Our accuracy, measured live. Run it yourself.
Two public benchmarks run against the same engine that answers every client match. Nothing is special-cased for the test. Press the button and the numbers come back with a receipt anyone can re-check.
The public sets, live.
A representative CRM file shows field performance: realistic noise with web domains at real-world fill. An adversarial set shows the honest floor: built to break us with no domains to match on. Each is around 200 rows and deliberately nasty, built from public CC0 GLEIF data.
Point an agent at the same endpoint or press the button. Results cache per reference pack, so running it again on the same pack lands the identical receipt and two parties can compare numbers without trusting either one.
Run both public benchmarks yourself, live against the serving pack. No key needed. A representative CRM file shows field performance and an adversarial set shows the honest floor. Both are small and deliberately nasty. Results cache per reference pack so every run returns the identical receipts.
Both numbers print together. The match rate says how much work the engine did. The wrong-entity rate says whether you can trust it. A close call routed to review counts as neither right nor wrong.
Both sets plant fake companies with no real counterpart, and the no-match line counts the engine refusing to invent a match for them. A 100 percent refusal means zero hallucinated matches.
37% fewer unmatched records.
Between two engine versions the share of records we could not match fell from 8 percent to 5 percent on a held-out test. Measured, not projected.
Matched 92.0% on v8.10 and 95.0% on the v8.11 deployed arm: unmatched fell 8.0% to 5.0%.
The full measurement and its method are published in the measurement record.
ER-7: the public company-ER benchmark.
The public benchmark for company matching. On the clean set our engine is right 99.8% of the time on what it links, while tools that see nothing but public company names merge roughly half of what they link. Rules were frozen before any number existed and the datasets ship CC0 so anyone can regenerate them.
How we count sources.
By sources of record we mean distinct underlying registration authorities witnessed on our ingested records: the company registers, filing systems, and authority codes our records actually carry. Counted and never estimated.
Measured: 937 distinct registration authorities counted against the 2026-07-14 GLEIF golden copy with format-verified codes and sentinels excluded. Direct feeds and referenced registries only grow the union. Where the site rounds it rounds down.
The landing ledger binds the served count from /api/stats and activates only when the api serves it. The coverage atlas catalogues the pull-sources jurisdiction by jurisdiction.