The method.

01 · The spine

One sentence, five parts

Every dataset in the system reduces to the same sentence: subject, place, time, value, receipt. A population figure, a tender notice, a committee decision and an award outcome are all the same shape once harvested. That is the whole data model, and it is why the machinery above the data — bands, joins, detectors, ledgers — never changes when a dataset is added. A new dataset is a harvester plus definitions, nothing more.

The subjects are institutions, places and documents. The place grid is the commune — the smallest unit the system knows, by law, forever.

02 · The receipt

Receipt or it does not exist

Every row carries the URL it was fetched from, the timestamp of the fetch, and a cryptographic hash of the fetched content. These are not conventions; they are column constraints. The database refuses a row without them, so a claim without a receipt is not merely against policy — it is unrepresentable. Anyone holding a row can fetch the source, hash it, and compare.

The same discipline reaches the surface: every figure rendered on this site opens into its receipt — the view it was read from, the moment of the read, and the verbatim request that produced it, replayable by any reader through the same public door.

03 · The harvest

The harvest never calls a model

Harvesters fetch from the official registers, parse deterministically, and store verbatim. There is no language model anywhere in the ingest path — not for extraction, not for cleanup, not for summarization. Every stored string came off a register exactly as the register published it. A model in that path could silently rewrite the record, and a record that might have been rewritten cannot be audited; so the model is simply not given the pen.

The model has four jobs, all of them downstream of the record and each of them checked: parsing documents into structure with verbatim quotes, compiling a reader's question into register joins, explaining nominated residue in plain language with a second pass checking the first, and investigating nominations with every sentence held against a receipt. When you ask this system a question, the question compiles into named filters and joins — and the joins answer. Never the model.

04 · Three states

A read has three states

Rows are the record. Empty means the harvest holds nothing, and the surface says so plainly. A failed read says READ FAILED, and the surface stays silent about the value — never a zero, never a stale figure, never a spinner pretending a value is on its way. The pages of this site obey the same rule: when the live read behind a figure fails, the slot says so instead of showing you something soothing.

The distinction matters because a catch that returns empty is a false all-clear. A system that renders zero when its pipe breaks will one day tell you a commune filed no decisions when the truth is that nobody looked.

05 · Detectors

Detectors are versioned like law

The patterns the system watches for are computed by detectors, and detectors are versioned the way statutes are: a change is a new version with a dated ruling, never a silent rewrite. Every claim a detector makes registers the version that made it, in falsifiable form, with a check date. Matured claims are rechecked against fresh harvest; the precision of every version publishes with the misses included, and every matured miss is classified and joins that detector's golden test set. A proposed refinement runs in shadow beside the standing version on the same matured claims and is promoted by receipted comparison, never by taste.

The claims ledger opens with the first claim, in public, and every version's curve stays on the record permanently. A curve that starts modestly and climbs is a better trust story than a flat impressive number, because it is visibly real.

06 · The gate

The gate

One country at a time, one layer at a time. No second country and no next dataset until the layer below is green on Denmark — green meaning harvested, receipted, checked, and covered by the acceptance tests. This is why the coverage line on every page says Denmark and says it at full size. A system whose marketing outruns its harvest has already told its first lie, and this one does not intend to.

07 · The refusals

What the machine never does

It never models a person. No table in the spine represents a human being, no join can build one, and where a register prints a name it renders exactly as published, judgment-free. It never prints a composite: no score, index, grade or rank of places — the verdict happens in the reader. It never invents a forecast: the forward field holds the institutions' own projections, their track record, labelled arithmetic, and the documented history of peers that hit the same wall — nothing else. It never sells a finding or its removal: the subject of a record reads the same page as everyone. And it never decorates: no invented precision, no illustrative numbers, no figure anywhere on the surface that was not read from the spine at render time.

The full list, with the reasoning each rule was paid for with, is on the laws page.