The Stop Rule: Decide What Enough Means Before You Begin

Research expands to fill the uncertainty available. One more review leads to another comparison. One more AI draft reveals another possible revision. A purchase becomes a spreadsheet, a weekend decision becomes a forecast project and a useful investigation becomes a way to postpone exposure to the result.

A stop rule is a condition written before the work begins that says when to decide, ship, pause, escalate or abandon. It does not force premature certainty. It makes the cost of further inquiry visible and prevents the finish line from moving every time discomfort appears.

Why “I will know when I have enough” fails

Enough is difficult to recognize from inside an active search. New information creates new branches. The last source makes unfamiliar gaps more salient. The cost already invested makes stopping feel like waste. If the decision carries identity or status, continued analysis can protect the fantasy that a perfect answer still exists.

Digital systems make the supply effectively endless. Search engines, recommendation feeds and generative tools can always produce another candidate. The constraint has moved from access to termination. Without a finish condition, apparent thoroughness may be only unpriced attention.

Herbert Simon's work on bounded rationality rejected the fantasy of a chooser with complete information and unlimited computation. A practical decision uses an aspiration level: a condition good enough for the purpose under real limits. Satisficing is not choosing badly on purpose. It is designing a search that can end.

There is no universal stopping rule

A low-cost restaurant choice and a medical treatment decision should not share a threshold. A reversible choice can stop after a small representative sample. A safety-critical or legally consequential decision may require professional review, formal evidence standards or a procedure defined by regulation.

This guide applies to ordinary research, shopping, creative work, low-risk experiments and planning. It does not replace clinical trial stopping boundaries, emergency procedures, legal deadlines, safety inspections or professional standards. In formal research, repeated unplanned looks at results can distort statistical inference; stopping rules need appropriate methodology, not a productivity slogan.

Use the reversibility test first. Increase the evidence and review burden when harm is severe, difficult to detect, hard to repair or imposed on other people.

Step 1: write the decision in one sentence

“Research laptops” has no end. “Choose one laptop under the approved budget that runs these three applications, connects to this display and can be returned within the trial period” can terminate.

Name the actor, choice, deadline and consequence. If the work is exploratory, name the artifact: “produce a two-page brief that identifies the three plausible explanations, the strongest evidence for each and the next discriminating test.” Exploration can be open-ended in spirit while bounded in output.

Separate the decision from the desire to feel certain. The stop rule exists to support an action, not eliminate every possibility of regret.

Step 2: define minimum sufficient evidence

Write what must be true before a decision is allowed. For a purchase, this might include exact compatibility, total cost, return conditions and two independent accounts of a known tradeoff. For a factual brief, it might require one primary source for each material claim, a documented disagreement and a check for newer guidance.

Use source classes, not a raw source count. Ten pages repeating one press release are one evidence path. A manufacturer's manual can establish compatibility but not independent durability. A regulator can define a rule but may not answer a personal value judgment.

Define disqualifiers separately. If a product lacks the required port, a plan exceeds the hard budget or a source cannot support the central claim, stop considering that path. Disqualifiers save more time than elaborate rankings of unacceptable options.

Step 3: set the search budget

Choose a clock limit, source limit, iteration limit or spending limit appropriate to the decision. Examples: forty-five minutes and five strong sources; three vendor quotes; two AI drafting rounds followed by human editing; one weekend prototype; a fixed amount of money that can be lost without creating a second problem.

A budget is not the evidence threshold. Both conditions matter. If the threshold is met early, stop. If time expires before it is met, do not quietly lower the standard. Escalate, defer, narrow the question or record that the evidence is insufficient.

Time-boxing alone can reward a rushed answer. Evidence criteria alone can produce endless search. The stop rule needs the pair.

Step 4: define five exit branches

1. Decide. The minimum evidence is met and no disqualifier remains.
2. Stop early for dominance. One option meets every requirement and further search is unlikely to change the choice.
3. Stop early for futility. No available option can meet a hard requirement within the budget.
4. Escalate. Stakes, uncertainty or authority exceed the decision-maker's competence.
5. Pause and review later. A named future event—new test result, contract date, policy update or price-independent need—could materially change the answer.

