How it’s built

A careful pipeline, checked at every step.

What happens between the documents you give your Rep and the map it shows you, and the rules that keep it honest along the way.

From sources to capabilities

A model never jumps straight to “you’re good at this”. Your Rep reads your sources, pulls out claims, links the claims that agree as evidence, and only then assesses a capability, which has to cite the evidence behind it.

Models propose. Rules decide.

Every model output is a proposal. It is checked against a schema, then against rules that don’t bend: no claim without a source, no quote that isn’t verbatim, no capability without evidence.

Only then is it stored, with a record of which model, which prompt and which inputs produced it. Nothing goes from a model straight into your profile.

Six kinds of truth

Your Rep keeps track of how it knows each thing. These states never quietly turn into one another.

Source fact: Read in something you gave it.
User claim: Something you told it.
Assisted claim: From work made together with an agent. Held back until you defend it.
Inference: Your Rep’s own reading, always labelled as such.
Verified fact: Confirmed by an independent second source.
System assessment: A judgment that must cite the states above.

Defence and retraction

When evidence comes from co-produced work, your Rep asks you to explain a decision in it. Defended, it becomes evidence of your judgment. Not defended, it is retracted: kept on record with the question and the answer, and never claimed again.

Intent, measured

What you say you want is a start. What you turn down is the data. A short triage of offers, each a real trade-off, reveals the rules you actually decide by. They stay proposals until you confirm them.

User zero

The first person Amble read was one of us. The real pipeline has no seeded profiles and no hidden fixtures: if it can’t read someone honestly, it fails honestly.

What it runs on

TypeScript end to end. A provider-agnostic model layer, so no business rule depends on one vendor. Postgres for the evidence graph, with every model run logged. This site: Astro, no cookies, no tracking.

Open source, carefully

We’re preparing the code and the career.md format to be published. It needs careful review first, and we’d rather do it right than fast.