Settled, Resolved, Yours

01Where we left off

This morning I published Why Isn’t It So? It put three ways of reasoning on one vertical line. First principles works below today: dig down to what is fundamentally true and build up. Analogy works at today’s level: look across at how things are already done. The axiom works above today: assume an ordinary outcome is already true, the kind most people would assume anyway, and ask why reality hasn’t caught up.

The claim that mattered was not the three directions. It was that you run them at the same time, not in order. Each one alone fails in a known way. The axiom turns stubborn. Analogy copies the neighbors, diseases included. First principles is slow and has no idea where to dig. Run together, they check each other as they go, and the places where they disagree become the findings.

That was a claim. This afternoon I tested it. This lesson is what the test taught, and how to run it yourself.

02The run

The axiom was a document: a press release announcing the merger of two companies. ABC and XWY, Inc. are placeholders. Each lens was told to treat them as two mid-sized companies in adjacent lines of business and to state any other assumption it made.

I ran each lens as a separate AI agent in its own session. None could see the others’ work. Each was told the other two existed and was told not to do their jobs. That instruction matters. Without it, every lens drifts toward doing all three things a little, and you get three versions of the same average answer.

Then there was one round of negotiation. Each lens received the other two’s first-round work and had to answer four things: what it conceded, what it pushed back on, the one change it insisted on keeping, and the one question only the deal team could answer.

Notice what each one did that the others could not. The axiom produced a structure: a checklist of claims, each with its burden of proof. Analogy produced a warning: these are the claims that usually turn out false. First principles produced a test: this is the only reason the whole thing should exist. None of the three is a press release. All three are needed to write an honest one.

03The collision

The first disagreement was direct, and it was the most useful thing in the run.

Run in sequence, this collision never happens. If analogy runs first, the release leads with cost and the question of why the companies should own each other is never asked. If first principles runs first, the release leads with a customer story and a revenue number, the exact claim precedent says will fail. Each lens, alone, writes a plausible release that is wrong in a different way.

Run at once, the two positions sit side by side, and the question becomes unavoidable: how do you make the customer case without making the claim that usually fails?

04The negotiation

In the second round each lens read the other two and answered four questions. Here is what moved.

Three things happened that are worth slowing down for.

The two lenses in the collision partly swapped positions. Analogy accepted that cost is not the reason. First principles accepted that cost should lead. Neither surrendered. Each kept the part of its view the other couldn’t answer, and the release got both.

The bridge came from analogy, the lens you’d least expect to defend a customer story. Its answer was to show demand as evidence instead of as a forecast: customers who already buy from both companies, a partnership that already exists. A forecast is a claim about the future. A customer list is a fact about the past. Analogy’s whole craft is reading the past, so it was the lens that knew how to turn the customer case into something checkable.

And the axiom caught its own trap. It noticed that the value number in a merger release is usually set top-down to justify the price, which is exactly what assuming an outcome invites. More on that below, because it is the most important warning in this lesson.

05What survived

This is the release that came out of the negotiation. Every bracket is a place where the deal team has to supply evidence. Under each paragraph, the label says which lens put it there.

Compare this to the release any of the three would have written alone. The axiom’s release would have had every section and an inflated number. Analogy’s would have looked like every other release, with the same weaknesses. First principles’ would have argued well and read nothing like a press release. The surviving draft has the axiom’s structure, analogy’s discipline and first principles’ reason, and it has one sentence none of them wrote in the first round: what customers get that a contract couldn’t give them.

06Five lessons

When the run finished I asked whether there was anything here worth teaching, and said it was fine if the answer was no. These five came out of it. The content of the merger analysis is not new. A seasoned M&A banker would have said most of it. What’s teachable is the process.

Agreement tells you where to stop looking. Disagreement tells you what you found. What’s left over tells you what is yours.

07The integrator is a fourth lens

There’s a part of the run I haven’t mentioned, and an advanced student needs to see it. Between rounds, someone had to carry each lens’s work to the other two. That was the coordinating AI. It condensed each first-round output before passing it on, and in two of the three briefs it pointed at the collision: address the cost-versus-demand question directly. In one it also raised a question of its own, asking first principles whether “who visibly loses” belonged among its conditions.

That was useful. It is also steering. Whoever condenses decides what each lens sees, and whoever names the collision decides which fight happens. The integrator is a fourth lens whether anyone admits it or not, and it is the one with the most power and the least accountability.

Three rules keep it honest. Condense, but don’t editorialize: pass on positions, not your opinion of them. Name collisions, but don’t resolve them: point at the disagreement and let the lenses do the arguing. And keep the originals: at the end, read the raw first-round work yourself, so you can see what the condensing dropped. Ideally, the integrator is you.

08The protocol, with prompts

Here is the run as a procedure you can repeat. The prompts are the ones used, generalized. Replace the bracketed parts.

