A precommitted kill criterion is a governance instrument that depends on a political condition: the sponsor has to survive triggering it. Where that condition fails, the criterion is written, agreed, and quietly not applied.
The mechanism it has to defeat
Escalation of commitment is the documented mechanism, and its shape is specific. In a role-play experiment manipulating responsibility and financial feedback, the highest group mean allocation of subsequent funding to a previously selected division occurred among participants who had made the initial choice and then received negative results.
The people most likely to fund a failing course are the ones who chose it, after learning it is failing. That is precisely the population holding sponsorship of a prototype.
So the criterion is not fighting inertia or poor information. It is fighting an incentive that grows stronger as the evidence gets worse, and it is doing so through a person whose credibility is attached to the original choice.
Sunk cost has a precise treatment
Appraisal excludes already-incurred, unchangeable costs from the decision about what to do next, while retaining the opportunity cost of continuing to use resources already paid for.
Both halves matter. A sponsor citing 50,000 already spent has offered nothing about the future. But a prepaid resource can still have an alternative use, and a future commitment attached to it can still be relevant — so the answer is not to ignore everything already purchased.
The comparison that decides it is future consequences under each option: operating and transition work, contractual commitments, user reliance, information obligations, salvage, and forgone alternatives. Past spending is preserved for accountability and excluded from the benefit side.
Suppression is investigated, not asserted
Political suppression occurs when organisational incentives prevent adverse evidence or a justified stop recommendation from reaching an accountable decision. It is a candidate mechanism rather than a default explanation.
The investigation compares the original criterion, the full evidence, the reporting versions, and the sequence of decisions, and identifies who selected the material shown to decision makers and what changed between versions.
Four leads worth following: unexplained removal of adverse cases, repeated extensions with no learning question, criteria rewritten after results, and instruction to present an unresolved study as successful. Each is a lead requiring context.
A review deck that drops the failed task slice from the original acceptance criterion, leaving only the successful one, has a change log showing selective presentation. Whether that was an editing error, a change disclosed elsewhere, or deliberate pressure needs further evidence, and the response is to restore the criterion history before deciding whether continuation is justified — not to infer motive from a slide.
The competing explanations are tested with the same seriousness. Continuing can be justified by a credible new test, valuable partial use, an obligation to support existing users, or genuine uncertainty in a flawed measurement. Disagreement about values or costs is different from alteration of evidence.
An analogous pattern is documented in research reporting, where an archive of conducted studies showed an association between result strength and whether work was written up or published, with some authors describing anticipated difficulty or reduced priority after null findings. Different setting, same structural pressure.
Three mechanisms that help
Dissent with access. Reviewers need a genuine challenge role and enough access to inspect the underlying evidence. Nominal independence without access adds little, which is the standard failure of an oversight function.
A route to report a negative result. It has to be possible to report a valid negative finding without fabricating a success story — while preserving accountability for missing controls or poor work. Blameless framing that erases consequential actions and ownership fails in the other direction.
Incentives that do not reward only continuation. Where the only recognised outcomes are shipped and ongoing, the criterion is asking someone to volunteer for the one outcome the system punishes.
The memo carries the credibility
The kill memo is the durable record of why the work ended and how the remaining consequences are handled, and it is what makes an honest kill survivable.
It preserves the original hypothesis, criteria, and amendments; the evidence window, population, outcomes, and limitations; the adverse results, missing evidence, and disagreements. It distinguishes an unsupported hypothesis from an invalid or incomplete test of one — a distinction that protects both the team and the future reader.
It separates measured past expenditure from forecast future differences from realised savings. A four-month comparison forecasting 3,000 a month to continue against 2,000 to close yields a 10,000 forecast difference under those assumptions, and that number is neither realised savings nor a net benefit: it excludes future benefits, other alternatives, and any omitted commitments.
And it ends with closure tasks, owners, deadlines, retained obligations, salvage decisions, and the conditions under which the idea would be revisited. Retirement itself has substance — remaining user needs, replacement arrangements, communication, API transition lead times, and continued protection of retained information.
The rule
What stays fixed is that the decision authority, the criterion changes, and the dissent are all visible in the record. What changes is who holds the incentive to continue, and the criterion is designed against that person rather than against uncertainty.
Not to be confused with
Proof of motive. Escalation studies and publication-bias research illuminate mechanisms. They establish nothing about the motives of anyone in a particular project.
A guarantee. Writing a memo does not preserve credibility, and a review process does not eliminate bias. These are proposed safeguards, not an evaluated programme.