. We have spent a good deal of effort arguing that decisions need records — that an output without its sources is an assertion, that a verdict without minutes is not accountable, that a right to contest is empty without a record to contest from. All of that is true and necessary. But a record can clear every one of those bars and still fail the affected party, because completeness and legibility are different properties, and only one of them is usually designed for. A record can contain everything and communicate nothing.

The reason this matters is not aesthetic. A record that the affected party cannot read does not merely inconvenience them; it relocates power. The decision was already made by an institution with more expertise, more resources, and more time than the person on the receiving end. If the account of that decision is legible only to people with the same expertise, then exercising the right to contest requires hiring one of them — a lawyer, a technical expert, a translator standing between the person and their own case. The record was supposed to equalize. An illegible one re-imposes exactly the asymmetry it was meant to correct, and dresses the re-imposition as transparency.

Two readers, two records — or one that serves both

It helps to be precise about who a record is for, because a Decision Receipt has more than one reader and they do not need the same things. There is the expert reader — the auditor, the regulator, the opposing technical witness — who needs the record to be exhaustive, exact, and unforgiving: every input, every version, every timestamp, nothing rounded off. And there is the affected reader — the person denied, downgraded, flagged, or refused — who needs first of all to understand what happened to them and why, in terms they can act on. These are not competing demands. They are layers of the same record.

The mistake is to assume that serving the expert reader automatically serves the affected one, that a sufficiently complete record is by definition a sufficiently clear one. It is not. Exhaustiveness and clarity pull in opposite directions: the more complete a record becomes, the more it tends toward the technical, and the technical is precisely what the affected reader cannot use. An institution that builds only for the auditor produces a record that is admissible and unreadable — defensible in a proceeding the affected party can barely enter. Both readers are owed something, and the something is different for each.

A record the affected party cannot read is not a window into the decision. It is a wall with the word "transparency" printed on it.

What the affected reader is owed is not a simplified record — simplification that drops the load-bearing facts is just a second injustice — but a legible surface over a complete one. The plain-language account of what was decided and on what basis, sitting on top of the exact, exhaustive record, each traceable to the other. The affected party reads the surface and understands their case. If they want to fight, they, or someone they bring, drops to the layer beneath and finds it all still there, unrounded. Legibility is not subtraction. It is a second face on the same artifact, written for the reader the first face forgot.

Comprehension is the start of standing

There is a tendency to treat the right to an explanation as satisfied the moment a record is produced and made available. Produced, available, therefore delivered. But a right to contest does not begin when a record changes hands; it begins when the person can understand what they are holding well enough to know whether they have a case. Comprehension is not a nicety that comes after standing — it is the first thing standing requires. You cannot decide to appeal a decision whose grounds you cannot make out. You cannot point to the error in a record you cannot read. An account that arrives in a form its intended reader cannot parse has not yet been delivered, however completely it was produced.

This sets a standard that is harder than it sounds, because it cannot be met by the producing institution grading its own work. Whether a record is legible is not a fact about the record; it is a fact about the reader meeting it. The only honest test is the one taken from the affected party's side: hand a representative reader the account and ask whether they can state, in their own words, what was decided, what it rested on, and where they would push if they wanted to push. If they cannot, the record has failed its second reader, no matter how it would fare with the first. Legibility is measured at the point of reception, not the point of production — which means it cannot be asserted by the institution, only demonstrated against a reader who owes it nothing.

None of this lowers the bar on completeness. The exact record still has to exist, signed and preserved and ready for the expert who will read it adversarially. The argument is only that completeness is necessary and not sufficient — that a record discharges its duty to the affected party not when it contains the truth but when it conveys it to the one person who most needs to understand it. We have learned to ask whether a decision kept its evidence. The next question, the one that decides whether the keeping was for the affected party or merely about them, is plainer and harder: can the person it ruled against read it? A record that cannot answer yes has done its bookkeeping and failed its reader, and those are not the same thing as having done its job.

— Dispatches · Summit Cognitive