Rinse and Repeat

Contents

1. The plan that starts with a deck
2. The loop
3. Loop 1: nobody asked it to call the customers
4. Getting the evidence: the calls, recorded
5. Loop 2: two CEOs, many candidates, many rounds
6. Loop 3: the evidence that broke the conclusion
7. What the loop is really doing
8. The answer may be “don’t”
9. Where the process failed
10. Running the loop on your project
11. Guardrails
12. What I expect
13. The ledger

01The plan that starts with a deck

Big, complicated projects seem to call for big, complicated first moves. So we plan forward from today, and planning forward is reasoning by analogy: we copy the workstreams from the last project that looked like this one. The result is a plan made of activities. Every activity is reasonable. Almost none of them is evidence.

That’s the trap. Activity feels like progress, and it’s easy to assign. Evidence is uncomfortable to get, because it usually lives outside the building with someone who might say no. So a project can run for months on activity before anyone collects the one fact that would have changed its direction.

The previous two lessons built a different way in. Why Isn’t It So? added a third direction of reasoning: assume the outcome is true and ask why it isn’t. Settled, Resolved, Yours ran all three directions at once on a merger press release and sorted what came out. This lesson is about what happens next. You don’t plan. You loop.

02The loop

Figure 01

One loop, run until the question is small enough to answer

1. Write the documentThe artifact that would exist if the outcome were true, with brackets for every fact you don’t have.
2. Run three lensesTop down, level and bottom up, in separate sessions, working at the same time.
3. Negotiate and editEach lens votes on the others’ edits. Anything with two of three votes goes in.
4. SortSettled, resolved, yours.
5. Find the next evidenceAsk each lens for the cheapest piece of evidence that would move the most, and who holds it.
6. Go get itUsually a conversation or a document. Capture it whole: a recording, a file, not a summary.
7. Feed it backThe new evidence goes to all three lenses, along with the current version of the document.
8. Run againUntil the next evidence costs more than the decision is worth, or what remains is judgment.

↺ step 8 returns to step 2

The important word in step 5 is find. You don’t choose the next action. You ask three lenses, working independently, what single piece of evidence would move the decision the most for the least cost. When they converge, you have it. When they don’t, the disagreement tells you something too.

The method never tells you what to do. It tells you what to find out. Doing comes after.

03Loop 1: nobody asked it to call the customers

In Part 2 the axiom was a press release announcing the merger of two placeholder companies, ABC and XWY, Inc. Three lenses worked on it, negotiated once, and sorted the result.

Here is the part I want advanced students to notice, because it’s the most important thing that happened. No lens was asked what to do. The briefs asked the top-down lens for claims and their burdens of proof, the level lens for precedent, and the bottom-up lens for the minimum conditions that must be true. The only forward-looking question came at the end of the negotiation: name the one question only the deal team can answer. Customers did appear in the briefs, as one item on the bottom-up lens’s list of things to build from. Nobody pointed at them as the answer.

All three answered with a question about customers. Do they already buy from both companies? Will they stay? Will they want the combination? The answer to that question sits with customers. So the first action of a very complicated project turned out to be the simplest thing in it: the two CEOs pick up the phone.

That action wasn’t in a plan. It was discovered, because three independent directions all pointed at the same missing fact. When everything depends on one unknown, that unknown is load-bearing. A planning meeting would have listed customer outreach as one item among forty. The convergence put it first.

04Getting the evidence: the calls, recorded

Once the method finds the action, how you carry it out decides whether the evidence survives the trip back. Three choices matter.

The CEOs make the calls themselves

Not sales, not consultants. A top customer takes the CEO’s call. The decision in the “yours” bin belongs to the CEOs, so the evidence should reach them firsthand. Each CEO calls their own accounts, because neither can vouch for the other’s relationships. A delegated call comes back as someone’s summary.

Ask about the need before you talk about the deal

Before an announcement, what can be said is a question for counsel. A safe structure asks about the need, not the deal. How do you buy equipment and service today? Who decides each? What breaks between your two vendors? If one company did both, what would you stop doing? What would make you move your business? Each question maps to a lens. The axiom asks what would be different. Analogy asks what happened last time. First principles asks why they buy separately at all.

Record the call, with consent, and feed the transcript back whole

I’d record every one of these calls on a Plaud. Tell the customer you’re recording and why: you want their words exactly, not your memory of them. Plaud plays a disclosure beep at the start of a recording, and its own guidance is to inform everyone on the call and get their consent. The recorder produces a speaker-labeled transcript. Through Plaud’s MCP connector, an assistant like Claude can read the full transcripts directly, with no copying and pasting.

