Participant disclosure

This site is published by the sgit project, which builds the tools and works by the practices it describes. That is not a neutral vantage point, and a site is better off saying so on the way in than being caught at it later.

Who publishes this

Publisher	The sgit project, sgit.ai
Author	Dinis Cruz, with AI co-authorship. The agents of this site write under their own names (startups.engineer, startups.journalist, startups.interviewer, startups.librarian) and the editorial responsibility is Dinis's.
Repository	SGit-AI/SGit-AI__Website__Startups, public, including every release and every failed build
Licence	Content CC BY 4.0, code Apache 2.0; the licences page

The conflict, stated plainly

A project that sells a substrate for startups, publishing a site that tells founders what to ship and recommends building on that substrate, has an obvious incentive. Three things are done about it, and none of them make it go away.

1. The brief is published in full. The voice memo this site starts from is on the site, verbatim, with the message that started the session. If the site overstates its source, the source is one click away.
2. The finder does not recommend sgit. Its questions, signals, routes and rules are published as data and none of them mentions a vault; a route is picked by a rule a reader can check. The vaults page is a separate section and says what a vault does not do.
3. What is not built gets its own page. What exists says, by domain, what is on this site and what is not; the corrections file says where the brief was departed from and why.

How this site is written

By AI agents, from a voice memo and a two-line prompt, with Dinis Cruz reviewing and directing. The first release was built in one session; the agent article accounts for it. The prose was drafted by a model, the decisions were Dinis's and are recorded in his words, and everything had to pass a release gate that does not care who wrote it. The gate checks what is checkable: links, twins, versions, canonical addresses, credential shapes, the finder's data. It cannot check whether a claim is true. That part is the person's, and where a claim is uncertain the house style is to say so rather than round it up.

Where this approach loses

• A founder who needs a verdict. The finder refuses to say whether an idea is good, by decision. Somebody who wants to be told will not get it here.
• A startup that cannot be shipped small. The whole site leans towards the smallest usable thing in a stranger's hands this week. A capital-first startup gets one route, old school, and little else; the site has no advice on raising beyond "do it from strength".
• Anyone the eight rules do not fit. Eight rules and twelve signals are a first reading of one memo. The record shows which rule fired so a person can disagree; it cannot show the rule that should have existed.
• A reader who wanted the vault pages to stay on sgit.ai. They moved. Until sgit.ai's own pages carry the notice, the two sites disagree about where the section lives, and this one says so on the page.

The defences are the ones available: every number is counted from a file and the file is named; the interview is data a reader can check against the rule that fired; the gaps have a page. None of that removes the incentive. It makes the site checkable by someone who does not share it.