The warranty on a decision

Consider two dishwashers, priced the same, sitting side by side in a showroom. One comes with a ten-year warranty; the other is sold as-is. Before you have read a single specification, before you know anything about the motors or the seals, you already know something about the two machines — not from the manufacturers' claims, which are identical and equally cheerful, but from the terms attached to them. The warranty tells you that one seller has done the arithmetic on how often their machine fails and concluded they can afford to pay when it does. The other seller has either done the same arithmetic and reached the opposite conclusion, or has declined to do it at all. Either way, the terms have told you more than the brochure.

This is one of the oldest results in the economics of information, and it is worth stating plainly because it runs against intuition. A warranty is not primarily a promise about the product. It is a signal about the seller's private belief in the product — and it is a credible signal for exactly one reason: it is expensive to offer if that belief is wrong. Michael Spence won a Nobel Prize for formalizing the general point. A signal carries information only when it is costlier to send falsely than truthfully. Anyone can say their product is good; saying so costs nothing and therefore reveals nothing. But agreeing to bear a defined cost when the product fails is cheap for the firm whose product rarely fails and ruinous for the firm whose product often does. The two firms, who would say identical things, cannot afford to do identical things. The warranty is the thing they cannot both afford.

The warranty as a signal, not a promise

Hold onto why this works, because the mechanism matters more than the conclusion. The information in a warranty does not come from its words. It comes from the fact that the seller has skin in it. A firm offering a generous warranty is not telling you their product is good; they are revealing a private estimate of its failure rate by putting money behind that estimate. You do not have to trust their honesty. You only have to trust their self-interest, which is the one thing a market lets you rely on completely. The warranty converts a claim you cannot check into a bet you can watch them place.

And the corollary is where it gets sharp. If the willingness to stand behind a product is a signal, then the refusal to stand behind it is a signal too — one sent in the same currency and just as hard to fake. The seller who disclaims all liability, who prices confidence into the pitch and pushes every downside onto the buyer, has told you something precise about their own estimate of the product. Not that it is certainly bad. But that they are not willing to bet it is good, on terms where being wrong would cost them. In a market, that reluctance is data. It is the seller declining, in the only language markets fully trust, to back their own brochure.

A vendor who will not stand behind a decision is not withholding a courtesy. They are pricing the decision — and the price they name, in the only currency markets believe, is zero.

Now turn this instrument toward the thing this series is about. Today, the vendors of automated decision systems sell confidence and disclaim liability. The pitch is fluent about accuracy and the contract is silent about consequences; the terms of service move every downside of a wrong decision onto the buyer, and past the buyer onto whoever the decision lands on. Read as a signal, this is not a neutral commercial default. It is the dishwasher sold as-is, in a showroom where every machine happens to be sold as-is, and where the buyers have gotten so used to it that they have stopped noticing the terms are telling them anything at all.

You cannot warrant what you cannot inspect

Here is the part that ties the warranty to everything else in this ledger. A warranty is only offerable by a firm that can actually assess the risk it is underwriting. You do not agree to pay for failures you have no way to count. Before a seller can warrant a dishwasher they must know, roughly, how dishwashers fail — which means they must be able to open one up, find the fault, and tell an honest failure from a customer who ran it empty for a decade. The warranty rests on an inspectable object. Take away the ability to inspect and the warranty becomes an uncalculated gamble that no rational firm would make and no honest one would offer.

The same precondition governs decisions, and it is far less often met. To warrant an individual decision — to agree to bear a defined cost if that decision cannot be justified on the record — a provider must be able to answer, after the fact, whether the decision was sound. That requires a decision that can be examined: replayed against the inputs that were actually in front of it, tested against the rules that were actually in force, contested by the party with the most reason to find the flaw. A decision that leaves no such trace cannot be assessed, and what cannot be assessed cannot be underwritten. So the record is not merely an accountability artifact, filed away in case someone complains. It is the precondition for the assurance instruments — warranties first, insurance after — that let the risk of a decision be priced and transferred instead of silently dumped on whoever is downstream. You cannot warrant what you cannot inspect, and you cannot inspect a decision that was never built to be examined.

This reframes what a Decision Receipt is worth. Its first value is to the affected party, who gets a surface to contest. But its economic value is larger and quieter: it is the thing that makes the decision underwritable. A provider who keeps genuine records has, without necessarily intending to, built the substrate on which someone can finally calculate the risk of their decisions and agree to carry it. A provider who keeps none has foreclosed that market for themselves. They could not offer a warranty even if they wanted to, because they have no way to know what they would be signing up to pay.

From disclaimer to warranty

Markets for goods that cannot be casually verified tend to mature along a predictable path. They begin sold as-is, confidence included and liability excluded, because at the start no one can tell the good products from the bad ones and so no one can charge for the difference. They mature when someone works out how to make quality visible — and the arrival of the warranty is the usual marker of that maturity, because the warranty is what becomes possible the moment quality can be assessed rather than merely asserted. The used-car market did not become trustworthy through the moral improvement of used-car dealers. It became trustworthy through inspections, certifications, and warranties: instruments that let a seller who actually had a good car prove it by betting on it, and thereby separate themselves from the seller who could only say so.

The market for machine decisions is somewhere near the beginning of that path, and the warranted decision is what its maturity will look like. Not a vendor claiming higher accuracy — every vendor claims that, and the claim reveals nothing. A vendor willing to stand behind a specific decision's soundness, to bear a defined cost if it cannot be justified on the record when someone comes to check. The ability to make that offer will not be evenly distributed, and that is precisely the point. It will belong to the providers who built records worth warranting, and it will be unavailable to the ones who built only disclaimers. The offer itself will do the sorting that no brochure can, because it is the one claim about a decision that is expensive to make falsely.

This is the through-line that connects the warranty to the certification that precedes it and the insurance that follows. Each is an instrument for turning an unverifiable quality into a priced and transferable risk, and each rests on the same foundation: a decision that can be inspected. So when a provider tells you their decisions are excellent but declines to stand behind any one of them, listen to the second half of the sentence. They are not being modest. They are telling you, in the language markets trust most, either that their decisions could not survive the inspection a warranty would invite — or that they never built the record that would let anyone, including themselves, find out.

— Dispatches · Summit Cognitive