The provenance of a prompt

When a generative system gives an answer, we are trained by long habit to study the answer. We read it, we weigh it, we decide whether it is good enough to act on, and if we keep anything at all we keep the output — the paragraph it wrote, the score it returned, the recommendation it made. This is a reasonable instinct inherited from a world of fixed reference works, where the answer was the thing and the question was a transient nuisance you discarded once it was satisfied. It is also, now, a quietly dangerous instinct, because with a model the answer is not a fact retrieved from a stable store. It is a thing produced, in the moment, by everything that was put in front of the model immediately before it spoke.

Consider what actually shapes a model's reply. There is the prompt itself — the precise wording of the question, which the model is acutely sensitive to in ways that no encyclopedia entry ever was. There is the retrieved context: the documents, records, and snippets that were pulled in to ground the response, often selected automatically and invisibly, often the real source of whatever authority the answer appears to carry. And there are the instructions in force — the system's standing directions about how to behave, what to refuse, whom to defer to, which constraints to honor. Change any of these and you can change the answer entirely, while the answer itself bears no visible mark of what produced it. The output arrives looking like a conclusion. It is really the last line of a conversation we did not write down.

The inputs are evidence

If a consequential decision rests on what a model said, then accounting for that decision means being able to show why the model said it. And the why does not live in the output. It lives in the inputs — in the exact prompt, the exact retrieved context, the exact instructions that were active at that instant. These are not metadata. They are not logistics. They are the evidence on which the decision actually turned, in the same sense that the documents in front of a human decision-maker are the evidence on which a human decision turns. To keep the answer and discard the inputs is to keep the verdict and burn the case file.

The reason this matters is that an answer, by itself, cannot be contested. Hand someone a model's output and ask them to dispute it, and they have nothing to grip. They can disagree with it, but disagreement is not refutation. To refute it they would need to know what it was responding to — whether the prompt smuggled in an assumption, whether the retrieved context was incomplete or skewed, whether the instructions tilted the system toward a particular kind of reply. Without the inputs, the most they can do is offer a competing assertion. The provenance of the answer is what converts it from an opinion into something a person can examine, attack, and either break or fail to break.

An output without its inputs is a verdict without a trial. You can read it, but you cannot question it.

There is a temporal trap here that deserves naming. The inputs to a model are unusually perishable. A retrieval index is updated; the document that was pulled in last week is no longer the document that would be pulled in today. A system's instructions are revised; the constraints that governed the answer are quietly replaced by new ones. The prompt, if it was assembled on the fly from a template and some live data, may never have existed as a fixed artifact at all unless something captured it. Reconstruct any of this after the fact and you are not recovering what happened — you are producing a plausible reenactment with today's parts, and calling it the original. The only honest record of a generative decision is one captured at the moment of the decision, because that is the only moment the real inputs exist.

What provenance asks of us

To take this seriously is to change what we preserve. A record that carries only the model's output is, for accountability purposes, nearly empty — it tells you what was decided but offers no surface on which to ask whether it should have been. A record worth the name carries the answer together with its provenance: the prompt as posed, the context as retrieved, the instructions as they stood. With those three, a decision becomes replayable. You can feed the same inputs through and see whether the same output emerges, and where it diverges the divergence is itself informative — it shows you precisely what has shifted since. Without them, replay is impossible, and a decision that cannot be replayed cannot really be defended; it can only be repeated more confidently.

This is not a counsel of suspicion toward generative systems. It is the opposite. A model whose answers travel with their provenance is more trustworthy than one whose answers arrive bare, because its answers can be checked rather than merely believed. The provenance is what gives someone standing to contest the decision, and a decision that can be contested and survives the contest has earned an authority that an unexamined one never can. We do not protect ourselves from these systems by refusing to use their outputs. We protect ourselves by insisting that the outputs come with the questions, the context, and the rules that made them — so that the record of a machine-mediated decision is a record of a decision, and not just a record of an answer.

So the discipline is simple to state and easy to neglect: when a model's answer is going to matter, keep what was fed to it with the same care you keep what it returned. Preserve the prompt, the context, and the instructions in force, frozen as they were. The answer is the part everyone looks at. The provenance is the part that lets anyone look back.

— Dispatches · Summit Cognitive