Evidence, not confidence
This is the rule the whole design rests on: a model asked to grade its own certainty will, and it will be wrong in the direction that makes it look useful. So no tool in the agent accepts a confidence, a score, or a sourceUrl offered as proof.
How it works
A tool reports what it observed — crm.signature-block, github.account-identity, linkedin.profile-match — and a ledger prices that observation. The weights, the combination rule, and the bands live in one file, versioned like code.
| Band | What it means | What happens |
|---|---|---|
| Verified | A primary source, strong enough to stand alone | Written to the record automatically |
| Probable | Real evidence, not conclusive | Stored as a suggestion — a rep decides |
| Below that | Weak or circumstantial | Nothing is written |
The bands are behavior, not labels. Probable is a correct outcome, not a failure: when four people named Marchetti work at the same company, the agent won't silently pick one.
Three things the write path enforces
The only route to a contact's fields enforces what a prompt cannot:
- Never overwrite a human. A value a rep typed always outranks the agent's own research.
- Never re-offer a dismissal. A suggestion you dismissed is remembered — it is never proposed again.
- Never write without a primary source. No primary source, no automatic write, regardless of how the evidence combined.
Suggested facts, in the app
Some fields on a contact's Overview tab show as a suggestion — a small card with the proposed value and the evidence behind it — rather than a set value.
- Read the rationale under the suggestion.
- Accept writes it to the record; Dismiss drops it.
- A dismissed suggestion is never re-offered.
Why guessing is the one failure that matters
A confidently wrong fact about a customer is worse than a blank field, because nobody can tell it is wrong. The same rule governs anything you ask the agent directly in the Agent tab: if it doesn't know something, it says so rather than filling the gap.
Two copies of an identity matcher — one in the API, one in the agent — once drifted until one of them matched every employer on earth. That outage is why all intelligence lives in the agent and nowhere else.