Glossary & Shared Vocabulary

Investigations keep running into the same things written differently: a register writes GES, a contract says Gesellschaft mit beschränkter Haftung, an analyst types GmbH. The glossary is Classifyre’s shared vocabulary. It says once what things are, and the semantic layer built on it lets every finding, asset and case carry that meaning — so a question about Personal data or Bank account reaches every piece of evidence, whatever detector or language it came from.

The meaning map

Entities & glossary opens on the map: every approved concept as a card, grouped by the scheme it belongs to, with the evidence behind it. It answers what is our data about at a glance, and it is where the glossary grows.

• A frame per scheme. Inside a frame, concepts read as an outline: a broader concept on top, its narrower concepts indented beneath it, the biggest first. A long scheme flows into several columns so the whole map fits the screen.
• A card per concept. The number is how many assets are about the concept (narrower concepts included). The bar under it compares that with the biggest concept on the map, and its colours are the severity of the findings behind it. An amber number counts proposals waiting for review. Zoomed out, a card shows only its name, larger.
• Lines. Solid lines join a broader concept to its narrower ones, dashed lines are related to, and a labelled line carries a verb of your own. Dotted lines join concepts found on the same assets more often than chance would put them together.
• Taxonomy or Network. Taxonomy is the outline described above. Network lets concepts that turn up together pull towards each other, while each scheme keeps its own neighbourhood. You can drag cards to untangle it.

Point at a concept and everything connected to it lights up, with the number of shared assets on each found together line. Click it to open the panel on the right:

From the panel: Open concept (its page), Add to case…, Watch (a new watch about the concept), Evidence, Copy link (the map, opened on this concept) and Review proposals when there are any.

Growing the glossary on the map

• + on a scheme’s frame adds a concept to that scheme. A new scheme shows an empty frame with Add the first concept.
• Edit changes the selected concept: name, definition, scheme, aliases and codes.
• Narrower adds a concept underneath the selected one.
• Relate… then a click on another concept: choose is a kind of (either way round), is related to, or a verb in your own words. The new line appears at once.
• Point at a relation in the panel and × removes it.

New concepts are approved straight away, so they appear on the map before any data is linked to them (with an empty bar).

Bringing a concept into a case

Add to case… picks an open case or starts a new one, then:

• marks the case as about the concept, so the concept lists the case and the case’s Meaning panel lists the concept;
• pins the concept’s card on the case board;
• optionally opens a new watch about the concept, already linked to the case, so new findings about it reach the case by themselves.

Filters and the case highlight

Narrow the map by scheme, by source (counts are recomputed for those sources only) or by a minimum number of assets. Highlight a case rings the concepts its evidence and its own about links touch, and dims the rest.

Ways in

• From a concept page: Show on map.
• From a scheme (Schemes tab): Show on map, filtered to that scheme.
• From a case: Case file → Meaning → On the map, with the case highlighted.
• From Explore, which draws the concepts two documents share as labelled boxes between them; select one and Open in the glossary.
• A copied link opens the same view: layout, filters, highlighted case and selected concept.

The counts on the map come from a summary that is rebuilt automatically after scans and glossary changes (and on demand with Rebuild in the map’s legend). Two concepts count as found together once they share enough assets — twenty by default. A small workspace can lower that with the cooccurrenceMinSupport setting of the suggestions settings API.

Concepts and entities

Every term has a kind:

• A concept is a kind of thing: GmbH, IBAN, Force majeure, Eigenmittelquote. Concepts have a definition and usually belong to a scheme.
• An entity is one particular thing: ACME Holding GmbH, Jane Doe, a project codename. Entities keep the entity types you know — person, organisation, location, reference.

Each term has a stable key (gmbh, bank-account-iban). Keys are what connectors, bindings, filters and agents refer to, so renaming a term never breaks anything: an old key keeps resolving to the term it named.

Besides its name, a term has three kinds of labels:

Single letters and short abbreviations belong in hidden aliases — as an alias, E would match half the corpus.

Schemes and relations

A scheme is a controlled vocabulary: the legal forms of a register, the GDPR categories of personal data. Concepts within it form a tree through broader relations — GmbH is a kind of Kapitalgesellschaft, which is a kind of Gesellschaft. Other relations say a concept is related to, part of or an instance of another, or carry a verb of your own. Cycles are refused, with the path that would close them.

Evidence rolls up the tree: asking about Kapitalgesellschaft with include narrower reaches every GmbH and AG too.

Draft, approved, deprecated

Terms, relations and bindings move through the same lifecycle. Only approved items drive meaning: a draft term can be looked up and discussed, but nothing links to it until it is approved. Deprecating a term (optionally naming its replacement) stops new links while keeping history.

Terms you add are approved at once. Terms proposed by AI agents, imports marked as drafts and suggestions from the system start as drafts and wait in the review queue.

Entities: from mentions to things

An entity is more than a glossary entry: it is a thing with a name, aliases and identifiers, and every finding that carries one of them is a mention. ACME Holding GmbH in an e-mail, the register record for ACME and the VAT number ATU12345678 on an invoice become one thing with one page.

Mentions are never stored per finding. An entity keeps a list of confirmed values — its name and aliases under every name label of its type, plus the identifiers you or a source gave it — and a mention is any finding whose normalised value is on that list. A value confirmed today therefore shows the documents scanned last year, and a new scan adds its matches without a resolution step.

