Certified where measured. Disclosed everywhere.

Every cohort classification we serve carries a per-jurisdiction, per-cohort-family certification status — one of exactly three states, and the third is the default:

Certified — a real precision measurement of ≥99% exists for that jurisdiction and cohort family, backed by a certification artifact: the measured precision, the sample size, the measurement date, the cohorts covered, and who measured it, with the method and primary sources cited in the body. Where a certification predates this system and the exact value was not recorded, the artifact says so explicitly (value_basis: gate-floor) — an inference is never presented as a measurement.

Measured below the bar — a real measurement exists and came in under ≥99%. We serve the data anyway and disclose the actual number and date, inline, wherever those results appear — in the API response envelope and on the cohort pages. We don’t hide data behind a failed measurement; we attach the measurement.

Not yet measured — the default for everything else. If no measurement has been filed, the membership discloses as unmeasured automatically. There is no silent state: absence of evidence is itself disclosed.

As of the status map published 2026-08-26 13:27 UTC: 51 memberships are certified — measured, with a sample large enough that the lower bound of the 95% confidence interval clears 99%. 37 are measured but indeterminate: a real measurement exists and the interval spans the bar, so we can neither certify them nor say they failed. Each one discloses its own sample size and interval. 3 are disclosed below the bar at their real measured values. 10 rest on a founder attestation rather than a sample we can show you. And 1771 are not yet measured — across 144 jurisdictions.

The indeterminate count went up on 2026-08-19, and nothing got worse that day. Certification used to compare the measured precision against the bar and ignore the sample size sitting beside it — so a perfect result on 30 permits certified, even though 30/30 has a 95% interval of 88.7%–100%. The bar is now the interval’s lower bound, which at 99% needs roughly 400 adjudicated permits. Memberships that moved were always this uncertain; the label is what changed. A measurement program is working through them, and memberships with no independent publisher signal to audit against stay honestly unmeasured rather than being forced through a manufactured measurement.

Certification measures precision. It does not measure scope. Of the permit records we hold for a jurisdiction, only some carry the work markers our cohort rules need — the rest are published without a residential or commercial signal we can classify on. Every membership in the status map therefore also carries two numbers: how much of that jurisdiction’s corpus is eligible at all, and how much of the eligible part we have placed in a cohort. The first is a fact about what the source publishes; the second is a fact about us. Across the 48 jurisdictions holding a certified membership, that first figure ranges from 20.0% (naperville) to 100.0% (cambridge). We state it per jurisdiction because there is no honest average of it.

The shape of that loss, across those same jurisdictions: 14,255,562 permits held, of which 8,820,709 carry a classifiable work marker, of which 8,406,574 are placed in a cohort. Both ratios are row-grain, and the counts are published beside the percentages in the same file so you can recompute them rather than take our rounding. Measured from the classification snapshot of 2026-08-26 11:22 UTC.

These counts are recomputed from the status map every time it is published, and they move only when that timestamp moves. If you noted a number here and see a different one later, check the timestamp above: a different timestamp means the map was republished — a re-measurement, a new artifact, or a rule change, all of which we describe on this page. The same timestamp with a different count would be a bug, and we would want to hear about it. The map itself is public and dated, so you never have to take our word for which of those happened.

Verify it yourself. The complete status map — every jurisdiction, every family, every value and date — is public: certification-status/latest.json. The numbers on this page are computed from that same file at render time; they cannot drift from what you can fetch. API responses on /v1/permits/target carry the same statuses inline (certification_disclosure) for exactly the jurisdictions and families in each page of results.

The certification requirement is enforced in CI, not policy: a jurisdiction cannot enter our curated surface without a valid certification artifact, a below-bar measurement can never admit anything, and the measurement floor is machine-checked on every change. To our knowledge, no one else in this market publishes per-jurisdiction classification precision at all.