Feed the lenses the transcript, not the summary. Part 2 showed why: whoever condenses the evidence is steering it. A CEO’s memory of a call is a summary too, and it’s written by hope, which is exactly the bias the axiom invites. The transcript is the evidence. A summary is just a fourth lens.

Figure 02

What the eight simulated calls actually said

• Only three of eight would give a quoteThe rest declined for policy reasons: legal, a state university, a health system, a franchisor. Real customers mostly won’t endorse a vendor in public.
• The largest dealer was already hedging“I’m not going anywhere tomorrow. But this week I’m calling the other two lines I carry to ask what their dealer program looks like.” He wanted written terms before any announcement.
• One sentence held the whole dealA contract cafeteria operator described what happens when an ABC unit under warranty goes down: “we lose two days arguing whose problem it is.”
• Customers raised leadership without being askedOne of ABC’s own customers, on the call with ABC’s CEO, said: “If Sam’s still on the service side, that calms a lot of people down.” Another opened the call with XWY’s CEO by asking: “Who’s in charge?”

05Loop 2: two CEOs, many candidates, many rounds

The second loop changed the setup in three ways. The customer transcripts became the evidence. Instead of one lens deciding “the one concrete thing the combination makes possible,” each lens drew its own candidates from the interviews, and they competed. And we didn’t choose a CEO. We ran the release twice in parallel, once with ABC’s CEO, Dana, running the combined company and once with XWY’s CEO, Sam, with three lenses on each track. Each track ran five rounds of voting, and a separate editor agent applied only the edits that got two of three votes.

Figure 03

Each CEO choice protects the constituency that CEO doesn’t own

Dana as CEO

The manufacturer on top, so the release protects service customers

• Customers keep choosing who services their ABC equipment
• No recommending replacement of a competitor’s unit when it can be repaired
• Technicians not paid on ABC equipment sales
• ABC warranty calls get no priority over other calls
• Sam named in the subhead as leading service

Sam as CEO

The service side on top, so the release protects dealers

• Dated written dealer terms before the announcement
• The terms say where a service call on ABC equipment goes
• Dana’s quote is about dealers, with the same date gate
• Dispatch by response-time commitment, not by brand
• Dana named in the subhead as leading products and dealers

06Loop 3: the evidence that broke the conclusion

A separate agent wrote ABC’s warranty file without knowing what the lenses believed. It contained a fact that undid loop 2’s central claim. In 2024, ABC had given its largest dealer standing warranty approval by contract for nine months. The wait for approval fell from 1.9 days to 0.2. Downtime barely moved, from 6.5 days to 5.0, because the real wait was for parts. Cost per claim rose, though perhaps half of that rise came from an unrelated hinge defect. Finance didn’t renew the pilot. XWY asked for the same arrangement in 2025 and never got an answer.

In the first step, all six lenses on both tracks independently declared iteration 2’s reason false. “A commercial agreement could not provide this” was wrong. A contract had provided it.

Figure 04

The same sentence, three loops apart

Loop 1 · the template

After closing, customers will be able to [the one concrete thing], which a commercial agreement between the two companies could not provide because [reason].

Loop 2 · after the customer calls

A commercial agreement could not provide this, because under a contract the manufacturer paying for warranty repairs keeps the right to check each claim … and today that check is where the delay happens.

Loop 3 · after the warranty file

ABC has tried standing approval under a commercial agreement; it shortened the wait but was not renewed because of cost. … After closing, one company carries both the warranty cost and the service bill.

Look at what the loop didn’t do. It didn’t defend its earlier answer. Neither track argued for the old claim. Both narrowed it to something the evidence would support: under a contract, the company that pays for speed isn’t the company that benefits from it, and that split is what ended the pilot.

Three other things happened in loop 3 that are worth teaching.

A dead candidate came back. Parts on the trucks had been killed in loop 2 because a contract could do it. The warranty file showed parts were the biggest source of downtime. Both tracks put parts back, not as the reason to merge but as a commitment. Evidence can revive what argument buried.

The CEO question moved again. The friction turned out to sit on ABC’s own warranty desk, under Dana: the lapsed pilot, the unanswered request, labor times not updated since 2021. Track A required Dana’s quote to acknowledge that ABC “did not act on standing authorization before this agreement.” The file also showed that XWY was ABC’s slowest warranty servicer, so Track X refused to let the release imply XWY was the better one.

