Questions and answers

What is the idea, in one sentence?

A consequential automated decision earns trust not by explaining itself but by carrying a record a stranger could examine, replay, and contest. Everything else is detail about where that standard applies, at what cost, and within what limits. The doctrine line underneath it does not move: if a decision cannot be replayed, it should not be trusted.

What does "admissible" mean for a decision?

Admissibility is the property of a decision whose record could be taken by someone who was not in the room and used to arrive at the same result. The word is borrowed from evidence law: an admissible exhibit is one built to be examined, challenged, and weighed by a party who does not trust the side that produced it. A decision is admissible when its record passes that same test — not when it is reassuring, but when it could be shown to be wrong and survives the attempt.

Why does a record have to be built to be contested?

Because a record whose only function is to reassure has no way of proving it deserves to. The value of a record is not in how settling it is; it is in how cleanly it could be shown to be wrong. A polished account that offers nothing to argue with is not neutral — it is suspect, because if the decision genuinely happened the way the record says, exposing the inputs and the rules costs the author nothing.

What is a Decision Receipt?

A Decision Receipt is the record a consequential decision leaves behind, built to be contested rather than merely believed. It carries the real evidence the decision rested on — in the order it was consulted — the rules that were active at the time, and enough state to replay the decision and check whether the outcome follows. A receipt that carries those things is more authoritative than any memo, precisely because it invites the attack and withstands it.

Does this mean every decision needs a receipt?

No, and a standard that demanded one would be self-defeating. The overwhelming majority of automated decisions are trivial, reversible, and never contested; building full admissibility into them would be waste dressed as rigor. The discipline is to draw the line honestly and spend the cost only on consequential decisions — the ones that land hard on a specific person, the ones someone might reasonably demand to contest.

What are the four properties of an admissible decision?

Provenance, replay, rules, and standing. Provenance preserves the actual inputs the decision rested on, frozen as they stood. Replay lets the decision be re-run against those inputs to check that the outcome follows. The governing rules are captured in the form they held at decision time. And standing puts someone on the other side with the entitlement to demand the record and a forum to contest it. A record missing any one of them is incomplete.

What does provenance preserve, and why frozen?

Provenance preserves the real sources a decision consulted, in the order it consulted them, captured as they stood at the moment of use rather than summarized later. Frozen, because inputs drift: a record reassembled from data that has since changed confronts a different decision than the one that was actually made. Order matters too — a fact read before a contradicting fact does different work than the same fact read after.

Why capture the rules as they stood at the time?

Because policies drift, and a record that depends on the current policy to explain a past decision can be quietly rewritten by changing nothing about itself. The rules have to be carried inside the record in full, frozen as they were at the instant of decision, so the question did this decision follow the rules it was bound by can be answered without anyone deciding after the fact which version to quote.

What is standing, and why is it one of the properties?

Standing is the entitlement of the affected party to demand the record and a forum in which to contest the outcome. A record is only half of accountability; the other half is a contestant with the means to actually challenge it. A record answerable to no one is not accountability — which is also why standing cuts the other way, against a record open to everyone. It is built for the person with a reason to look.

Can you really replay a modern model? Aren't they non-deterministic?

Left unconstrained, yes — sampling, hardware, and silent model updates all introduce variation. But determinism is not a property you discover in a system; it is a constraint you impose on the decisions that need it: pin the model version, fix the sampling parameters or record the seed, capture the inputs, freeze the rules. Where full determinism cannot be reached, the record degrades honestly — it states what was fixed and what was not.

If you cannot open the model's black box, what does replay prove?

Admissibility draws the boundary at the decision, not the model's interior. It does not claim to capture a large model's internal reasoning, which may not be fully capturable at all. What it makes contestable is the inputs, the rules, the outcome, and enough determinism to re-run the path between them. Accountability lands at the boundary, not in the weights — and a record of the boundary is not a record of nothing.

Isn't this just logging?

