What Growth-OS will not do, and why

A campaign in Growth-OS is an audience, a hypothesis, a message, an action and a result. One that cannot say what it was trying to learn is not a campaign, and the record will not let you pretend otherwise. Angles come from the evidence themes we actually recorded, and an angle with nothing behind it is labelled a hunch in the type rather than dressed up as evidence.

A recipient snapshot freezes who the audience criteria selected and who suppression removed, each excluded account carrying the rule that decided it. Criteria the audience grammar cannot read refuse the freeze instead of quietly widening back to every account. Then it stops. No email, message or invitation leaves Growth-OS.

Fifteen sources are declared on the Sources page. Each one states what truth it owns, what it would read, what Growth-OS may write back, what kind of credential it needs and where its authorization boundary sits. That declaration is design, not a connection.

Adapters exist for most of them and were built against fixtures. No provider was called from this repository during development and no credential for one existed there. The Sources page is the only place that reports what is actually connected in your workspace, and until somebody adds a key and runs a sync, the honest answer is that nothing is.

An account is compared against a chosen ICP revision. Each weighted dimension carries a status and a reason you can read. An unknown input leaves the denominator rather than counting against the account, a criterion nobody supplied an input for is omitted from both totals and named to the reader, and no score appears at all until enough of the declared weight can be evaluated.

Every score is a share of the weight that was compared, so the screen states its own denominator and two accounts are only comparable through it. The result is pinned to the ICP record and revision it was computed against. Disqualifiers are raised for a person to look at and reject nothing on their own. There is no trained model here and no prediction of whether a deal closes.

Ask searches the project with indexed lexical retrieval and reranks a bounded candidate set, so an older matching record stays findable. Passages keep record versions, dated evidence and review metadata. Every answer cites the records it used.

What the citation proves is narrow, and worth saying out loud: it proves a source was supplied, not that the source supports every sentence in the answer. Ask cannot browse the internet, cannot mark a source verified and cannot save a proposal as research. Read the answer, open the citation, then decide.

The operating figures on Today each carry a numerator, a denominator, a window, a source, a definition in prose and the record ids behind the number. A share of an empty denominator is its own state and is never rendered as zero percent, because those two things mean opposite things.

They describe how the company is operating: what is waiting, what is stale, what has no owner. They are not a scoreboard of individuals, and no figure is built to be read as one. If a number looks bad, the record ids under it tell you which records to fix.

“A tool that quietly does more than it claims is worse than one that does less.”