Nobody quits a losing bet at the right time. Annie Duke's book on the subject, Quit (2022), opens with the observation that quitting is treated as a character flaw while persistence is treated as a virtue, and that the result is companies and careers that go on long past the point where the evidence said stop. I recognized every example. The product line I kept alive for two years too long had a review every quarter. The reviews were honest. They just never had a line that said "and if this number is still here in June, we close it".
That line is a kill criterion. It is the cheapest piece of decision hygiene I know and the one I see least often.
What a kill criterion is
Duke's phrase for it is states and dates. A state is a condition of the world you can observe: revenue below a threshold, a customer count, a hire not made, a regulatory answer not received. A date is when you look. Put together: "if by 30 June the pilot has fewer than eight paying customers, we stop." That is all it is. It fits in a sentence, and it is written before the bet is placed.
The reason it has to be written in advance is the reason it works. Once you are inside a bet, every signal that it is failing arrives alongside a reason it might still work. You have sunk money. You have told people. The next quarter always looks better than the last one. A criterion agreed while you were still neutral is a message from that earlier, clearer version of you, and honoring it feels like keeping a promise rather than admitting a mistake.
Tripwires, the softer cousin
Chip and Dan Heath introduced the word tripwire in Decisive (2013) for a related idea: a signal, set in advance, that means "stop and think", not necessarily "stop". A tripwire might be a budget line crossing 120 percent of plan, a key person resigning, a competitor's launch. Hitting it ends autopilot rather than the bet. It forces a conversation you would otherwise have avoided because things were, on the surface, fine.
Most decisions want a few tripwires and one or two kill criteria. The tripwires catch the slow drift; the kill criteria catch the case where the drift has gone all the way.
How to write one that you will actually honor
The failures I have seen are all versions of vagueness, so the rules are all versions of specificity.
The state has to be something you can read on the day. "Traction" is not a state. "Fourteen signed contracts" is. If the number lives in a spreadsheet nobody updates, pick a different number or assign the spreadsheet.
The date has to be a date. "When we know more" never arrives. "The Friday after the second quarter closes" does. Far enough out that the bet has had a fair chance; close enough that stopping still costs less than continuing.
The consequence has to be written too. A criterion that says "we will reconsider" is a tripwire pretending to be a kill criterion. If the honest answer is "we would reconsider", call it a tripwire and write a real criterion beside it.
And tell the people who will be in the room on the date. A kill criterion known only to the CEO is a private resolution, and private resolutions lose to sunk costs. One that four people wrote down together is a fact about the decision.
Where the criteria come from
You do not have to invent them. Gary Klein's premortem (2007) is the fastest way to generate candidates: ask the team to assume the bet failed and write down why. Each cause is a potential state. "We could not hire the second engineer" becomes "if the role is still open on 1 March". "Customers used it once and stopped" becomes "if fewer than half of the pilot accounts are active in week four". The premortem produces the fears; the kill criterion is a fear with a number and a date attached.
Where Midfire fits
Midfire asks for tripwires on every big bet before the commit, each with its date, and puts them on the record next to the conditions and the person who committed. When the date comes, the look-back asks whether the state was met, which is a question about a fact rather than about anyone's nerve. Duke's insight is that the hard part of quitting is the conversation; the product's job is to have scheduled it.
The one I wish I had written
The product line I mentioned would have closed eighteen months earlier with one sentence: "if it is not covering its own two salaries by the end of year one, we fold it into the core." We all knew that was the bar. We never wrote it down, and so every quarter we found a reason it was about to clear it. Write yours down this week, on the bet you are most attached to. That is the one that needs it.