“Keep looking” is not a branch unless it states what missing evidence is being sought and how much more search it earns.

Step 5: add guardrails that outrank progress

A guardrail is a condition that stops the activity even if the primary goal is improving. A personal experiment stops if sleep or safety deteriorates. An AI workflow stops if it exposes private data or loses source traceability. A home project stops before work enters electrical, structural or permit territory beyond the plan.

Guardrails should be observable. “Stop if this feels unhealthy” is weaker than “stop after two nights below the minimum sleep window” when that threshold is appropriate and low risk. High-stakes guardrails should come from qualified guidance, not improvised numbers.

Put the authority beside the trigger: who may stop the work, who must be notified and what safe state follows? A stop signal without an action path becomes a warning everyone can admire and ignore.

Step 6: allow changes—but preserve them

Precommitment should not become stubbornness. New facts can expose a bad rule. If the budget changes, a safety issue appears or the decision scope expands, revise the rule. Record what changed and why before continuing.

This is the practical lesson of research preregistration. The Center for Open Science describes preregistration as specifying a research plan in advance. It does not ban exploration. It makes the distinction between planned and later decisions visible. Ordinary decisions benefit from the same honesty.

Do not rewrite the original condition after seeing the result and pretend it was always the plan. Keep the first version, the revision and the evidence that justified it.

Use a stop rule for AI iteration

Generative systems remove the natural friction of producing another draft. That can improve work or create an infinite surface for polishing. Define the acceptance test outside the model: factual support, required sections, tone constraints, prohibited content and the human who owns the final decision.

Limit self-critique loops. A model can always suggest another weakness, and its critique may be as unsupported as its draft. After the defined rounds, verify consequential claims against primary sources and move to accountable human judgment. For a broader workflow, use the AI workflow audit.

The one-card stop-rule template

• Decision or artifact: What exactly will exist at the end?
• Minimum evidence: Which facts, tests or approvals must be present?
• Disqualifiers: What makes an option unacceptable immediately?
• Budget: How much time, money, searching or iteration is authorized?
• Early success: What makes further search unlikely to change the choice?
• Futility: What shows the path cannot meet the need?
• Guardrails: Which conditions stop the work regardless of progress?
• Escalation: Which uncertainty requires a specialist or another owner?
• Review trigger: What future event justifies reopening the decision?

A worked ordinary example

Decision: choose a home router for a two-story house by Friday. Minimum evidence: current ISP compatibility, enough Ethernet ports, published security-support policy, return window and a placement plan. Budget: ninety minutes, official documentation plus two independent technical sources, three candidates. Disqualifier: subscription required for basic security updates. Escalation: existing wiring or business-network requirements make the setup unclear.

The rule may end with “buy nothing.” If moving the current router and testing the dead zone resolves the problem, more shopping is not diligence. It is a new activity requiring its own justification. The home Wi-Fi guide provides the architecture decision after the stopping discipline is in place.

The bottom line

A stop rule does not prove that a decision is correct. It proves that the decision process had a visible boundary. That makes errors easier to audit, revisions easier to explain and attention harder to consume without limit.

Define enough before the search teaches anxiety how to move the finish line. When the rule is met, decide. When it is not, escalate or narrow the problem. Do not call endlessness rigor.

Sources and scope notes

• Herbert A. Simon, A Behavioral Model of Rational Choice (1955), for bounded choice under limited information and computation.
• Center for Open Science, Preregistration, for documenting plans before results are known.
• NIST, Sequential Testing to Guarantee the Necessary Sample Size in Clinical Trials, as an example of why formal statistical stopping requires dedicated methods.
• For high-impact choices, start with The Reversibility Test and use qualified professional standards where they apply.
• For preserving the rationale and later outcome, use the decision-journal protocol.