The words we use, and exactly what they mean.
Several terms on this site carry a specific, load-bearing meaning rather than a loose one — "bounding" and "adjudicated" are two of them. Each is defined here exactly once, and used the same way on the platform page, the API docs,published accuracy and coverage.
- Adapter
- The piece of the pipeline that reads one specific public register or portal and maps its columns onto our field vocabulary. An adapter can be implemented and "live-schema verified" — meaning it correctly reads the source today — before it is cleared for customer use, because reading a source and having the right to serve its data as a product are separate questions.
- Adjudicated corpus
- A set of expectations — "project X should have capital cost Y" — built mechanically from an authoritative register, used to score discovery recall and value accuracy. Nothing in it is typed by hand or judged by a model; the register itself is the adjudicator, named on every scorecard.
- Append-only
- The rule that a new value for a fact never replaces the old one. Both remain, each with its own evidence and dates, so "what did we believe in March?" has an answer and a developer’s said-versus-did is measurable rather than erased.
- Bounding
- The block on every API response stating whether the answer is complete, what scope it was narrowed to, what was left out, and what it cost. A truncated result says so explicitly instead of looking like a short but complete one.
- Canary
- A recurring, automated production check confirming a source adapter still imports correctly against the live register. Paid availability for a geography requires seven consecutive daily canaries to pass, not just a one-time import.
- Evidence chain
- The document, version and exact quoted words behind a stored fact. A value with no evidence chain is either discarded or explicitly flagged as inferred and ranked below anything that carries one.
- False assertion rate
- One of the four published accuracy metrics: facts asserted that the authority register says do not exist. Lower is better, and it is reported per geography rather than folded into a single score.
- Field vocabulary
- The versioned set of predicates (capital cost, commissioning date, and so on) that every extracted fact is mapped onto. Each column of each source is mapped onto it once, in reviewed code, rather than guessed per document.
- Knowledge-ready
- A source adapter that reads correctly and produces mapped facts, but has not yet cleared the legal and evaluation gates required before those facts can be sold to a customer. Knowledge-ready is a technical state, not a commercial one.
- Launch gate
- The combined set of conditions a geography must clear before paid checkout opens for it: legally usable customer data, a fresh complete import, seven consecutive daily canaries, and a recent independent scorecard with useful denominators. No single strong condition substitutes for another.
- Package
- A discrete piece of work inside a larger project — the scaffolding contract, the boiler replacement, the turbine order — tracked separately from the project it belongs to, because a lead worth calling is usually stated at package granularity, not project granularity.
- Parity (field coverage)
- The count of fields an incumbent typically sells that this system also answers, reported apart from the count of fields it answers that no incumbent offers at all. The two are never added into one number.
- Round-tripping
- What happens when a register is scored against itself: the system finds close to 100% of what it was also given as input, which measures agreement with the source, not discovery. Expectations supported only by an ingested register are excluded from recall and accuracy for this reason.
- Said-versus-did
- A comparison between what a developer or contractor originally announced and what the append-only record shows actually happened, made possible only because superseded values are retained rather than overwritten.
- Scorecard
- The published, per-geography accuracy artifact: expectations, adjudicator, metrics and exclusions for one register and one evaluation run. Every scorecard stays readable at its own address once published, and none are averaged together.
- Source class
- One of twelve categories of public evidence the architecture can represent — company registries, planning authorities, environmental permits and so on. A source class being representable is not a claim that it is currently producing customer data.
Why definitions live on one page
A term used loosely on a marketing page and precisely in the API is worse than not defining it at all, because it lets a reader believe the loose version. Everything above is the same word, same meaning, wherever it appears on this site — including in the frequently asked questions and the per-geography coverage detail.