Published September 13, 2026
The rules that must be fixed before the first customer can bend them
Nine essays have made an argument. This one converts it into commitments, because an argument that costs nothing to hold is not worth much — and because the specific failure mode of this thesis is not being wrong. It is being right and then, one reasonable decision at a time, becoming the thing it described.
The claim of this essay: every rule below will be tested by somebody offering money to break it, and the ones worth writing down are the few where the offer will be tempting.
1
Why a Constitution and Not Principles
Most companies publish values, and most values are unfalsifiable. Customer obsession, bias for action, integrity: nobody can breach them because nobody can say precisely what would count as a breach. They cost nothing, which is why they are so widely held.
A constitution is different in one respect. It names things the company will refuse when refusing is expensive. A rule that will never be tested is decoration. The test of whether a clause belongs here is simple: can I imagine the meeting where somebody makes a persuasive case for the exception, and can I imagine wanting to agree?
That test removes most of what would otherwise be on the list. It leaves four refusals, four promises and one number.
One structural condition sits behind all of them and cannot be written as a rule, because a rule can be overruled by the organisation that hosts it. This company must remain organisationally separate from any incumbent business whose sales model, customer commitments or existing architecture could override what follows. The Company the Price Builds explained why: the functions being removed are not costs inside the old organisation, they are the old organisation. A constitution written inside a company that can suspend it is a memo.
Figure 1. Nine essays reduce to this.
2
Four Refusals
No customer-specific fork. A capability enters the product only when it is common to a defined segment, expressible as a reusable specification, testable, and operable on the shared core. A customer may configure its own rules. The company will not maintain a branch for one customer, however large.
Why this will be tested: the first meaningful contract will come with a requirement that is almost general. Agreeing once produces revenue and a maintenance obligation that never appears on an invoice. Agree three times and the company is an AI services firm with a product-shaped brochure — which is the failure mode Inside the Foundry named, and the one that arrives disguised as traction.
No customer large enough to rewrite the model. No single customer may hold enough revenue that its departure would change what the company builds. This constrains what can be accepted, not only what is pursued.
Why this will be tested: an enterprise buyer will offer more than the next hundred merchants combined, and will want procurement, a security review, a named account team and a roadmap commitment. Each is reasonable on its own. Together they are the old company shape, purchased at a premium and paid for later.
No second vertical before the first core compounds. Products two and three are adjacent jobs on the same core. Expansion into a new market waits for the number in section four.
Why this will be tested: a second vertical always looks like growth, always looks urgent, and conveniently postpones the only measurement that decides whether the production system is real.
No quiet removal of the promise. The obligation described in The Expensive Last 10% — accountability, migration, incident ownership, consent provenance, reversibility — may be automated, industrialised and made cheaper. It may not be deleted to reach a price.
Why this will be tested: it is the easiest margin in the business and the last one anybody notices. Removing the obligation does not lower the cost; it moves the cost to a customer who has not been told they are carrying it. This is the clause most likely to be breached by accident, through a series of individually sensible economies.
3
Four Promises
One bill. Adjacent jobs share one core, one memory of the business and one invoice. Capabilities remain separately understandable and separately switchable; the customer adds a capability, never another company to manage.
Eight jobs. A defined set of jobs, done properly, on shared foundations. Not a suite pursuing completeness, and not a thin app pretending its neighbours do not exist.
A tenth of the price. Priced against the incumbent bill being replaced, not against each product in isolation. Published, monthly, with no contract required, no seat tax where seats create no cost, and metering that is visible and capped by default.
A safe way out. Full export of data and of the reconstructed operating specification, a documented account of what stops working, and a transition window. Available as a documented capability rather than as a retention conversation.
One operating rule stands behind all four, and it is the one that makes them checkable rather than merely stated: the obligation is inspectable. Which specifications are current and when each was last reviewed. Which dependencies are monitored. What authority the system holds and who granted it. How a consequential action is reconstructed, and how quickly it can be reversed.
At a low price this is not a courtesy. A cheap product and an unpriced risk look identical from outside, and published operating evidence is the only thing that tells them apart.
4
One Number
Everything in this series rests on a claim that can be settled with evidence, and the claim is not that the software will be good. Impressive first products are now within reach of any capable team with agents; 2026 will be full of them, and none of them proves anything about a production system.
The claim is that production compounds. Inside the Foundry set the standard and this constitution freezes it as the single test:
Is the second product materially cheaper, faster and safer to produce than the first — and the third cheaper still?
One test, not a dashboard. The measures in The Company the Price Builds are instrumentation for running the company; this is the condition on which the thesis stands or falls. It resolves into four figures published together: elapsed time from specification to dependable operation, proportion of the product drawn from existing machinery, defects reaching customers, and support minutes per customer per product.
And a commitment about the commitment, because this is where such tests usually fail. The numeric threshold and its date will be published before the first product enters production — before anyone knows whether it will be met. A gate set after the results are visible is not a gate. Once published it does not move, and the current constitutional date is March 2027.
Three conclusions are possible and all three will be stated plainly: the machinery compounded; parts of it compounded but not enough; or a good AI software company was built and a foundry was not. The third is not a small outcome. It is a different one from the one being claimed, and the difference matters more than any individual product.
Two clarifications keep the test honest. It measures the machine, not demand — product-market fit cannot compensate for production that does not compound, and a successful first product proves nothing about the second. And reuse cannot be bought with quality: a threshold met alongside rising escaped defects, rising support load or rising operating cost has not been met.
5
A Debt, Dated
Inside the Foundry closed by promising that the next essay would show the dials. Six essays have followed and none has. The promise is unpaid, and a constitution that lists proof measures while owing an earlier one would be exactly the kind of document this essay opened by dismissing.
So the position, stated plainly: the dials do not exist yet, because the second product does not exist yet. They cannot be estimated, modelled or previewed without becoming the thing they are meant to prevent — a system grading its own homework in advance. The four figures above will be published within one quarter of the second product reaching dependable operation, favourable or not, alongside the assumptions behind them.
Until then, everything in these ten essays is what it has always claimed to be: a description of a machine, offered before the machine has run long enough to be judged.
6
Amendment
A constitution that cannot change is a superstition, and one that changes quietly is decoration. So the rule for changing it: this document is versioned and dated, every previous version stays visible, and an amendment requires a written account of what changed, what evidence required the change, and what follows for customers.
One clause governs the rest. No amendment may retroactively erase a missed commitment. A rule may be abandoned because reality disproved it — that is what evidence is for — but the rule, the failure and the reasoning remain part of the record. Governance is credible only when it preserves its own provenance, which is the same standard this series has applied to every specification it has described.
Three failures would make the document worthless, and readers can watch for all three. Silence: if the second product ships and no numbers follow, the omission is the answer. Amendment under pressure: a rule revised in the quarter it became expensive has not been revised, it has been abandoned with paperwork. Proliferation: if this grows to twenty clauses it has become a policy manual, and the four that mattered will be harder to find than they are today.
It is worth being concrete about how this ends badly, because none of it arrives labelled. The first custom request will look reasonable and small. The first enterprise customer will seem too important to decline. The first migration that needs a project team will be described as an exception. The first support queue will be easier to staff than to eliminate. The first missed gate will invite a better metric. Every one of those arguments will be intelligent, local, and supported by revenue — which is the entire reason the rules have to exist before the arguments arrive.
The series has argued that software production has changed, that the obligation around software has not, that a company built on both propositions must be shaped differently, and that a market should be chosen on distribution rather than on margin. Those arguments are now finished. What follows is arithmetic — and the useful thing about arithmetic is that it does not care how well the essays were written.
