Our accuracy, published. Ask anyone else for theirs.
Scored against corpora built mechanically from authoritative registers. Nothing on this page was typed by hand and nothing was judged by a model — a model grading a model measures agreement, not accuracy.
There is no overall figure, here or anywhere in the product. One number across geographies would let a strong result in one carry a weak result in another, and the geography you are asking about is the one it would hide.
No scorecard has been published yet.
Rather than a placeholder or a sample figure. A number on this page is a number we are asking you to hold us to, so an illustrative one would be the exact failure this page exists to argue against.
Scorecards appear per geography as each corpus is adjudicated, and every run stays readable at its own address once published.
How these are produced
A corpus is built mechanically from a register — for Great Britain, the NESO Transmission Entry Capacity Register, which every transmission-connected project must appear in before it can connect. Each column is mapped onto our field vocabulary once, in reviewed code, and every expectation records which column of which snapshot it came from.
The adjudicator is therefore the register itself, named on the scorecard. Not a person who typed it, and not a model that guessed it — an authority you can go and check.
The trap, and what we do about it
Once we also ingest a register, scoring against it stops measuring discovery and starts measuring whether our pipeline round-tripped its own input. That reports something near 100%: the most flattering number this system could produce, and the least true.
So expectations supported only by the register that authored the corpus are excluded, and the exclusion is stated above the numbers rather than in a footnote. A fact another source also asserts still counts — that is corroboration, not circularity — and a fact we never found at all stays a miss, because hiding a discovery gap behind an anti-flattery guard would be the same failure by a quieter route.
What these numbers cannot tell you
A register states facts; it is not the prose our extractor reads. So this measures discovery recall and value accuracy. It cannot measure evidence quality, because registers carry no quotes — that limit is recorded on every corpus rather than left for a reader to discover.
Why nobody else shows you this
Measuring accuracy after the fact needs two things: a retained record of what you believed and when, and an adjudicated corpus to compare against. A store that overwrites has neither. It is not that incumbents have declined to publish accuracy — it is that their data model cannot compute it.
Questions people actually ask
How is CapexSignal’s accuracy measured?
Against an adjudicated corpus built mechanically from an authoritative register — for Great Britain, the NESO Transmission Entry Capacity Register. Nothing is typed by hand and nothing is judged by a model, because a model grading a model measures agreement rather than accuracy.
Why is there no single overall accuracy figure?
Because it would be misleading. A single number across geographies lets a strong result in one carry a weak result in another, and the geography a customer is asking about is precisely the one it would hide. Every figure here is per geography and none are combined.
Do you score yourself against data you also ingest?
No. When a register is also one of our sources, expectations supported only by that register are excluded from recall and accuracy, and the exclusion is reported on the scorecard. Scoring them would measure whether our pipeline round-tripped its own input and would report close to 100%.
Why do competitors not publish accuracy?
Because their stores overwrite. Measuring accuracy needs a retained record of what was believed and when, plus an adjudicated corpus to compare against. A table of current-value project records cannot produce either after the fact.