A new fight came up and was settled on the evidence. Two lenses moved ABC’s warranty programs under Sam. The third argued that the desk that approves the largest dealer’s claims can’t report to that dealer’s service competitor. Both conceded, and the vote went 3 to 0. The warranty desk stays neutral.

07What the loop is really doing

Step back from the merger and look at the shape. Each loop didn’t produce a bigger answer. It produced a smaller question.

Figure 05

The question, loop by loop

1. Before loop 1Should ABC and XWY merge?
2. After loop 1What would customers get that a contract couldn’t give them?
3. After loop 2Does the warranty seam actually need ownership?
4. After loop 3Did standing approval itself raise warranty cost, or was it the hinge defect?

That’s the real work of the loop. A big question can’t be answered, so people answer it by judgment, which in practice means by analogy and by hope. A small question can be answered with evidence. The loop keeps shrinking the question until the answer is a number someone already has.

That’s also why each discovered action was cheaper than the last. Loop 1 needed a dozen phone calls and a week. Loop 2 needed a document and a day. Loop 3 needs a spreadsheet and an afternoon. Cheaper evidence isn’t a coincidence. It’s what a question looks like once it has been shrunk.

Figure 06

Three loops at a glance

Loop	Evidence fed in	What the lenses discovered	Next action they converged on
1	None. Just the bracketed release.	Every claim rests on customers. Only owning, not contracting, needs a reason.	Have both CEOs call their top customers.
2	Eight recorded customer calls	The concrete thing is the warranty seam. The CEO choice decides which promises are needed. Both tracks build a split mandate.	Pull ABC’s warranty terms and claim records.
3	ABC’s warranty file	A contract already did it once. The claim narrows to “tried and not sustained.” Parts come back. The deal leans toward a contract.	Re-cut the 2024 pilot’s cost without the hinge claims.

08The answer may be “don’t”

At the end of loop 3, all six lenses, on two tracks that never saw each other, gave the same deal verdict: undecided, leaning toward a contract for the customer benefit the release leads with. Every lever a customer would feel, from faster approval to parts on the truck to updated labor times, is something a contract or ABC alone could do. One lens put it plainly: “The release is now honest, but it does not make the case for a merger on customer grounds.”

That isn’t the method failing. It’s the method doing its most valuable work. Part 1 warned that the axiom can be false and that the diagnosis must be allowed to come back terminal. Here the terminal diagnosis is close, and it comes with instructions. If the re-cut shows that fast approval was cheap, the honest move is to renew the pilot and update the labor times. That’s a contract, not a merger.

Count what it took to get there: a dozen phone calls, one document pull, and soon one spreadsheet. Compare that with what it costs to find out the same thing a year after closing, with integration under way, the largest dealer gone to a rival line, and the customers rebidding their service.

The loop may still come back with a yes. If fast approval itself raised costs, the split between who pays and who benefits is structural, and the narrowed reason to merge holds. Either way, the decision now turns on a number, not on a story.

09Where the process failed

An advanced student should see the machinery’s mistakes as clearly as its findings. Three failures came up, and each one is a design lesson.

• Split votes can keep a falsehood in placeIn loop 3 all three lenses on one track agreed the old “could not” paragraph was false. But each backed a different replacement, none reached two votes, and the false paragraph survived a full round. Agreeing that something is wrong isn’t the same as agreeing on what replaces it.
• Generosity can do it tooIn the final round, all three lenses wanted the unmeasured phrase “first visit” removed. Each withdrew its own fix to back someone else’s. Two fixes tied at two votes each, the tie rule applied neither, and “first visit” is still in that track’s final release even though every lens voted against it. It was left there rather than fixed by hand, because an integrator who fixes things quietly is steering.
• The tracks weren’t equally carefulOne track bracketed “approved by both boards” and put a date on the dealer-terms claim, because on the call the CEO had told the dealer “we haven’t worked that out.” The other track listed both as preconditions but left the text ungated.

The fix for the first two is a rule, not more argument. When all three lenses vote to change the same sentence, the editor applies the version with the most support and breaks a tie by taking the earliest proposal. Watch for a related thing: by the last round, one lens had worked the problem out for itself and backed only one text “so the votes would not tie.” The participants were learning the rules as they went.

There’s a fourth thing to keep in view: the coordinating AI that carried work between rounds. Twice it passed on a neutral process note, that support had split. It never chose a text. That’s the line from Part 2: name the collision, don’t resolve it.

10Running the loop on your project

Start with the Part 2 protocol, including its round-one briefs. These are the additions that make it a loop.

