Precommitted kill criteria are recorded conditions for ending an initiative or a defined scope, agreed before anyone reviews the evidence those conditions will govern. They distinguish termination from three other outcomes that get confused with it: pausing, narrowing the scope, and commissioning a further bounded test.
The timing is the entire mechanism. Criteria written after the results cannot be departed from, because there is nothing to depart from, and the departure is the thing worth seeing.
A stop is not a finding
A stop decision and a conclusion about value require separate explanations, and merging them corrupts both.
Consider a prototype that meets its task-time target while no team accepts operating responsibility by the agreed date. The correct decision is to stop progression under the ownership criterion. The correct record retains the task-time result within its tested scope and states that progression stopped for want of an owner. Rewriting that as “no user value” destroys a valid finding and misinforms whoever revisits the question later. The reverse error is equally common: a project stopped on budget grounds gets reported as evidence the approach does not work.
Termination is therefore not proof of no value, and continuation is not proof of value. Each needs its own sentence.
Six trigger families
Each trigger needs an evidence source, a reviewer, a specified response, an authority, and a review time. The families differ in what they observe and in what response they permit.
Insufficient value. Evidence and remaining uncertainty inconsistent with a worthwhile outcome under the tested conditions. Response: end the scope or reject the intervention.
Inconclusive evidence. Missing or inadequate observations, or uncertainty that cannot answer the decision question. Response: close the test without a value conclusion, and decide explicitly whether another test is justified. This is the family that goes missing, and its absence forces inconclusive results into the first family.
Resource limit. The available time or money, and who can change that allocation. Response: stop at the limit, or record a separately justified extension.
Feasibility barrier. A required capability and evidence it cannot be delivered within the accepted scope. Response: stop, narrow, or investigate an alternative.
Operational harm. An observable event requiring interruption, and who can act immediately without waiting for a meeting. Response: pause, contain, review restart conditions.
Missing ownership. The operating or receiving responsibility required, and the date by which it must be accepted. Response: hold the next commitment, or end the work if the dependency stays unresolved.
The last is the one that distinguishes prototype governance from evaluation governance. Every other family asks whether the thing works. This one asks whether anyone will run it, and it is answerable long before the evidence arrives.
Staged investment, and what an option actually is
A real option is an opportunity, not an obligation, to make a later investment or operating choice in response to what becomes known. Staged investment separates commitments so that later choice stays available. Application-design prototyping has been analysed in exactly these terms since the late 1990s.
Before describing an initial spend as buying an option, five things need answers. What later action does it preserve — deferring, expanding, changing, or abandoning? Who can choose it, and what capability, resources, or agreements make it available to them? Does the spend create new access, reveal decision-relevant information, or build reusable capability, as distinct from simply being spent? What becomes known by testing rather than by waiting? And when does the opportunity erode — a supplier offer expiring, a team dispersing, a price rising?
Valuation is a separate step and a harder one. The known limitations are that project value and variance are difficult to estimate, that the price-process assumptions imported from financial options are frequently unsuitable, and that the duration of the right is itself uncertain. Identifying the flexibility is useful on its own. Attaching a number to it requires defending all three.
The rule
What stays fixed is that every trigger names its evidence, its owner, its response, and its moment. What changes is the values, and those cannot be copied between projects, because the threshold is a function of what the change costs the organisation that is paying for it.
Where precommitment fails
The criterion has a number and no owner. “Stop if accuracy is below 85 percent” specifies nothing about who looks, when, or what happens next. In practice the review does not occur, and the number is retrieved afterwards to justify whatever was decided.
The stop is treated as a verdict on the idea. Covered above, and worth its own trigger family precisely because it recurs. The closure record has to separate what the evidence established from why work ended.
Not to be confused with
A quota. No optimal proportion of projects to kill is established by anything. A programme that kills a fixed share is running a quota, not a criterion.
Evidence that written triggers work. Research on scientific decision-making in entrepreneurial settings found higher termination rates in trained groups than in controls. That result concerns a training intervention, not written triggers isolated from it, and it does not establish that recording criteria alone changes behaviour.
Irreversibility. Precommitment makes exceptions inspectable. It does not make the first judgement final, and a restart is legitimate where the new ownership decision is made and the earlier evidence is rechecked for whether it still applies.