Method · v1 · 2026-08-29

How a document
becomes a record

This register holds 20,989 records, of which 3,688 carry a page of their own. Underneath them are 19,645 sourced facts drawn from 40,396 captured documents. This page is how that happens, what the rules are, and where each one is enforced.

None of it is aspirational. Every rule below is a database constraint, a threshold script or a versioned method identifier, and each is named so it can be checked rather than believed.

01 · The unit of work

The document, not the person

Work does not begin by deciding someone is worth recording. It begins with a document — a programme, a licence roll, a venue calendar, a newspaper listing, a photograph of a flyer — and the document is decomposed into whatever it happens to contain. A person who appears once, in a lineup, on page eleven, becomes a real record at that moment, with a source, because the page said so.

That ordering is what keeps the register from being a list of people somebody thought were important. It is also what makes coverage legible: a gap here is a document nobody has captured yet, and gaps of that kind can be closed by going and looking.

What a capture carries

  • The original URL
    Where the document lived when it was found.
  • A Wayback snapshot, taken as raw bytes
    The file as it was served, not the archive’s later rendering of it. Those differ, and only one of them is the document.
  • A sha256 of those bytes
    A web capture without a hash cannot back a fact. That is a database CHECK, not a habit.
  • A second, independent submission
    Filed to the archive at capture time, so the register is not the only party attesting the document existed.
  • The date the bytes were pulled
    The retrieved half of a citation. A capture is a moment, and the register says which one.
  • A local copy on disk
    So a reading can be re-checked against the exact file it came from, years later, if the host and the archive are both gone.
02 · Evidence lanes

Three lanes, kept apart

Ordered by how much weight a reader should put on the row.

The central distinction in the register. What a document said, what a subject said about themselves, and what we decided something is are three different kinds of statement, and they are stored, coloured and counted separately so that none of them can borrow another’s authority.

Fact lane

A document said so, and the document is on file.

Every fact-lane credit carries a pointer to the source it rests on and a locator inside that source — a page number, a lot number, a line. The database refuses the write without one. This is the only lane whose rows a reader should treat as documented.

Claim lane

The subject said so. We have not verified it.

A biography, an artist statement, a résumé line. Recorded because it is real evidence of how someone describes their own practice, and separated because self-description is not documentation. Abstract works take claim-lane kinds only.

Classification lane

We decided so, against the class tree.

That a room is a comedy club, that a practice is barbering. Editorial by construction — it is the register reading the evidence rather than a document speaking — so it is never counted as fact. A record whose only credits are classifications holds an opinion and nothing underneath, and does not earn a page.

03 · Verification

Every fact records how it was verified

Ordered by identifier.

Not a “verified” flag. Each credit carries a versioned method identifier saying what kind of act the verification was — so that if a method later proves unsound, every fact resting on it can be found and retracted as a set. A boolean cannot be recalled, because nothing records what it meant.

  • ADC-VS-1The document’s own words

    Read from a capture. The page states the fact and the capture is hashed, so the reading can be re-checked against the exact bytes it came from.

  • ADC-VS-4Editorial

    The register’s reading, not a document’s claim. "This room’s descriptor is Comedy club, therefore the acts booked in it are comedians" is a VS-4 inference — reasonable, and not something anybody wrote down.

  • ADC-VS-5Transcribed from an image

    A flyer, a poster, a programme page. Filed with the file’s sha256, so the transcription and the artefact stay attached.

  • ADC-VS-6Inferred from evidence inside the document

    Where the document does not state the value outright. Comedy clubs sell a weekend rather than a night — "7/24, 7/25, 7/26" with the year nowhere on the page — so the year is anchored on the artwork’s own timestamp inside the page. Measured across 352 such dates: none fell before its anchor, median 39 days after, 95th percentile 78. Good evidence, and still not the document speaking, which is exactly why it is not filed as VS-1.

04 · What the database refuses

Six rules that are not policy

Ordered as they are numbered in the schema.

These are enforced as database constraints rather than as application rules, which means they hold when the console is bypassed, when a script is wrong, and when somebody writes a row by hand. The most consequential is the fifth: a fact-lane write with no source pointer does not fail review, it fails to commit.