An ordinary log is written by the system, for the system's operators, to answer what happened during normal running. A Decision Receipt is written for a reader who does not trust the system. It preserves the actual inputs frozen as they stood rather than asserting that a decision occurred; it lets you re-run the decision rather than read about it; and it answers to the person with standing to contest it, not the operator. The materials are familiar; the specification is different, and the specification is the point.

Isn't explainability enough?

Explanation is a useful capability and worth keeping — for debugging, for detecting bias, for operator intuition. But an explanation is the system's own account of itself: post-hoc, produced by the interested party, and in the general case unfalsifiable. It can even be sincere and name the wrong cause. Every accountability regime ever built has a place for testimony, and none rests on it alone — least of all on the testimony of the party whose conduct is in question.

Isn't this too expensive to scale?

It would be, if the claim were that every decision must be made admissible — that claim is economically absurd, and we agree. But the cost is scoped to consequential decisions, a small fraction of the total. And it has a shape worth noting: it is paid once, at the moment of decision, not repeatedly at every later moment of doubt. Admissibility is cheap to maintain and expensive only to exercise — and the exercise is rare, because most decisions are never contested.

Doesn't this slow everything down?

It adds work at the moment of consequential decisions — but those are rarely the latency-critical ones, and a decision that lands hard on a person can usually afford the milliseconds it takes to capture its basis. "Slows everything down" also measures the wrong clock. The speed that matters in an accountability failure is the days or quarters lost when a challenged decision cannot be reconstructed at all. The cost is paid once, up front; its absence is paid repeatedly.

Is this surveillance?

It can be, built carelessly — a sprawling store of sensitive data retained beyond its purpose is exactly the harm to avoid. The answer is to notice who the record is for. Admissibility is an instrument of the governed, not the governor: a record built for contestability should be scoped to the decision it documents, retained only as long as the contest window, and exposed to the challenger, not the world. Built that way, it replaces an opaque system that watches people with a bounded record those people can use to push back.

Does any regulation require this?

No instrument we know of mandates this specific architecture, and anyone who tells you the law currently requires it is selling you something with a false label. The claim is about direction, not the letter of a rule. The obligations that recur across jurisdictions — record-keeping, transparency, oversight, an increasingly explicit right to contest — each have a weak satisfiable form and a durable one. Admissibility is a bet on how they get read once they are actually tested.

Can't I just trust the vendor to handle this?

You can outsource building the record layer; a good vendor should be building toward exactly this. What you cannot outsource is being answerable. When a regulator, a court, or an affected person asks why a decision was made about someone, the organization that deployed the system answers — not the model vendor three layers up. "The model said so" is increasingly not a legal answer. Trust the vendor the way admissibility trusts anyone: on a record you could check, not on their word.

Who is this for?

Two audiences. The people who build and deploy systems that make consequential decisions, who will be the ones answering when a decision is challenged and would rather be early than retrofit under pressure. And the people on the other side of those decisions — the affected party, the regulator, the contestant — who deserve a record they can use to push back rather than a polished account they can only resent.

Isn't the author selling something?

Brian builds in this space; his company works on the record layer for automated decisions. That is a reason to discount what is being sold, and not a reason to discount the argument — because the argument can be taken and built by anyone, bought from anyone, or rejected outright. The ask is narrower than "buy this." It is "stop treating the record as optional where the stakes are real."

What is Dispatches, Warrant, and Admissible Reality?

Three forms of one argument. Admissible Reality is the long-form book — the case for decision admissibility made in full. Warrant is the podcast, where the same questions are tested in conversation. Dispatches is this series of short essays, written in the order the questions arrived. All three are edited by Brian C. Long.

Where should I start reading?

Begin with the reader's guide, which names five essays to start with and five themed paths through the rest — Evidence, Provenance, Standing & Contestability, Method & Replay, and Legitimacy & Accountability. If you read only one, read The receipt and the rebuttal: it states the test at the heart of everything else.