Build or buy the record

Every capable engineering organization's first instinct, faced with a new infrastructure need, is to ask whether it should just build the thing. Usually this is the right instinct to interrogate honestly, and often the right answer is yes — the organization understands its own systems, controls its own roadmap, and avoids paying a vendor's margin for something it could stand up itself. The standard build-versus-buy framework handles this well: weigh the total cost of ownership against the license, discount for focus and opportunity cost, decide. But the decision record breaks the standard framework, and it breaks it in a way that is easy to miss because the break is not about cost at all. It is about who the artifact is for, and that single difference inverts the usual logic.

The framework that usually applies

For most infrastructure, the thing being built serves the organization that builds it. A data pipeline, an internal tool, a deployment system — these are consumed by the same people who make them, and their quality is judged by those same people. If you build your own and it works for you, it works, full stop; there is no external examiner whose opinion of the artifact determines its value. This is why the build calculus reduces so cleanly to cost and focus: the only stakeholders are inside the building, so the only questions are whether you can build it well enough and whether your time is better spent elsewhere. When the answer to both is favorable, building wins, and rightly.

The instinct to apply this same reasoning to the decision record is overwhelming, because on the surface it looks like ordinary infrastructure. It captures events, it stores them, it retrieves them; an engineering team could obviously build that, probably in a quarter, probably better-fitted to its own stack than anything bought. And if the decision record were consumed by the same people who build it, that reasoning would be correct and this essay would be over. But it is not consumed by them. It is consumed by someone outside — the party who contests a decision, the regulator who examines it, the court that weighs it — and that person judges the record by a criterion the builders cannot satisfy from inside.

The value of a record is set not by how well it is built but by how much an adversarial outsider believes it. And belief in a record falls as the recordkeeper's stake in its contents rises.

The structural problem: you cannot be the trusted keeper of the record that judges you

Here is the difficulty in one sentence: the institution making the decisions is the party whose conduct the record will be used to evaluate, which means the institution is the last party that should be in sole control of the record. This is not a slur on anyone's integrity. It is the same principle that separates the auditor from the audited, the referee from the team, the investigator from the operator. We do not require those separations because we assume bad faith; we require them because a record kept by an interested party is worth less as evidence regardless of whether it was actually tampered with, since the mere ability to have tampered with it is enough to discount it. An outsider examining a decision does not ask only "is this record accurate?" They ask "could this record have been shaped, after the fact, by the party it exonerates?" If the answer is yes — if the institution authored the record, controls the system that holds it, and could in principle revise it — then the record carries the built-in discount of self-interest, and no amount of engineering quality removes that discount, because the discount is not about quality. It is about custody.

This is what the standard build calculus cannot see. A team that builds its own decision record can build it perfectly — well-structured, complete, faithful — and still produce an artifact that is worth less than a bought one, not because the bought one is better engineered but because the bought one is kept differently: by a party with no stake in what any individual decision says, under an arrangement the institution cannot silently alter, in a form designed to be believed by someone who assumes the institution would revise it if it could. The value being purchased is not the capture. It is the independence of the custody. And independence is, definitionally, the one property an in-house build cannot have, because in-house means kept by the interested party.

The two hidden costs of building

Even setting the custody problem aside, the build calculus for a decision record understates cost in two specific ways.

The first is the standardization cost. An internal record is built to the institution's own idea of what matters, in the institution's own shape. But the whole point of the record is to be read by outsiders, and outsiders — regulators, courts, counterparties — increasingly expect records that mean the same thing across institutions, because that is what lets them examine, compare, and rely on them. A bespoke internal format wins nothing from the convergence the category is moving toward; worse, it strands the institution on a private standard at exactly the moment a common one is forming, so that the eventual cost of building includes the later cost of conforming to whatever the outside world settles on. You do not just build the record once. You build it, and then you rebuild it to match the standard you did not help set.

The second is the durability cost, which is really a maintenance cost in disguise. A decision record is only useful if it faithfully reflects what was true at the moment of decision, years later, after the systems that produced it have changed, been deprecated, or been replaced. Keeping a record that remains a faithful account across all that change — surviving migrations, schema changes, and the ordinary churn of an engineering organization — is not a build. It is a permanent operating commitment, and it competes forever with every other priority the team has. Internal records rot for exactly this reason: they are always someone's second job, and the moment the person who understood the format leaves, the record's fidelity to the original decision quietly degrades. A vendor whose entire existence is the record does not have a second job. Keeping the record faithful is the only job.

Where building is still right

None of this is an argument that buying always wins; it is an argument that the decision turns on a different axis than usual. The honest test is not "can we build it?" — a capable team almost always can — but "will an adversarial outsider credit a record we keep ourselves?" Where the decisions at issue are genuinely low-stakes, never contested, and answerable to no one outside the building, the custody problem does not bind, and building may be entirely sensible; there is no outsider to convince. The custody problem only becomes decisive as the stakes rise and the audience moves outside — which is precisely the population of decisions for which the record matters at all. So the rule that falls out is uncomfortable but clean: for the decisions where a record is worth having, it is worth having in a form you did not solely author and cannot silently change; and for the decisions where you could credibly keep the record yourself, the record was not going to be contested anyway. Build-versus-buy, for evidence, is not a cost question with a custody footnote. It is a custody question with a cost footnote — and getting the order right is the whole decision.

— Dispatches · Summit Cognitive