  1. 01

    Maker-fact credits are invalid against abstract works.

    credits_abstract_takes_claim_kinds_only

  2. 02

    Recognition binds to an edition, never to a series.

    credits_recognition_binds_edition

  3. 03

    Credentials bind to an organization with a conferral date.

    credits_credential_binds_dated_org

  4. 04

    Intervals are never overwritten — a second interval inserts rather than collides.

    structural: no unique key on (agent, target, kind, role)

  5. 05

    No source pointer, no commit on a fact-lane write.

    credits_fact_requires_source, edges_requires_source

  6. 06

    No rights class, no render.

    media_assets.rights NOT NULL + generated render_posture

Enforced in lib/db/src/schema/.

05 · What earns a URL

The threshold, published

20,989 records exist; 3,688 have a page. The difference is a scored threshold, and it is published here because a bar nobody can see looks like a judgement about who matters. It is not one. Everything below it still exists, still renders and still gets read — as a row on the hub it belongs to, with its facts and its sources intact.

The bar is about whether a page would answer a question somebody arrived with. Fewer pages that answer one are worth more than many that do not, and the thin tail is not neutral: a page carrying a single unsourced fragment dilutes the register it sits in.

RecordClears at
A person or an organization
3 sourced fact-lane credits, across 2 distinct neighbours in the graph

Facts are counted in both directions — a venue’s evidence points inward, an act’s points outward. Counting one direction only left act-organizations looking empty when they were not.

An edition
2 sourced facts, across 3 neighbours

Two rather than three because the source is a YEAR | WINNER | 1ST ALTERNATE table: the state preliminaries publish the top two and nothing more. Two is not a thin reading of that document — two is the document.

A series
2 editions beneath it

A hub earns its status by having rows to land. The calendar reader mints a series for every listing it sees, and "Labor Day Party" is not a series; it is a Monday.

A building
2 of 3 evidence lanes — a place, a stated type, or 3+ photographs

The credit bar is unreachable here by construction: every building carries exactly one fact-lane credit, the firm that designed it, because the record is minted from the firm’s own project list. Authorship is the definition of the row, not evidence of depth.

A class or a place
Exempt — never scored

These are the destinations demoted rows land on. Scoring them demotes the landing site: a class page whose whole value is its roster would be demoted to a row on itself.

Derived by scripts/src/registry/derive-indexable.cjs, re-read on every run.

06 · Permanence

Identifiers outlive the names on them

Every record carries a permanent identifier — ADC-P-000482 — and its URL carries the number as well as the name. The number is the half the register controls. A person changes the name they work under, a room is renamed, a reading is corrected: the slug changes and the number does not, so the old URL redirects permanently instead of breaking.

Two records that turn out to be one are merged, and a merge is not a deletion. The retired identifier keeps answering, and it says which record it became, when, and why. Where the successor has a page the old URL redirects to it; where it does not, the old URL renders the explanation rather than redirecting into a dead end.

A page the register publishes and then withdraws is a third case, and it answers 410 Gone — not 404, which would say the page might return, and not a redirect, which would say something else stands in its place. The reason is stated on the page, dated.

07 · Being wrong

Corrections are appended, never silent

A register that edits itself quietly is worth less than one that is occasionally wrong in public. Corrections are appended to the record and dated; the previous state is not overwritten. Every record shows when it was created, when it was last verified and which version you are reading, and every page is citable at that version.

The versioned method identifiers are the same idea at a different scale. Because each fact records how it was verified rather than merely that it was, an entire class of facts can be retracted at once if the method behind them turns out to be unsound — which is the capability a “verified” checkbox quietly gives up.

08 · What this does not do

The limits are part of the method

  • Absence is not evidence. A person with no record here was not judged and found wanting; more likely nobody has captured a document that names them yet. Coverage is stated per place, as a survey status — not surveyed, partial, substantial or complete — precisely so that a thin page cannot be mistaken for a thin subject.

  • Classification is ours. Where the register says a practice is barbering or a room is a comedy club, that is an editorial reading against the class tree, and it is filed in its own lane so it can never be counted as something a document said.

  • A capture is a moment, not a fact about the present. Every citation carries the date the bytes were pulled. A venue that was open when its page was captured may be closed today, and the register says when it looked rather than implying it still holds.

  • Where two documents disagree, both entries stand and the disagreement is visible on the page. Reconciling them would mean asserting something neither document says.

  • The register documents practice; it does not appraise it. There is no rating, no ranking, and no editorial claim that one body of work is better than another.

Counted at export · 2026-08-29

20,989 records · 3,688 with pages · 19,645 sourced facts · 4,280 classifications · 40,396 documents · 7,399 citations.