Methodology

Methodology changelog

Semver. Methodology § 16: a published score changes only through an entry here (the method changed) or through the corrections log (a fact changed). Every entry is written in plain language — what changed, which cards moved, why — and is rendered on each site's methodology page from this file.

1.1.1 — 2026-09-11

No weight, threshold, rubric or flag changed. Every card keeps the points it had. What changed is who may be shown, how the gap in a card is measured, how the site's fixed wording is guarded, and who decides.

Two numbers instead of one. A card used to carry a single coverage figure, and a card under 75 % was held back as pending. That single figure counted a seller's own silence as our gap: a seller who publishes no prices closes the capital-to-scale axis, a seller who publishes no list of finished buildings closes the completion axis — and under the old rule both pushed the card into pending, so the least transparent sellers dropped out of the rating entirely. That is the same defect the method has now met four times (court counts, the delisting procedure, the insolvency row, and this): the opaque seller must never come out ahead of, or vanish beside, the transparent one. From this version checked measures us — did we finish every check the country's registries allow — and must be 1.0 for a card to score; an axis closed by the seller's own non-disclosure, or by a registry that would not answer three times over a week, counts as checked with zero points. confirmed measures the seller — how much of the public record a registry actually confirmed — and is the figure shown next to the rank. Cards that moved: in Bulgaria, most of the 44 cards that were pending under the old number are scored under the new one; their points did not move, their visibility did.

Pending is not published at all. No page, no row in the table, no sitemap entry, no noindex stub. A stub is a publication too.

The fixed wording is guarded by its hash, and the lexicon matches whole word forms. The approved templates, disclaimers, band names and policy texts are listed by SHA-256; the build no longer parses them word by word, so a forbidden term inside our own approved sentence cannot fail the build and cannot tempt anyone to rewrite the sentence around the gate. The lexicon matches word forms, not substrings — a forbidden Bulgarian adjective is not matched inside the longer, ordinary noun that happens to contain it. A forbidden term inside a quotation from a seller's own page drops the quotation, logs the term and keeps the evidence — the criterion still scores, because the fact rests on the URL, the date and the hash, and the quotation was only ever a display. Bulgarian versions of every fixed string are approved and used verbatim; English stays the normative original, shown once on the methodology page.

A link to a related company is a fact, with its own id. The continuity axis pays ten points for "linked in the register to the disclosed group", so a link without a fact behind it was a score derived from nothing. Every linked_entities[] entry now needs a legal_entity fact on the seller's record — the related company's number, registration date and status, and the counts of shared officers and shareholders, never names — or it is neither scored nor shown.

Officers' names are stored as salted hashes. Linking needs only equality, so the data holds a hash of the normalised name, never the name; the clear text lives in a temporary file during a linking pass and is not committed.

The human gate and the wave gate. "B1 = 0" in the human gate means found_inactive — a company found in the register and not active — and such a card must be checked for the sixth red flag (still selling under the brand), otherwise the gate held the card for a flag nobody looked for. A wave is published when six listed conditions are met, and a person is not one of them: no axis unknown in the wave roster, the known-leader test green, zero unreviewed flags and inactive-entity cards, imprint and policies live, a test letter answered from the corrections address, and the live-site check green including the forbidden-string gate.

The wave roster is picked by market presence and frozen before collection. Twenty to thirty sellers, chosen by how much they build in the city, written down with source and date before any collection starts. Picking by what we had already collected would rank only the sellers who disclose.

Money, stated at its real edge. Forbidden: payment from a rated company, from its broker or agent acting for it, or any payment that changes what a card says. Allowed and said plainly: a check commissioned by a buyer for their own purchase; labelled advertising by services that sell to buyers (lawyer, notary, escrow agent, surveyor, mortgage broker), only on the guide and check pages and never on a rating page, the index or a card; licensing the corpus. The offer is live from the first wave. The 90-day economic stop counter counts enquiries as well as payments, from the day the wave is published.

Gate 0 decides which asset a domain carries, not whether one is built. Tier 1: the gate passed, a full rating. Tier 2: the gate failed but sellers publish sites, a disclosure register — Layer A only, every card disclosure_only, a fixed notice on every page that the country's registries are closed to free checking — plus a guide and the check offer. Tier 3: no seller sites either, the domain is unused.

Who decides. From this version the operator owns the method. There is no approval loop and no memo asking anyone for a decision. An ambiguity is resolved by a five-step ladder — invariant, provenance, anti-inversion, publish less, reversibility — written as one line in decisions.md, and work never pauses for an answer.

1.1.0 — 2026-09-09

Three things changed, none of them a weight.

Where a fact may come from. Every source now carries what it demands of the person asking: nothing (none), an e-mail login (account), a card payment in the payer's name (payer), a resident's electronic identity (national-eid), or a visit in person or through a local representative (in-person). Only the first two may feed a card. A paid extract or an e-ID login leaves a record of who asked about which companies inside the country where those companies live — the one trail the publisher's whole design is built not to leave. So a source that costs money or identity is not budgeted; it is not used, and the axis it would have fed simply leaves the country's denominator, exactly as if the registry did not exist. No seller loses points for it. A paid document exists in one place only: the dossier a buyer orders under their own name for their own purchase. Every fact records its level in retrieval.identity, and a fact above account fails the build.

Which countries can be rated. Gate 0 now asks two questions instead of three: is there a free (or login-only) register of companies, and is there a free register that says who built a finished building — not just that a building stands at an address. Court portals no longer gate admission: courts left the score in 1.0.0 (§ 7), and what cannot move a score cannot decide a country. A country passes when at least 60 % of the Layer B points it could offer are collectable free and anonymously, completion with attribution among them. Spain fails this today (8 of 36 free); Bulgaria passes (Sofia first, because its attributable completion register is municipal).

One new row in B4. An insolvency proceeding that ended because the company performed its agreement with creditors, every resolution dated in the register, scores the full 8 points and raises no flag. It is shown on the card with its dates and case number. Without this row a company that went through insolvency and paid everyone back would be scored like an opaque liquidation for ever — the transparent loser to the opaque, a confound this methodology already removed twice. One Spanish card moves from "awaiting review" to this row; nothing else moves.

1.0.0 — 2026-09-07

Initial methodology. methodology.json is extracted verbatim from docs/handoff/devreg-methodology.md § 21 and is not edited by hand; a doctor check compares the two on every run.