Seven systems held the truth about one business.
None of them agreed.

Two systems both measured membership, and they measured different things. The closed membership revenue in the CRM's entire history was a fraction of what the point of sale books from the membership base in a single year. Neither is wrong: the CRM holds new inquiries and the point of sale holds the installed base. Anyone reading the pipeline as the state of the membership base would have been off by an order of magnitude.

The records measured assignment, not effort. The CRM owner field records who a deal was assigned to, not who worked or closed it. Grading anyone on it would grade them on the wrong thing, so the system does not.

The loss data could not explain the losses. Loss reasons were rarely recorded, so price, product, timing and neglect were indistinguishable. The answer was instrumentation: a required reason on every close from here on.

One activity record in five was filed to a person instead of a deal. A sweep of every record showed re-filing them would change the last-contact date on a handful of deals. The rate was identical for every person and every activity type, a habit rather than a discipline problem. The cleanup was skipped; the order of the team's morning changed instead.

One place had three names. Normalizing to a fixed list showed most deals had never recorded a location at all. That is a data-entry job, and it now has an owner.

Each system recorded cash at a different moment. The point of sale sees cash when it is taken and the books when it is reconciled. The thirteen-week forecast reads inflows where cash is recorded first, and says so on its face.

Three calendars ran at once. A fiscal year that does not follow the calendar, Monday-to-Sunday management weeks, and a calendar tax year. The company's own week map had construction errors in future years. One calendar table now serves every model.

Numbers moved while they were being reviewed. The lost-deal count changed in the hours between a page being built and being reviewed. Re-pulling everything before every meeting is a treadmill. Stamping every page with its pull time is the fix.

Some sources are still manual. The point of sale exports by hand. Until a daily export runs, the shop pages are stamped snapshots and are labelled that way.

In a founder-led business, sales run through the founder's relationships, so any finding about the pipeline can be read as a finding about the person who built it. The argument was sequenced the way a long-tenured operator needs to hear it: what is sound before what is broken.

1. The machine works. The membership base grows, and inquiries arrive every month without being asked for. Say that first.
2. Demand is cheap at today's volume. Enough to justify scaling the test. Not enough to call the economics proven, since paid-inquiry conversion was unmeasured.
3. The clock is the constraint. Every recorded win closed within a few weeks of inquiry, and many open deals had gone longer without contact. Aging, not dead: each needs an explicit disposition.
4. Wins are limited by attention. Stated as follow-up capacity, never as anyone's calendar, because the CRM cannot support a claim about a person.
5. So earn the right to spend. Gates first, then the ramp.

Gates written as service levels. Three capacity tests gate the ad budget: the full sales team closes a membership start to finish; 95% of new inquiries contacted within 48 hours, two weeks running; 95% of open deals carrying a dated next step, two weeks running. A gate written as perfection gets abandoned the first week somebody is away. Instrumentation gates nothing and started immediately.

Area	Owner
Campaign execution	Marketing, to a written campaign standard, five numbers weekly
Pipeline response	Owner: response time, and a dated next step on every deal
Assigned deals	Sales team, from the daily queue
Shop	Shop manager: discount ceilings, targets, inventory
Member records	Membership staff: contract and renewal state
The instrument	Me: what gets measured, definitions, warehouse, pages

Adoption follows sequence, not instruction. The team had been asked for months to log from the deal record and to create a deal before calling anyone. Both failed because people worked from the phone, the inbox and the contact list. The fix was a third instruction: start the morning on the pipeline page. It is a deal queue with a dated instruction on every row, so the first two habits come with it.

Facts, procedures and dashboards live in different kinds of documents. When facts were written into procedures, the facts moved and the procedures kept the old numbers.

On one evening the pipeline figures changed materially within seven hours. The pages were corrected that night and two documents were not, so for a day the project held retired numbers. The structure below exists because of that day.

Operating canon

One document that routes every question: settled facts not to be relitigated, open questions each with a named owner and a closing condition, the precedence rule, the page index and the cadence.

Knowledge ledger

Every claim logged and typed as observation, assumption, risk, opportunity, decision or data quality, with source, confidence, dollar impact, owner and a required next-review date. Entries are superseded, never deleted.

Predict, then score

Before the month's actuals are pulled, a prediction is written down; after, it is scored confirmed or refuted. A note cannot be wrong. A prediction can, and that is what turns a review into a test.

Data dictionary

The point of sale's codes, crosswalks, caveats and merge procedure in plain language, so the knowledge does not depend on me.

Open-inputs register

Every missing input with the placeholder used meanwhile, who owes it, and what changes when it lands. Pages built on placeholders show the code on the page.

The board

Everything built and to build, scored on effort, value and severity, with the inputs each person owes. Updated at the end of every session that ships or files something.

Procedures as reusable instructions

The nightly refresh, the Monday funnel review, the cash refresh and the pre-flight alignment audit are written as instructions Claude runs the same way every time. A procedure that lives in someone's head is done differently each time.

The first build produced too many pages: six working pages about one funnel, a control room built to settle one day's disagreements, a setup guide and a separate ad review. Each was correct when built. Together they let the same number appear in three places and drift.

Asking why there were six made the real finding plain. Each page answered for a slice of the funnel someone owned, and the end-to-end number had no single owner. Four pages were retired at once, kept but not refreshed or shared. A cover page now routes readers, the question test governs every new page, and the ownership question sits at the top of the open list. It needs a name. It does not need a hire.