Accuracy

The real numbers.

Not a marketing claim — the last measured run, dated, on a photo set the scanner was never tuned against. This page updates every time that number does.

What the scanner gets right

KitID is built category-first: what class of gear something is, its camo, and its rough price tier are the part we're confident enough to lead with. Brand and model are a bonus — shown only with an honest confidence band, never guessed past what the photo actually shows.

What users tell us after

Every result asks “Right?” — the verdicts below are every answer we've logged, not a sample.

n = 98 verdicts logged

“Close” means a user marked the category or general item right but the specific brand or model wrong or missing — the honest middle between a clean hit and a real miss. Every rejection and every close call feeds the correction queue: pick “or <candidate>” on a result and, once it's confirmed, that photo becomes part of the next eval run.

How we measure

We hold a set of photos back and never tune the scanner on them — that's the only way a number here means anything. Every merge to the scanner has to clear hard invariants first: zero wrong brands, zero overclaims, and a confidence cap when the visual “tell” between two lookalike products isn't actually visible in the photo. Brand and model recall are still informational at this sample size — we treat them as a read, not a ship gate, until the labeled set is bigger. Category recall is the number we're building toward next, because it's the layer the product leads with.