The ontology

The record is a small number of first-class objects with a typed graph between them. Sources are documents — a web page, a PDF, a scan, a book, a file — stored as immutable, content-addressed snapshots. A source is cut into chunks, ~1,200 characters on paragraph boundaries, each remembering its byte offsets in the snapshot. Claims are atomic statements extracted from chunks. Evidence ties a claim to an exact quote in a chunk, with the quote's offsets and whether the bytes were found. Entities are the people, organisations, places, events and things claims mention, resolved from the names sources use. Questions are the frontier — what the loop works. Patterns are what the slow loop finds in the whole record.

Everything else is derived from those and recomputed as they change: pages (spans of an uploaded document's snapshot with a digest each), surfaces (the record along one axis), trajectories (momentum and entropy of a thread), standing answers, the narrative, breaks, the ledger of what each stage cost. Notes — your margin — are the one table deliberately outside all of it: never evidence, never cited, never graded.

The graph is one table, links, where every edge has a from-object, a to-object and a kind. Which kinds exist, and what each means, is in the last section of this page.

An upload goes through the same ontology as a page the loop fetched; the only extra table is pages. The document's text becomes one snapshot, so evidence offsets are ordinary, and each page is a char span over it. Every stage is metered in the ledger (ocr · extract per chunk · summarize per page · document · advance), and the wall-clock of each is kept on the source.

Epistemic grade is about the statement: fact — the source states it as so; inference — the source reasons to it; hypothesis — the source proposes it; speculation — the source guesses. It is set at extraction from what the source licenses and enforced as a database constraint. It never climbs: the skeptic can downgrade, nothing upgrades.

Status is about the record: every claim starts provisional. It becomes corroborated when it is a fact, its quote verified, and its verified supporting sources span two or more independent lineages — copies of one account collapse into one (see Lineage). It becomes contested when evidence refutes it, or when none of its evidence verifies. Retracted is a person's ruling, and the loop never touches a retracted claim again. You can set any status yourself; your ruling outlives the worker's later pushes.

Evidence stance is how a quote bears on a claim: supports, refutes, or merely mentions. Corroboration counts only verified supporting evidence; one refuting quote is enough to contest.

The extractor does not name entities; it emits mention slugs — kebab-case strings like gabe-valdez or officer-gabe-valdez — and each becomes a mentions link from the chunk and the claim. The resolver then groups slugs into entity rows, computed, not written, by rules applied in order: one slug's tokens contained in another's (a bare single token must sit in the surname position — valdez joins officer-gabe-valdez, michael does not join every Michael in the corpus); an exact match once titles are stripped (dr, sheriff, senator…); a shared surname-like last token with compatible given names, nicknames included.

Your identity decisions constrain it: a pair ruled same is joined first; a pair ruled different may never share a group, whatever a rule would do. The canonical name is the fullest slug in the group; the kind is inferred from its last token — organisation suffixes, place suffixes, a person if it carried a title or is two or three plain tokens, else unknown. Entity ids follow the canonical name, so re-resolution every cycle reproduces the same ids and links: no growth, no reshuffling. Junk slugs — archive and library chrome, domain fragments — never become entities.

Places are geocoded; entities with a match are given a Wikidata id, which later serves as an identifier for identity questions. Two entities that might be one person are not merged by the system: they become an identity pattern with the evidence for and against, and nothing merges until you rule.

A relation is never written straight into the graph. It is structure on a claim — so its quote is verified like any other, corroboration counts independent lineages, a contradiction can fire on two disagreeing amounts, and a person can retract it. The extractor adds one only when the text states it (an attempt, an offer, a denied ownership is not a relation), with the text's own amount and currency, never converted, and the year the text gives.

Direction is where an extractor goes wrong, and a confident wrong arrow is worse than no edge: the skeptic reads the quote and confirms, corrects or drops the arrow — bounded per source. Passive spellings are folded so a fact has exactly one directed edge: B was paid by A is pays(A, B).

After resolution, relations are materialised as entity → entity links, one per (from, to, kind), whose weight is the number of independent lineages behind it — not the number of claims. Following the money from one entity walks those edges and names the leads: a sink (money in, none out), a spring (money out, none in), an edge resting on a single lineage, an edge backed by a contested claim.

Publication dates are inferred strongest-first: the page's own metadata, a date in the URL path, a date in the title, then text anchored to publication language — a byline, a dateline, a copyright line. A bare year in the body is never accepted. A date you set by hand outlives the worker's pushes. Rights default closed: every source is copyrighted unless it comes from a .gov or .mil host (government) or is a fixture; a copyrighted source's text never leaves the investigation in a package — the claims travel with the content hash.

Kinship is textual: two sources sharing enough distinctive six-word shingles (three, or twelve on the same site, where template chrome is shared) get an undirected kinship link weighted by how much they share. Kin at or above 0.5 fold into one lineage everywhere the system counts independent sources — the skeptic, corroboration, relation weights, breaks, the answers' ladder. Five publications repeating one wire story are one lineage.