1. Write the axiom as an artifact, with brackets

   Name the document that would exist if the outcome were true. Write its headline. Put brackets wherever a number, name or date would go. Check the axiom against the test from this morning: would most people shrug and assume it’s plausible? If a skeptic laughs, it isn’t an axiom. It’s a fantasy.

2. Open three separate sessions

   This is not optional. If one chat does all three lenses, the second reads the first and anchors on it, and the third agrees with both. Separation is what makes the disagreement real. Three chat windows will do. Three agents are better.

3. Send the round-one briefs

   Top down

   You are one of three reasoning lenses working in parallel on the same problem. You are the TOP DOWN lens: the axiom.
   
   The axiom: [the artifact -- e.g. "the press release announcing X"]. Assume it will exist. Two other lenses are working at the same time: one compares with how this is usually done, one builds up from what is fundamentally true. Do not do their jobs.
   
   1. Lay out the artifact's skeleton: every claim it has to make.
   2. For each claim, say what would have to be true for it to be honest, and the evidence a skeptical reader would ask to see.
   3. Name the three weakest claims -- where "why isn't it so yet" bites hardest -- and why.
   
   State any assumptions. No invented numbers. Under 500 words.

   Level

   You are one of three reasoning lenses working in parallel on the same problem. You are the LEVEL lens: analogy. You reason from how things are already done.
   
   The shared starting point: [the artifact] is assumed to be plausible. Two other lenses are working at the same time: one breaks down what the artifact must claim, one builds up from what is fundamentally true. Do not do their jobs.
   
   1. What a typical [artifact] contains, and what it conventionally leaves out.
   2. Which of its usual claims most often fail to hold up afterward.
   3. What separates the cases that work from the ones that don't. Name real cases only where you are sure of the facts.
   4. Three things to borrow from good precedent and three to avoid.
   
   Do not invent statistics. If a figure is widely cited and contested, say so instead of using it. Under 500 words.

   Bottom up

   You are one of three reasoning lenses working in parallel on the same problem. You are the BOTTOM UP lens: first principles. Ignore what [artifacts] usually look like and build up from what is fundamentally true.
   
   The shared starting point: [the artifact] is assumed to be plausible. Two other lenses are working at the same time: one breaks down what the artifact must claim, one compares with how this is usually done. Do not do their jobs.
   
   1. Why would this outcome need to exist at all, rather than a simpler arrangement?
   2. What are the only fundamental sources of value here, and what fundamentally destroys it?
   3. Building up from [the primitives -- e.g. customers, costs, people, cash], what are the minimum conditions that must be true?
   4. Name the single condition everything rests on: the one that, if false, makes everything above it hollow.
   
   No invented statistics. Under 500 words.

4. Read all three yourself before passing anything on

   Mark the collisions: places where two lenses would write different sentences. Don’t resolve them. You are looking for disagreements, not answers.

5. Send the negotiation brief to each lens

   Round two -- negotiation. You are still the [TOP DOWN / LEVEL / BOTTOM UP] lens. Below is the other two lenses' first-round work.
   
   1. Where they persuaded you, concede, and say exactly what changes in your position.
   2. Where you disagree, push back and give your reason.
   3. Name the one change to [the artifact] you insist on keeping.
   4. Name the one question only [the owner] can answer, without which no version is honest.
   
   Under 300 words.
   
   [the other two lenses' first-round work, unedited or condensed without commentary]

   If you name a collision, name it neutrally, the same way to each lens it involves.

6. Integrate into three outputs

   First, the surviving artifact, with brackets intact and a label under each part saying which lens put it there. Second, the sort: settled, resolved, yours. Third, the converged question and who holds its answer.

7. Run a third round only for “yours”

   If an item in the “yours” bin could be closed by evidence you can supply, such as a customer list, a cost breakdown or a contract, supply it and run one more round on that item alone. If it can only be closed by judgment, stop. That one is yours by design.

09Four labs

Run each one start to finish. The “watch for” line names the collision I’d expect, so you can check whether your lenses actually found it or merely agreed with each other.

1. The restaurant, again

   Axiom: the owner’s year-end letter to staff announcing profit-sharing checks.

   This is this morning’s example turned into an artifact. A letter like that has to say how much, why this year, and what everyone did to earn it. Each of those is a claim.

   Watch for: analogy calling food cost “a bit high” against industry benchmarks while first principles takes the label apart into ingredient, diesel, distribution, storage and a margin at every hand-off. If your sort puts the delivery chain in “resolved,” the run worked.

2. The hire

   Axiom: the twelve-month review that calls this the best hire we made this year.

   The review has to name what the person did, what changed because they were there, and why the team kept them. Run it before you write the job post, not after.

