One question answered.One still open.

Two ideas, two questions

On the ideas page I put two ideas up for scrutiny. GuestGraph resolves the strangers a hotel's systems store into one guest: the one who booked, checked in, ate, joined the wifi and left a review. CompanyGraph writes what a company knows as a graph of Markdown files that people and AI agents can both rely on. Both are open core: the substance is Apache 2.0, and exactly one part of each could ever earn money.

The page asks two questions of both. How much of an idea must exist before it can be validated — and what does that cost in hours and focus? What is a realistic outcome — revenue, learning or a clean no? Fifteen weeks of building later, the first has an answer with a number. The second has a test, and the test has not been passed yet.

Enough that it answers without me

More than I thought. A slide proves you can explain an idea. It does not prove the idea holds. An idea is validated only when someone who owes it nothing can try it while I am not in the room.

For CompanyGraph that meant the vocabulary and a check that fails when a reference points at nothing. It meant real instances rather than invented ones, and a chat and an MCP server that answer from them. For GuestGraph it meant the engine and one real hotel system connected, Apaleo, because identity resolution against invented data proves only that the invented data was consistent.

A slide proves you can explain an idea. Only a question answered without you proves it holds.

It happened on Sep 29

In a workshop on Sep 29, 2026, three people asked CompanyGraph what it could not answer yet: does it only work with Claude, and can a graph database like Neo4j hold a model? They were not my questions. That was the point.

An agent in the Writer seat drafted the entries, and I approved each one in the Owner seat. Thirty minutes later the chat answered both questions, and ten of those minutes were mine, by my own count; nobody timed it. The answers stand in the model as two objectives, a company keeps its model with whichever agent it chooses and any graph database can hold it. Both are targets. Nothing loads a model into a graph database today.

What it cost

Between Jun 9 and Sep 24, 2026, the two ideas and this site took about 434 hours of mine. That is an estimate from the days git shows work, not a timesheet: 26 days at five hours before Aug 17, and 38 days at eight hours from then on. The AI subscription came to about 650 francs.

The money is the small part. Focus is where the cost went. I wrote few of the lines, but I decided what nearly all of them say. Building stopped being the bottleneck in the first weeks, and every hour it saved went into deciding: what an entry claims, what a page may say, what gets declined. The workshop's thirty minutes were the same pattern, small: the agent did twenty, and the ten that stayed mine were the deciding.

A realistic outcome

Revenue: none. That is not a missed target. It is the current state. Nothing is sold, there is no rate and there is no proposal. CompanyGraph's LinkedIn page is typed Nonprofit, because that is the one type that claims no business.

A clean no: not yet either, and I have made it hard to dodge. Both models carry the question as a strategic objective of its own, and each says a no settles it as completely as a yes. CompanyGraph's asks whether it is a business. GuestGraph's reads:

Whether the scattered-guest problem is real enough in a hotel's operation that someone would pay to have it solved is settled, in either direction, by hotels rather than by us.

Learning is the realistic outcome so far.

What was learned

The first test is adoption, not revenue. Every CompanyGraph instance so far is ours, so every question the vocabulary has been asked was asked by someone who already knew the answer. The objective that comes first is a company we have never met keeping its own model, current, without us in the room. That has not happened yet.

Validating cheaply is a decision of its own. The skills an instance ships are written for Claude first, and that is recorded: supporting every agent before anyone outside keeps a model would be work spent on an idea that may still get a clean no.

And validation changes the person doing it. Building both ideas in the open answered the question my career break started with, what I like doing rather than what I am used to being hired for. It is why I chose an architect role from October.

What would settle it

Three things, each small on purpose. One company we have never met that keeps its own model. One engagement billed by the day, on the terms the billing page states. One hotel that pays for the hosted service. One of each is enough, and a larger one would not answer the question better.

If the honest answer is no

If you are that company or that hotel, or if you think the honest answer is no, tell me: a no is worth more than a polite yes.

And if you have a question about CompanyGraph, ask the chat before you ask me. If it can answer, the idea holds a little better. If it cannot, you have found a gap in the model, and closing it takes about thirty minutes.