Every list is built by a machine you can inspect.

Colder's own index of local businesses: name, category, website and phone for 3,814 businesses across 27 metros, kept current, de-duplicated, and screened so that chains, franchises and store-locator pages never reach a list. Nothing in it is scraped from Google, a directory, a forum or social media.

27 metros · 3,814 businesses · chain screen

Five dated government feeds, each pulled on its own clock and normalised into one ledger: USDOT registrations, NPI enumerations, food-service permits, contractor licences and child-care licences. Every row keeps the filed phone, the date of the event and a link back to the record. Future-dated and duplicate filings are held back, so a new row is new. 40,622 records held today.

fmcsa 24,233 · nppes 10,140 · permits 2,781 · contractors 2,466 · child care 1,002

Every record resolves to a state and a county, so a list can be a whole state or a single county, and the builder shows the live count for either before a row is spent. A business a customer has already received is never delivered to that customer again, on any play, in any territory.

2,622 counties with rows · live counts · never a repeat

Rows that name a person are flagged and left out unless the customer opts in, and then ship with manual-dial terms. In the four states whose data-broker laws reach a person's name, the row carries the official's title and not the name. Consumer data is never collected, bought or sold. Every row states which rules it was built under.

person flag · four-state name rule · manual-dial terms

Each candidate's homepage is opened in a real browser at phone width and read the way a customer would see it: what it says, what it links to, what it loads, what it asks for. 703 homepages read to date, each one recorded so the next check starts from what was actually on the page.

phone width · settled page · 703 renders

Eight deterministic checks that run the same way every time and quote their evidence: hiring, fleet, years in business, family-owned, online store, several locations, 24/7 service, financing offered. A program decides, not a model; a buyer can re-run any check by hand in ten seconds and get the same answer.

8 detectors · quoted evidence · no model decides

A language model reads the detector findings, decides whether the business is the buyer's kind of customer, and writes the one sentence a rep can say out loud. It is not allowed to state a fact. Anything it writes that contradicts a check is thrown away before it reaches a list.

one sentence · under 28 words · contradiction guard

An independent fetch of the raw page source from a different machine re-runs the same checks. If the two machines disagree the row is dropped. Nothing ships with a caveat. Site rows only; a record row's evidence is the record.

raw source · re-run checks · drop on disagreement

From the homepage and the about or team page: what they do, two to four things literally on their site, who to ask for when the site names someone, the founding year when it states one, and an opener that asks a question. Every content word must appear on the page; names and years must appear in the raw source too. Nothing it cannot quote ships.

word-level grounding · ask-for · no time-relative facts

Record lists build the moment they are saved; site lists build on a schedule, one row at a time so a crash loses nothing. Each row shows its sentence, evidence, tap-to-call number, opener, and keeps its outcome: called, interested, callback, not a fit, closed. Those outcomes are the ledger every later list is measured against.

statuses · notes · callbacks · csv