1. Keep one evidence folder

   Every transcript, file and number goes into one place every lens can read, unedited. Summaries are allowed only as extra files, never in place of the originals.

2. Make every round a vote

   Each lens reads the other two lenses’ last round unedited, votes on each proposal (accept, reject, or amend with exact text), and proposes at most four new edits. A separate editor applies anything with two of three votes and adds nothing of its own.

   Round [n]. You are still the [TOP DOWN / LEVEL / BOTTOM UP] lens.
   Read: the current version of [the artifact], the editor's tally, the other two lenses' last-round work (unedited), and the evidence folder.
   1. Vote on every proposal from the other two lenses: ACCEPT / REJECT / AMEND -> "exact text", with a one-line reason.
   2. Rank the live candidates for the concrete thing.
   3. Propose up to 4 edits, quoting the exact old text and the exact new text. Only claims the evidence supports.
   4. NEXT EVIDENCE: the single cheapest piece of evidence that would move the most right now, and who holds it.
   Under 500 words.

3. Give the editor a tie-break

   Two of three applies an edit. If several versions of the same change compete, apply the one with the most support. If they tie, take the earliest. If all three lenses want a sentence changed, it changes.

4. End each loop with verdicts

   In the last round, no new proposals. Each lens gives a final sort, a verdict on the decision itself, a verdict on the hardest judgment call (here, the CEO), and the next evidence. Read where they agree before you read where they differ.

5. Go get the evidence, then loop

   If all three converge on one piece of evidence, that’s your action for the week. If they split, take the cheapest one first. Feed whatever you collect back in full.

6. Stop when the next evidence costs more than the decision is worth

   Or when everything left in the “yours” bin is judgment rather than fact. At that point, the loop has done its work, and the decision is the owner’s.

11Guardrails

Recording consent. Federal law in the United States allows recording a call with one party’s consent, but about eleven states require the consent of everyone on the call, among them California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania and Washington. On calls that cross state lines, assume the stricter rule applies. The simple practice is the one Plaud recommends: tell everyone and get their consent before you record. None of this is legal advice.

Confidentiality. Before an announcement, the existence of a deal is confidential, and for a public company it’s the kind of information securities law cares about. Counsel decides what can be said, to whom, and under what agreement. Asking about the need instead of the deal is a design choice, not a loophole.

Customer data. Transcripts hold your customers’ words and business details. Decide who can read them and where they’re stored before the first call, not after.

Listen yourself. The AI reads every transcript. You should still read the ones that matter. If the lenses quote a line, go find it in the original.

12What I expect

1. Kickoff decks give way to the bracketed document

   Big projects will start from the artifact that would exist if they succeeded. The first status report will be a count of the brackets still open.

2. Recorded conversations become the main input to planning

   When an AI can read every word of every customer call, a leader’s conversations stop being color and start being data. The calendar shifts toward the calls only the leader can make.

3. “Next evidence” replaces “next steps”

   Meetings will end with the cheapest fact that would move the most, and who holds it, instead of a list of tasks. Fewer things will get done, and more of them will matter.

4. Projects die earlier and more cheaply

   The diagnosis that a project shouldn’t happen will arrive in the second week, for the price of a few phone calls, instead of the second year. Teams will learn to treat that as a win.

13The ledger

Already true

Over three loops on two parallel tracks, six AI lenses independently converged on the same next action each time: call the customers, pull the warranty file, re-cut one number. Each action was cheaper than the last. When new evidence contradicted the earlier conclusion, both tracks narrowed their claim instead of defending it, and ended on the same verdict without seeing each other.

What has to happen

People have to start big projects from the bracketed outcome instead of a plan, collect evidence the way the loop asks for it, record conversations with consent and feed in the originals, use a vote rule with a tie-break, and treat “don’t do it” as a successful result.

Where I am probably wrong

Every piece of evidence here was simulated by AI, and every lens, simulator and editor ran on the same underlying model. Agreement between the tracks is weaker than agreement between independent teams. Real customers surprise more, real files are messier, and real CEOs negotiate their own roles in ways no lens can. If students run the loop on real projects and the “next evidence” keeps coming back expensive, vague or different on every run, then the discovery I’m describing was an artifact of a tidy simulation. In that case I’ll have taught a ritual.

Back to the kickoff room. Forty slides and nine workstreams. Now picture the alternative on the table: one press release, half of it in brackets, and three lenses that have already said which bracket matters most and who can fill it. The project hasn’t started yet. It has something better: a phone number.

Find the fact. Get it. Run it again.