How we measure brand-detection precision.

PermitCore detects the manufacturer brand named in a permit’s free-text description — the surface behind the /brands directory, the permit-page brand chip, and the Market Explorer brand filter. Every surfaced mention passes a precision gate: ≥99% precision on the commercial-ICP brand subset (100% on the commercial-HVAC slice, n=116, all NYC), measured via a bidirectional-sample audit (2026-06-30) using gazetteer matching + negated-context filtering. When we name a manufacturer on a permit, we’re right ≥99 times out of 100. The 100% commercial-HVAC figure is a small, concentrated slice (116 permits, all New York City) — the entire pure-commercial-manufacturer pocket in the current corpus.

Precision here is the standard definition — of the brand mentions we surface, the fraction that are correct (true positives ÷ (true positives + false positives)). It is a precision claim, not a coverage claim: to keep what we show trustworthy, the detector is tuned to stay silent when it isn’t confident, so it will miss some real brand mentions (lower recall) rather than surface wrong ones. When a permit’s text names no brand, we show nothing — we don’t infer or fabricate one.

What this figure is not. ≥99% is our brand-detection precision — when we name a manufacturer, that attribution is right. It is not a claim about segment composition: how cleanly a permit is routed into a segment like Commercial HVAC is a separate measure, and today it is lower (some non-HVAC work is still mis-routed into that segment; a fix is in progress). So “≥99% brand precision” means we correctly identify the manufacturer — not that every permit in a segment belongs to it.

That trade is deliberate: for replacement-targeting and competitive-displacement workflows, a brand attribution you can act on beats a longer list you have to second-guess.