The order of the questions decides what the analysis is worth

Last updated 5 September 2026

The AGI²⁺ pre-commitment review works through a fixed order: the frame first, then the decisive uncertainty, then one question at a time, then the evidence, then the decision structure. Order matters because each answer changes how the next question is read. Fix the order and the analysis lands on the right problem.

Start with the frame

The first question a leadership team brings is rarely the one that decides the outcome. It became sayable, survived three meetings and got written into the agenda, carrying assumptions that stay invisible because everyone in the room shares them.

You are about to choose which of three countries to enter first, a reasonable question and expensive to answer well. But every document you hold, from the memo to the volume model, assumes you will ship product and a local company will sell it, and nobody wrote that down. If the opening is really a partnership structure rather than an export chain, the ranking answers a question you are not facing.

So the review declines to answer it, and builds a frame first: the real decision, the objectives actually in play, what sits outside the boundary, and the date by which this must be settled. That frame returns as a strawman, filled in and wrong in places, made to be corrected, and a gate holds the session there until every load-bearing actor, mechanism and authority condition is present.

When to stop

The complaint executives actually voice is not about question quality: it is that they have been analysing something for six weeks and still cannot decide, or that they moved quickly and missed something. Both are failures of timing, and every structured method that died on time cost died here.

Readiness is the answer, and it runs as a live test rather than a stage you arrive at. After every material change to the frame or to a belief, the review asks whether one more question is worth what it costs and can still be resolved before the deadline. When the answer turns, the analysis stops, with plenty still unknown and only one thing established: those unknowns will not change what you do.

There is a second stopping condition that matters more in practice, because sometimes the crux is genuinely unresolvable inside your timetable: the counterparty will not disclose terms before you commit, or the behaviour you need to observe requires a market you have not entered.

The review then stops trying to close the uncertainty and asks which decision structure survives it staying open. Stage the commitment so the first tranche is small enough to lose. Buy the information with a pilot instead of a study. Write the exit trigger now, while nobody is yet defending the decision. Turn the one-way door into a two-way door, and the unresolved crux stops blocking you.

Most teams do not make that move unprompted, because the meeting was organised around resolving the uncertainty rather than designing past it. The same discipline retires work already underway: an open workstream that changes nothing whether its result is positive or negative has no decision value, and the review says so even when it is funded and half finished.

Reframing late

A late reframe is a path event rather than a footnote. The review marks every downstream conclusion resting on the old frame as stale and re-runs that path, so you never end up with a corrected problem statement sitting on analysis that still answers the old one.

After the decision

Signposts already do part of this work. Each one names an observable indicator, its source, the threshold and the action when the threshold is crossed, and it is written before the commitment, while nobody in the room is yet defending anything. The v26 architecture carries them forward with the commitments they belong to: what was agreed, who owns it, when it falls due, what should be observable by then, what is actually observed, what is overdue, and the trigger that forces reconsideration. Not a task tracker with different labels. A task tracker asks whether the work was done. This asks whether the belief that made the work worth doing still stands.

Then the checks. Verification sets what was expected against what is observed. Falsification looks for the observation that would show a live hypothesis to be wrong, rather than for more support. A mechanism-drift check asks whether the causal path the decision relied on still operates as it did, a different question from whether new information has arrived. An adaptation check asks whether the actors have changed their behaviour under pressure. A failed forecast has at least four readings: the hypothesis was wrong, the mechanism held anyway, an external counterforce pushed the other way, or an actor adapted to the decision. Treating the first as the only one is the common error.

Learning is where the loop on this page ends: what did the episode teach. Re-anchoring is the step after it, and it asks what is now the authoritative state the next cycle begins from. Architecture rather than shipped software today, and set out station by station on the Runtime page.

Whether it beat the obvious alternative

Every buyer asks the same question, and it is the right one: could a capable analyst with one strong model not do this? Sometimes, yes. So the comparison is written into the engagement on a sample: the same brief goes to a person with one strong model and simple source-and-counterargument discipline, and the two outputs are set against each other.

The report has to state which found more, including in the runs where the comparison wins. No competitor in an adjacent category reports that, for the obvious reason that the result can go against the vendor, and it is reported anyway because the alternative is asking you to take the claim on trust. The comparison has not yet been run on a completed engagement, and that is worth saying plainly. The method has been demonstrated on real decisions; the system running it is being built through the engagements, which is why scope stays at one decision.