Core concepts

Thisness separates the steps that most matching tools fold into one score. Each step leaves its own record, so a conclusion can always be traced back to what supports it.

Observations are not entities

An observation is one source record: a row in a vendor file, a CRM account, a registry entry. It carries its attributes and a source reference (type, id, locator), and it never changes. Observations of one kind of thing share an entity type such as organization, person or taxon.

An entity is a stable identity above observations. Its ID never changes when its name does, and its state is active, provisional (created from too little evidence to trust yet) or retired (merged into another). A membership says an observation belongs to an entity, from one ledger sequence number until another.

Similarity is not identity

Resolution runs in stages, each with its own records:

1. Candidates. Blocking keys (a normalized name key, a website domain, a registration number, a postcode) propose which pairs are worth comparing. Two records that share no key are never compared directly.
2. Evidence. Each comparison of an attribute yields a level (exact_key, token_typo, number_differs, ...) with a direction: it supports a match, contradicts it, or says nothing.
3. Pair decisions. Fellegi-Sunter weights, fitted per store by EM, turn the levels into a weight in bits. A pair is a MATCH, POSSIBLE_MATCH, NO_MATCH or REQUIRES_REVIEW. Scores are typed, so there is no universal confidence float.
4. Consensus. Strong pairs form components. A component is committed only if it holds together: no contradiction inside, nothing contested at its edge. Otherwise it goes to review.

Identity is not relationship

"Subsidiary of", "brand of", "same household as" are relationships between distinct entities, never merges. Some comparison levels are relationship signals: two companies whose names differ by a qualifier ("Haleon UK Finance" and "Haleon UK Capital") are capped at 0 bits and listed as possible relationships instead.

Everything is reversible

Every change to identity is a commit: a set of ledger events (ENTITY_CREATED, MEMBERSHIP_ACCEPTED, ENTITY_MERGED, ENTITY_SPLIT, RELATIONSHIP_CREATED, ...) that cites the decisions behind it. Nothing is destructively collapsed. A merge retires one entity into another and keeps its history; a split moves observations out into a new entity. The ledger alone can rebuild current state, and thisness check verifies that it does.

Models propose; policy decides

Rules, Fellegi-Sunter, language models and people all contribute decisions. Only a commit, citing those decisions, changes identity. A model's answer is one more decision with its own record: which model, what it was sent, what it said and what it cost.

Constraints record what a person has settled: a cannot_link between two observations holds against any later evidence, and a must_link keeps them together.

Reviews

When consensus cannot settle a cluster, it opens a review:

Reviews are ranked by impact. An answer is same, different, related (with a relation), or groups (a regrouping of the records), and it is stored with its reason. A review assistant can suggest answers, which a person accepts or overrides. See The Workbench.

Dossiers

A dossier is everything Thisness believes about one entity, assembled from stored records: each attribute's values and how many observations support each, contradictions, members, relationships, open reviews and lineage. thisness dossier ID prints it as context for a language model, within a character budget, withholding attributes a profile marks as local only.