Direction is earned, not guessed. A citation — source A naming source B's title verbatim, or its author with its year — is a directed cites link. A derivation is only written when strong kinship, date order and a judge's confirmation agree; similarity plus chronology alone never earns an arrow. When the corpus has many tellings of one account, drift measures whether the wording strengthened over the years while no new lineage arrived beneath it.

The store is the terrain — claims, evidence, entities, dated facts, documents. A surface is a reading of that terrain along one axis, made of cells, each carrying its claim mass (density = its claims over all claims), its verified share (coverage), its sources, its status mix, and the claims that sit there. Four axes are built in, and each also has a summary cell — placement rate, cell count, top cells, gap count.

How things land on a surface: the time axis places a claim in the year its text carries; the actor axis has one cell per person or organisation entity, and a claim lands there through its resolved mentions; the place axis has one cell per place entity, geocoded or not; the document axis has one cell per source. Every entity of the right kind gets a cell even with zero claims, because an empty cell is a gap too. Surfaces are recomputed every cycle and after every upload, never edited.

Gaps are read off the surfaces by fixed rules and emitted as gap patterns, so the slow loop can turn them into frontier questions. A surface gap that later closes is a break.

A trajectory is a thread with a clock. Two kinds: a question thread — a frontier question with its lineage of child questions, the sources found through them, and the claims those sources evidence — and an entity thread — an actor, place or thing and every claim naming it. Each cycle a thread gains mass: 1.0 per verified claim, 0.4 per unverified, ×1.5 when corroborated, ×0.7 when contested. Velocity is a moving average of mass gained, acceleration its change, and momentum = velocity · √mass · (1 + ¼ ln sources). Coherence is the cosine between successive cycles' centroids in embedding space.

Two entropies say what a thread is made of: over its sources (low = a monoculture) and over the entities it names (low = focused). Per claim, surprise is how improbable its entity combination is under the corpus's co-occurrence; novelty is one minus its nearest neighbour; a bridge is a claim joining two well-established entities that almost never co-occur.

The survey is mechanical — computed, not written — over the whole corpus, and every pattern has a stable id, so re-surveying updates strength and support rather than minting twins, and a pattern you dismissed is never resurrected. A pattern a live question was born of is kept even when it drops out of the cut. Then interrogation hands the strongest forty fresh patterns — not raw claims — to the director, one scale up, and the questions it proposes land on the frontier at priority 0.85, linked born_of the pattern.

Corroboration is the survey's mirror image and writes links, not patterns: claims from different lineages sharing an entity and a compatible attribute value, passing the same-fact judge, get a corroborates link, and a provisional claim so confirmed is promoted.

A question's origin sets its priority: the primary you wrote, 1.0; a secondary, 0.95; one born of a pattern, 0.85; a routine decomposition of a bigger question, 0.7; a gap question about an entity nobody has asked about, 0.6. Each cycle takes the highest open question, older first on a tie — unless the last one worked was pattern-born, in which case the next pick is forced to be something else, so emergence and coverage interleave. Statuses: open, active (being worked), closed (the director closed it, with a reason), dead (you killed it).

The loop stops when it has spent its cycles, when the frontier is empty, or when — for a window of consecutive cycles — new claims per source and novelty of new text both stay below their floors. Measured, not felt.

A standing answer is what the record says about each question, in writing, composed strictly from retrieved claims with inline citations, and graded by rule: its best cited rung on the evidence ladder, and a confidence — contested if a cited claim is contested and fewer than two corroborated ones outweigh it; answered if the best rung is corroborated or better; partial if verified; lead otherwise; none if nothing is cited. The cited basis is hashed: an unchanged basis means an untouched answer and no cost; a changed one means a new version with its delta in words.

A break is a moment the investigation's knowledge actually changed, detected deterministically after every cycle, survey and upload from the tables above. Kinds: independent (a claim now spans two lineages), echo (looks corroborated, is one lineage), contradiction, resolved, question_closed, load_bearing (an entity the graph's shape depends on), structure (a cluster more connected than chance, checked two ways), filled (a surface gap closed), converged and stalled, and the answer kinds first_answer, answered, answer_contested. At most three per pass are promoted to the tier that interrupts — notified, first on the Overview — and only at importance five or more; the rest go to the case log.

The narrative is an outline the director plans (re-planned when enough new material has arrived) and sections the writer composes strictly from in-scope claims, every factual sentence carrying its claim marker; three sections — open questions, contested, sources — are built from the store, computed, not written. A section's basis is hashed like an answer's: unchanged, untouched; changed, rewritten as a new version with the old one kept.

Notes are your margin — on any object, with an anchor to a char range, a time range or a point — and the one table with no epistemic ladder. The ledger records what every document and process cost: tokens and calls per stage, with units, so the cost of an OCR page or an extracted chunk falls out.

Everything relational is one table. Reading it is reading the investigation's structure.