Field Notes · Infrastructure & Economics August 21, 2026 · 5 min read

Before you buy the architecture, identify what must be true for it to work.

Consequential technical commitments arrive for approval as advocacy: a proposal explaining why the platform, the vendor, or the rebuild is the right choice. Advocacy is the wrong genre for capital protection. The discipline that actually protects the commitment is falsification: decompose it into the assumptions that must hold for the money to come back, grade each assumption by how it can be verified, rank by damage-if-false, and buy evidence for the expensive ones before signing. This note lays out the method and the failure patterns it prevents.

Advocacy is the default genre, and it hides the load-bearing parts

A proposal that reaches an executive for signature has already survived internal selection, which means it arrives argued for. Its structure is a case: the problem, the option chosen, the benefits, the cost. What that structure systematically omits is the commitment's dependency graph: the specific conditions under which the projected benefits actually materialize. Those conditions exist whether or not anyone writes them down. Signing the proposal means betting on all of them at once, and the organization usually discovers which ones were load-bearing at the moment one fails.

The corrective question is short: what must be true for this to work? Not "why is this good," which invites more advocacy, but "list the propositions whose falsity would make this money wasted." The question converts a persuasive document into a testable one, and it can be asked by any executive, of any proposal, without technical depth.

The assumption classes

Across architecture, vendor, infrastructure, and build-versus-buy decisions, the assumptions that carry commitments fall into a small number of recurring classes:

• Demand assumptions. The projected usage that justifies the capacity, the licenses, or the build. Frequently inherited from a growth model rather than measured, and frequently the assumption doing the most financial work.
• Technical behavior assumptions. The system performs as projected at the load, data volume, and integration surface it will actually face: the claims examined throughout this corpus, including how to decompress a vendor's labels into testable statements.
• Economic assumptions. The cost model's shape holds: unit costs at scale, pricing stability from providers, the exchange rate between committed spend and flexibility. Discounts have expiry dates; workloads have growth curves; the model's snapshot is not a contract.
• Organizational capability assumptions. The team can build, operate, and maintain what is being approved, with the people it actually has, alongside the work it already does. The most routinely overestimated class, because the proposal's author is usually describing a team at its best month.
• Counterparty durability assumptions. The vendor remains solvent, the product remains supported, the terms remain recognizable at renewal, the open-source project remains maintained. Multi-year commitments assume multi-year counterparties.

The method: enumerate, grade, rank, buy evidence, restructure

1. State the commitment precisely. Amount, duration, exit terms, and what is being foreclosed: the alternatives that become unavailable once this is signed. A commitment's real size includes its lock-in.
2. Enumerate what must be true. Write the assumptions as falsifiable statements. "The platform scales" is advocacy; "the platform sustains our projected peak concurrency at a cost per transaction within 20% of the model" is an assumption someone can test.
3. Grade each assumption's verifiability. Four grades: verified (evidence exists now); cheaply verifiable (a test, a reference architecture inspection, a measured pilot could establish it before signing); verifiable only in production (the truth arrives after commitment); unverifiable (a genuine bet). The grading is where proposals get honest, because advocacy tends to present grade-three and grade-four assumptions with grade-one confidence.
4. Rank by damage-if-false times uncertainty. An assumption that is probably true but catastrophic if false outranks a doubtful one with cheap consequences. The top of this ranking is where diligence hours and pilot budget belong.
5. Buy evidence before commitment. For the expensive cheaply-verifiable assumptions, spend the small money: the load test, the paid pilot, the customer references asked the specific question, the contract-terms review. Evidence purchased before signing is the cheapest insurance the decision will ever be offered.
6. Restructure around what cannot be verified. Grade-three and grade-four assumptions do not forbid the commitment; they shape its form. Shorter initial terms, staged expansion gated on measured results, exit clauses priced while leverage exists, and parallel-run periods are all ways of converting an unverifiable assumption into a survivable one. If the counterparty resists every such structure, that resistance is information about their own confidence.
7. Write the decision down. A short memo: the commitment, the assumptions, their grades, the evidence obtained, the bets knowingly taken, the conditions that would trigger reconsideration. The illustrative sample memorandum on this site shows the form. The memo's value compounds later: when an assumption fails, the organization knows whether it was a bad bet or a bad process, which determines what to fix.

The failure patterns this prevents

Three patterns account for much of the waste in large technical commitments, and each is a specific skipped step. The all-assumptions-at-once pattern: a multi-year commitment signed on an unexamined bundle, where any single failed assumption (demand below model, migration harder than projected, key discount expiring) strands the rest. The confidence-transfer pattern: verifiable claims accepted on vendor assurance because testing felt slow, with the truth arriving in production at production prices. And the silent-foreclosure pattern: the commitment's lock-in never priced, so the organization discovers at year two that leaving costs more than staying, which was knowable, and negotiable, at signature.

The strongest objection

The serious objection is that some commitments must be made under irreducible uncertainty: capacity bought ahead of demand, platforms chosen before the market resolves, first-mover positions that cannot wait for evidence. Correct, and the method handles it without modification. Falsification discipline does not forbid bets; it identifies which parts of the commitment are bets, prices them, and makes leadership take them knowingly. There is a real difference between an organization that committed early because the upside justified an explicit grade-four bet, and one that committed early because nobody graded anything. The first can size the bet and structure the exits. The second finds out what it bet when it loses. Speed is not the casualty either: the decomposition is days of work, not months, and it is faster than the quarter spent unwinding a commitment whose load-bearing assumption was never named.

Conclusion

Every expensive technical commitment is a bundle of propositions about demand, behavior, economics, capability, and counterparties. Approval by advocacy signs the bundle unread. The alternative costs days: name the propositions, grade how each can be known, test the ones that are expensive and knowable, restructure around the ones that are not, and record what was bet. Organizations that work this way still take risks. They stop taking risks by accident.

Related: Technical Due Diligence Should Test the Investment Thesis, the same discipline applied when the commitment is an acquisition.

← All Field Notes