• Exact links need no review. Names, aliases and identifiers link on their own, once the entity is approved. A draft entity lists its mentions so you can judge it, but links nothing.
• Spelling variants are candidates. After each scan, name-like values that sound or fold to the same spelling as an entity’s name (Acme Holding G.m.b.H.) are proposed with a score. Accept one and it becomes a mention; reject it and it is never proposed again. Candidates never confirm themselves and entities never merge on their own.
• An identifier has one owner. An IBAN, VAT or national ID confirmed for one entity cannot be confirmed for a second without a conflict review: take it from the first holder or reject it. A name may belong to several entities; the page flags it as ambiguous.
• Only identifier-like and name-like labels take part. Tag, date and free-text labels never do, so a common word cannot turn every document into a mention. Labels from your own detectors are declared as names or identifiers under Entities → Labels that take part.
• Merging moves values, references, relations, graph edges and watches from one entity to another. The first redirects to the second and a source that re-declares it on the next scan lands on the survivor. Splitting is not supported.

Where entities appear

• Make entity on a finding, in Where else found, in the shared-value panel of the Fingerprints graph, and on a case board’s details.
• The entity page: mentions over time, sources and assets, the most co-mentioned entities (computed on demand), the anchor record, cases, watches and pending candidates. Export mentions (CSV or JSON) answers “what do we hold about this person”.
• Review queue (Entities & glossary → Proposals): candidates and conflicts, best first, each with up to three occurrences.
• Watches on an entity (“anything new about ACME”) and case leads: a case about an entity suggests findings that newly mention it.
• Agents and MCP can search, create and update entities and propose verdicts. Only a person confirms an identifier conflict, accepts a candidate or merges.

Sources that declare entities

A register or a customer master does not merely mention a company: its record is the company. A custom connector says so with Asset(entity=Entity(...)), see custom connectors. Declared entities are approved from the start, anchored to the record and idempotent by its URN; what an operator later edits is kept.

The Entities switch

Entities are one of the three feature switches in Settings → Cleanup. They share the value index with duplicate detection, so values stay indexed while either is on. A new workspace has entities on; an existing one gets the switch off and turns it on to build values and counters from what is already there.

Giving data a meaning

Bindings

A binding says once what a detector output or metadata field means. It is a choice from the vocabulary your data actually produces — never a pattern:

• Every finding of this kind — IBAN_CODE from the PII detector means Bank account (IBAN).
• Only these values — the tag risk means Financial distress only when its value is hoch.
• Each value names a term — the tag legal_form holds register codes; each value is looked up among the codes and labels of the Rechtsformen scheme, so GES means GmbH and AG means Aktiengesellschaft.

The What does this mean? dialog opens from a finding, a term, or the Unbound vocabulary tab, picks a sensible default (a lookup when the values mostly name concepts of one scheme), and previews how many findings, assets and sources it would give meaning before you save. Saving needs no re-scan: links follow within moments. Mark an output No meaning (say, DATE_TIME) to take it off the worklist on purpose.

The Unbound vocabulary tab lists every output and field with open findings and no meaning yet, with the semantic coverage of your workspace on top.

Find in text

Some concepts never appear as a detector output — a PDF just says Eigenmittelquote. Find in text builds a tested REGEX detector from the concept’s labels (word boundaries, codes matched exactly) and binds its output to the concept.

Connector declarations

Connectors and augmentation notebooks can say what a record is with means(Ref.asset(...), Ref.term("gmbh")) or asset.means("gmbh") — see custom connectors. A declaration to a key nobody has defined yet is kept and shows up under Unknown references in the review queue; it links the moment the term exists.

Linking by hand

From a finding, an asset or a case, Link to a term records that it is about a concept. On a case board, drawing a link from evidence to a pinned concept does the same.

Starter packs

Starter packs install schemes, concepts and bindings for common needs — GDPR personal data, secrets and credentials, financial identifiers, DACH company law. Preview first; nothing you curated is overwritten.

Meaning in daily work

• Finding pages show a Meaning card: which concepts the evidence is about, through which binding or link, and what follows up the taxonomy. On an asset page the same list is in the Connections tab: select the asset on the canvas and its Details show its concepts by method, with Link by hand; the concepts it shares with other documents are drawn on the canvas.
• Term pages open on Where it is connected: the assets that carry the concept or entity most, the term itself between them, the other concepts they share, and what those bring in. Double-click any document to dive into everything it is connected to (Explore). The tabs list the evidence (by method, support and severity), the bindings, relations and history.
• Filters and watches accept glossary terms: search findings or assets by concept, and build an inquiry watch about a concept, with or without its narrower concepts.
• Exports carry a terms column.
• The meaning map (see above) is the glossary’s main view.
• Explore draws the concepts and entities two documents share as labelled boxes between them, in its By meaning view.
• Case boards list the concepts the case’s evidence is about, can pin them as cards, and open the map with the case highlighted.

The review queue

Everything that would change meaning without you deciding lands in one queue: agent drafts, proposed aliases, suggested bindings for unbound vocabulary, suggested document links (with embeddings on), suggested relations, and unknown term references. Accept, dismiss with a reason, or dismiss forever — a dismissed suggestion comes back only if its evidence changes materially.

Agents

Autopilot agents read the glossary, its vocabulary and meaning, and propose terms, relations, bindings and document links. Within guardrails you set in Settings → Autopilot, they may approve relations and bindings they proposed: only between approved terms, never against an operator’s binding, below an impact limit, within a daily budget, and only after previewing exactly what they approve. Every agent approval is in the undo log. Agents never decide on documents, terms or aliases — those stay yours.

Import, export and transfer

Import and export CSV or SKOS (JSON-LD), with a dry run first. Workspace archives carry schemes, terms, relations, bindings, manual links and your review decisions; derived data — the vocabulary inventory, semantic links, pending suggestions and the map — is rebuilt after an import. In Settings → Cleanup the curated glossary is protected, and the derived datasets can be cleaned and rebuild themselves.

Where this leads

The glossary is consumed everywhere findings, connections and cases are worked, and it is one of the things Autopilot agents can propose additions to, always subject to your review.