How the record is kept.
Two things together: what canopticon is trying to do, and how it checks what it puts on a page. A claim from this site is trustworthy because both are visible, and both can be checked.
The standard
Every documented fact rests on a real source. Argument is kept separate from fact — an essay argues, a document is the evidence. Confidence is shown, not hidden — an uncertain claim says so. Nothing is stated that can't be traced back to a document. The record is permanent and correctable: mistakes are logged, not erased.
What's in scope isn't a topic list — it's a question: is more said about this than has ever actually been checked? Wherever that's true, it applies, whether the subject is a politician's record, whether a therapy actually works, or whether a species is really recovering.
Checked by machine, not vouched for.
A claim on this site is trusted because of what happened to it before it was published, not because of who wrote it. Every documented fact runs a chain of checks:
A source document is stored first, exactly as received. Facts are then read out of it, each one carrying the exact quote it came from.
Every quote on every page is a literal match against the stored document — checked by machine, character for character, at the moment the page is built. Whitespace differences are tolerated; word substitutions are not. A quote that fails this check never reaches a page.
Wherever this site places a passage on a page — a lead passage, a finding, an answer to a question the record raised — it is selected from the document, never written for it. A model may propose a span; the match in check 2 decides. A proposed span that is not found word for word is discarded, and the page shows how many were.
A span that fails the match is removed before it is shown — never silently kept. Where the count matters, the page says how many were discarded, so a shaky answer never looks identical to a solid one.
Every receipt on every page opens to the verbatim text, the source document, and when it was read. Claims on this site are deep-linkable to their evidence.
What happens to a document
Uploading is publishing. The document is live at its own address at once, whole. Then the record reads it, one step at a time, while the operator's library page is open — nothing runs unattended, and nothing here calls a model when a page is rendered: whatever a model proposed was checked and frozen into a row first.
- Title — a filename or a working title is replaced with a line lifted verbatim from the document; the original is kept.
- Read — facts, events, entities and quotes are read out; every quote is checked word for word before it is stored.
- Figures — the numbers with a named source and a date, each checked the same way.
- Passage — one lead passage, selected from the document, never written for it.
- Section — filed under one of eight sections; "everything else" is a real answer.
- Card and index — a share card is drawn, and the verified passages are indexed so search can return them.
A draft essay may be opened from a reading and written by a model. It is never published as such: it is handed to the operator as a file whose first line names the model and the reading it came from, to be rewritten and uploaded under the operator's own name — or not.
Disputed stays disputed
A claim someone made ("X said Y") is never silently shown as an established fact. Disputed material is stored as disputed and shown as disputed. An essay published here is held to the same rule: the facts that complicate its own case are shown, not left out — it can't leave out what doesn't fit.
The stance
Forensic, not journalistic. The goal isn't to report what happened but to map what is — the structure beneath the event, the pattern beneath the story, the architecture beneath the policy. Individual facts matter because they reveal structure. Structure is the finding.
What's actually in the record
Two kinds of thing are published, and everything else is derived from them:
- Research documents — primary material, kept whole and published as itself. Its own summary and its own stated gaps are moved to the top; nothing is summarised away.
- Essays — arguments written to a fixed convention: a question per section, the answer in the first sentence, sources numbered at the end.
- Verified quotes — every span read out of a document, checked as a literal match against the stored text. Each has a receipt at its own address.
- Findings — a figure with what it measures, when, and who reported it. The best-evidenced shape a document can carry, and the one this site is built to be cited for.
- Links between documents — a subject appearing in two documents is a link from one to the other. Found by a query over the record, never composed.
- Leads — where the record is strange: two documents that contradict each other, a figure no document sources, a question the documents raise and do not answer. A lead is a question for the operator, not a page for a reader.
- Bridges — a subject that appears in two documents filed under different sections. Found by a query, never by a model, and never ranked by how often it recurs. The record cannot make the connection; it can put the two documents side by side and ask whether anyone has made it before.
- Readings — a short note the operator wrote on a passage of a document. The passage is verified like any quote; the note is the operator's own words, signed and dated. A document with a reading is marked read by the operator: the one thing the machine cannot supply.
- Lineage — what a document grew from and what grew from it. A brief researched from a question one document raised is that document's child, and says so.
Search asks the record, not a model: a question returns passages from documents on file, word for word, each with its receipt.
A page, its plain-text twin, its structured data, the sitemap and the feed are all generated from the same rows. They cannot drift apart because there is only one of them.