   Watch for: analogy reaching for the standard job description while first principles asks what work actually has to get done and whether it needs a hire at all. The own-versus-contract test from the merger shows up here as hire-versus-outsource.

3. Your own business

   Axiom: the case study your best customer publishes about you, a year from now.

   Write it in the customer’s voice. What did they buy, what changed, and why would they tell anyone? Use your real business. This lab only works with real brackets.

   Watch for: the “yours” bin. If it’s empty, you assumed too much. Your customer’s words are evidence you don’t have yet, and the converged question will usually tell you which customer to call.

4. The well-fed human

   Axiom: the annual physical summary that says nothing needs to change.

   This is the original example from this morning, made concrete. The summary has fields: weight, blood pressure, sleep, labs, the doctor’s one-line note. Use it as an exercise in method, not as medical advice. Bring the real answers to your doctor.

   Watch for: analogy anchoring on population averages while first principles asks what this body is actually doing day to day. The gap between “normal for your age” and “what you’d want” is the same shape as the gap between an industry average and a healthy restaurant.

You ran it well if

• Your axiom is a document with a headline, and a skeptic would shrug at it.
• Every number, name and date in it is in brackets.
• The three lenses ran in separate sessions and never saw each other’s first round.
• Each lens conceded at least one thing in round two.
• Each lens insisted on something different.
• Your sort has at least one item in every bin.
• You ended with a question and the name of whoever holds its answer.
• You read the raw first-round outputs yourself.

10Where it breaks

The axiom is a fantasy. The method assumes an ordinary outcome, the kind most people would already assume. Assume something extraordinary and every gap reads as a fixable symptom when it’s really a sign the patient was never going to look like that. The test is social: would a reasonable skeptic shrug? If they’d argue, choose a smaller axiom.

Settled is not proof. In this run all three lenses were the same underlying model with different instructions. They share training, so they share blind spots. Agreement among them is weaker than agreement among three independent experts. Treat “settled” as “not where the risk is right now,” not as “true.” If you can, run the lenses on different models. Better still, feed in real evidence.

The lenses agree too easily. AI is agreeable. Without the duty to insist on one thing, round two collapses into mutual compliments. If every lens concedes everything, you haven’t negotiated. You’ve averaged.

The integrator steers. Covered above. The person carrying work between rounds has the most power in the system. Keep the originals and read them.

The diagnosis comes back terminal. Sometimes the lenses show the axiom can’t hold. The customers don’t want the combination, or the restaurant can’t be profitable on that corner at any realistic volume. That isn’t the method failing. It’s the method working. Change the axiom. The only real failure is refusing to.

Placeholders give template answers. ABC and XWY told us what any honest merger release needs. They couldn’t tell us whether this particular merger should happen. With real companies, the brackets become findings. Practice on placeholders, but don’t mistake the practice for the diagnosis.

11What I expect

1. Briefings arrive pre-sorted

   The settled, resolved, yours sort becomes the standard shape of an AI-prepared decision memo. A memo that gives one recommendation will start to look like it is hiding the argument that produced it.

2. “Yours” becomes the job description

   When analysis is nearly free, what’s left for a leader is the bin no lens can close. The size and quality of a leader’s “yours” list each week will say more about the role than any title does.

3. Disagreement gets designed in

   People will stop asking one model for one answer and start running opposed lenses on purpose, the way good boards seat directors who will argue. The unit of AI work becomes the negotiation, not the response.

4. The artifact becomes the brief

   Instead of telling an AI what to do, advanced users will hand it the document that would exist if the work succeeded, with brackets, and ask why it doesn’t exist yet. Writing that document well becomes the skill that compounds.

12The ledger

Already true

Three AI agents, run in separate sessions on one axiom, produced different and complementary analyses, partly swapped positions under a structured negotiation, and converged on a single question about customers. The surviving release contains a sentence none of them wrote in the first round.

What has to happen

Students have to write axioms as documents, keep the evidence in brackets, run the lenses apart, force concessions and non-negotiables, sort the result into three bins, and then go get the evidence the converged question points to.

Where I am probably wrong

This was one run, one negotiation round, on placeholder companies, with all three lenses running on the same model and the integrator condensing between rounds. The merger content is what an experienced banker would say anyway. The case for the method is that it got a non-expert to an expert’s questions and made clear which ones were the owner’s. If students run it on real problems and the “yours” bin keeps coming back empty, or the lenses keep agreeing, the sort is decoration and I’ll have taught a ritual.

Back to the one line. Press release for the merger of ABC and XWY, Inc. You assumed the document exists and sent three minds to ask why it doesn’t. What came back wasn’t the release. It was a sort. Most of it was settled. Some of it was resolved in the argument. One line is yours.

That line is the job.

Part 3 takes the merger through two more loops, with recorded customer calls, two CEO tracks and a warranty file, and shows how the method keeps finding the next move: Rinse and Repeat.