Methodology

Every factual claim in a Note carries a label telling you exactly how well-supported it is. This isn't decoration — the labels are load-bearing. If you're deciding whether to commit bid-team hours off the back of a claim, you should know whether it's a fact from an official filing or a judgment call, and how confident that judgment call is.

The four tiers

OBSERVED

Stated directly in an official record — a TED notice, a national contract register, a company registry entry. No interpretation. If it's OBSERVED, the source document says it in those terms.

DERIVED

A mechanical computation over OBSERVED facts — a sum, a count, a share, a date difference. No judgment involved, but coverage matters: a sum over 2 of 5 known values is disclosed as a partial sum, not silently presented as the whole picture.

INFERENCE · CONFIDENCE

An interpretation that goes beyond what the records state outright — what a pattern probably means. Every inference carries an explicit confidence level (HIGH / MEDIUM-HIGH / MEDIUM / MEDIUM-LOW / LOW) and, where the evidence is genuinely ambiguous, the competing explanations we didn't rule out.

VERIFY

A concrete next step — a specific register, filing, or document — that would let you confirm or overturn the inference yourself, before you rely on it for a bid decision.

Sources

Every source is official and open-licensed. We don't scrape paywalled databases, and we don't represent third-party commercial data as our own.

Independent review, not self-check

Before a Note reaches a human for sign-off, it's checked by more than one AI model, from more than one provider, working independently of the model that drafted it and of each other. Each reviewer scores the Note and can flag it as not ready — a Note doesn't advance on the drafting model's own say-so. That's a structural check, not a single model marking its own work.

What we don't do

When we're wrong

If a later review finds a Note got something wrong, the correction is published in the Note itself, in place, not quietly edited away. Note #001 has a live example — an earlier version treated "1 tender received" as evidence of low competitor interest; a later review caught that this is a structural artifact of how a certain framework type reports call-offs, not a sign of an uncontested market. The correction is printed directly in the Note, with what changed and why.

Who checks the work

The full account of how much of this is automated and where analyst sign-off happens is on the Transparency page.