The Undo Is Manufactured, Not Measured

#ai #llm #tools #productivity

The Undo Is Manufactured, Not Measured

A board is asked to approve entry into one new market. The proposal is a full commitment: a local entity, a three-year lease, six hires, and a signing date at the end of the month. The alternative in the deck is to wait a year. Framed that way, the meeting becomes an argument about courage.

There was a third option, and it was not on the slide. Same market, one district instead of four, a six-month lease with a break at ninety days, two hires plus contractors, and a written agreement to read the numbers on a fixed date with the authority to stop. That option is missing from most strategy decks because it does not look like a decision. It looks like indecision with extra steps.

It is neither. It is the same decision, restructured so that being wrong is survivable, and restructuring it changes which regime the decision belongs to.

Reversibility is read off a decision, and it can also be built into one

Earlier in this series there is a list: hand the reversible, enumerable, high-volume calls to the millisecond layer, and keep the rest for the minutes. The list is right, and it is incomplete in a way that costs money. It treats reversibility as a property you discover in the question. Mostly it is a property you choose in the structure of the commitment.

A pilot is not a smaller decision. It is a decision with a built-in read date and a built-in exit, which converts a wrong answer from a loss into a purchase of information. Phasing is the same move on a longer clock. A break clause, an escrow, a pilot customer, a staggered hire, a lease with an option, a two-stage payment, a design that keeps the old system runnable for a quarter: each of them moves a decision one step back toward the cheap end, because each of them lowers what a wrong branch costs.

Which means the honest first question about a hard decision is not only "which option is stronger." It is "can I make being wrong cheaper than this?" That question is answerable in an afternoon, and it gets skipped because the framing arrives pre-packaged as two doors, one of which is labeled bold.

What an undo costs, and how to price it

Manufacturing a way out is not free, and pretending otherwise is how the discipline turns into an excuse for endless pilots.

It costs money directly: break fees, duplicated tooling, the premium on a shorter lease, the slower start. It costs position, because to a partner a phased commitment can read as a hedge, and sometimes it is one. It costs attention, because a staged decision is another thing to manage and revisit rather than a box that closes. And it can cost the thing you were buying in the first place: a full commitment is sometimes the only credible signal, and a tentative one is worth less to the person receiving it.

So price the exit against the error you are actually carrying. Our published bands, self-run on named benchmarks with the failures disclosed, are the price list. On JudgeBench, 620 judgments with the 6 first-verdict failures disclosed, calls reported at 90% confidence or above were right 99.6% of the time and calls in the 80–90% band 94.0%; raw accuracy came out at 92.5% against 92.2% for a plain direct baseline, a tie we report as a tie. If you act on a call in the 80–90 band, the residual is roughly six in a hundred, and the question becomes arithmetic: is the undo path cheaper than six percent of the downside? Act above 90 and the residual is four in a thousand, where an expensive exit is usually not worth buying. The band tells you how much insurance the decision needs. Without a stated confidence you are buying insurance blind, or, more often, buying none and calling it conviction.

The part that cannot be undone

The discipline has a boundary, and the boundary is a question with a short answer: name the thing that cannot be un-said, un-signed, or un-learned.

Some decisions are one-way doors, and no amount of structure changes that. A signature on somebody else's future. A public commitment made in front of the people it affects. A sold controlling stake. The first entry in a record that others will cite. Trust, once spent, is not re-purchased at the old price. Those decisions belong in the minutes no matter how small the numbers look, and the useful work there is not engineering an exit but slowing down enough to deserve the signature.

The symmetric error is expensive and common: treating a one-way door as though it had a handle. "We can always change it later" is true until the announcement goes out, the team reorganizes around it, or the counterparty makes plans of their own. Reversibility has to be verified against the other party's behavior, not against your own intention. A staged exit nobody has agreed to is not an exit; it is a hope with a calendar attached.

And the third error is the one sorting lists encourage: treating a structured decision as a failure of nerve. A team that buys an exit is often described as not having committed. What it has actually done is move the decision into the regime where a rate, a read date and a machine can carry it, and reserved the minutes for the doors that do not open again.

An exit that is not written down is not an exit

Here is where this connects to the rest of the series. The exit path is a document, or it does not exist.

It has to name four things: who can pull it, by when, at what cost, and on what evidence. A pilot with no predefined read date is not a pilot; it is a full commitment with a delay on the invoice. A review trigger of "if things go badly" is not a trigger; it is a mood. The read date is the load-bearing part, and it has to be set while the decision is still exciting, because by the time it matters, the people who would have to admit the mistake are the ones holding the calendar.

There is a second artifact the staged decision needs, and it is the confidence. A number attached to the pick is what makes the read date a test rather than a conversation. If the call was reported at 94% and the district came in at half the projection, that is a real signal about the instrument or the assumption. If the call was reported at 60%, the correct conclusion on the read date may be that the decision was a coin flip made on purpose, and the exit was the whole point. When both answers are defensible, our own benchmark says so: the deliberately constructed near-tie splits on ContextualJudgeBench, self-run under the official pairwise protocol where a pair counts only if both presentation orders are judged correctly, which is why the random floor is 25% rather than 50%, sit at 46–60%. That band is not a verdict you should strain to read. It is the strongest argument for buying the undo instead.

What we owe the decision you can still get out of

We build the slow end of this, so the bias is ours to declare. The case for minutes is not that deliberation is virtuous. It is that some decisions cannot be restructured, and those are the ones where a person has to sign and live with the result.

For everything else, the better move is usually structural rather than analytical: spend the afternoon making the wrong branch survivable, then hand the read to a rate and a date. Decider is for the part that structure cannot remove, one question and two candidate answers that both survive scrutiny, and it returns a pick with a calibrated confidence and a written argument, so the exit you built has something to be read against.

Buy the undo where you can. Spend the signature where you cannot.