Open source · Apache-2.0 · Python 3.12+

Identity resolution for
imperfect observations

Thisness decides which records refer to the same thing, and keeps the evidence, the decision and the history behind every conclusion, so any of them can be explained, reviewed or reversed.

Thisness WorkbenchOpen the explorer

Are these the same entity, different entities, or related ones?

6 records in 2 entities today. In the Workbench you drag cards between columns to regroup; every answer is stored with its reason and can be undone.

Suggestiondifferentgpt-6-luna

The records form two internally consistent groups: “THE GYM LIMITED”/“Gym” and “THE GYM GROUP PLC” (C1–C6). The shared postal location and partial name overlap provide only weak cross-group support, and the distinct registered names do not establish that these are the same organization or that they have a specific relationship (C13–C15).

R1, R2, R3R4, R5, R6its grouping

GGymactive3

THE GYM LIMITEDR1

2ND FLOOR, ARDING & HOBBS, 7 ST. JOHN'S ROAD, LONDON, SW11 1QN

companies_house:rec000965

GymR2

2ND FLOOR, ARDING & HOBBS, LONDON, SW11 1QN

accounts_payable:rec000966

GymR3

LONDON, SW11 1QN

crm:rec000967

GTHE GYM GROUP PLCactive3

THE GYM GROUP PLCR4

2ND FLOOR, ARDING & HOBBS, 7 ST. JOHN'S ROAD, LONDON, SW11 1QN

companies_house:rec001361

THE GYM GROUP PLCR5

2ND FLOOR, ARDING & HOBBS, LONDON, SW11 1QN

gleif:rec001362

The Gym Group PlcR6

LONDON, SW111QN

accounts_payable:rec001363

Comparisons, strongest first

Records

Evidence

Weight

Verdict

C1

R4 ↔ R5

name exact key+10.2location postal+8legal form same+0.6

18.8

match

C2

R4 ↔ R6

name exact key+10.2location postal+8legal form same+0.6

18.8

match

C3

R5 ↔ R6

name exact key+10.2location postal+8legal form same+0.6

18.8

match

C4

R1 ↔ R2

name exact key+10.2location postal+8

18.2

match

C5

R1 ↔ R3

name exact key+10.2location postal+8

18.2

match

C6

R2 ↔ R3

name exact key+10.2location postal+8

18.2

match

C7

R2 ↔ R4

name token subset0location postal+8

8

possible

C8

R2 ↔ R5

name token subset0location postal+8

8

possible

C9

R2 ↔ R6

name token subset0location postal+8

8

possible

C10

R3 ↔ R4

name token subset0location postal+8

8

possible

C11

R3 ↔ R5

name token subset0location postal+8

8

possible

C12

R3 ↔ R6

name token subset0location postal+8

8

possible

Read-only. In the Workbench you answer here:

Real UK companies, resolved with no model, and a model's suggested answer to an open review. Hover a label in its rationale to see the records and comparisons it cites.

Hesperus is Phosphorus

Two observations.
One planet.

For centuries the evening star and the morning star were two things. Then observation showed they were the same planet: identity established by evidence rather than by definition.

Thisness does that for data. It takes observations from many sources (“IBM”, “IBM Corp.”, an invoice line, a contract party) and decides what they refer to, with a record of why.

O1Hesperusseen at dusk, in the west

O2Phosphorusseen at dawn, in the east

orbit sameperiod sametime seen differs, says nothing

V

Venus2 observations · committed with its evidence

active

Principles

Every conclusion keeps its reasons

Observations are not entities

A source record is an observation of something. Entities are stable identities above them, with IDs that never change when names do.

Similarity is not identity

Candidates, evidence, decisions and identity are separate steps with separate records. Scores are typed; there is no universal confidence float.

Identity is not relationship

“Subsidiary of” and “brand of” relate distinct entities. A group's sibling companies are never merged for sharing a name.

Everything is reversible

Merges and splits are ledger events over memberships. Nothing is destructively collapsed, and the ledger alone rebuilds current state.

Models propose; policy decides

Rules, Fellegi-Sunter, language models and people all contribute decisions. Only a commit, citing them, changes identity.

Read the design

Seven design documents, from the core model to real organizations, with every measurement.

The design documents

Three domains ship

Records in, entities out

A domain is a resolution profile: attributes, comparisons, constraints and the guidance a model is given.

Organizations

Haleon UK Finance Limited

HALEON UK FINANCE LTD

HALEON UK FINANCE LIMITaED

one entity

Haleon UK Finance Limited

Haleon UK Capital PLC, same address: related, not the same

People

Robert J. Smith, 1971-04-02

Bob Smith

R. J. Smith, same email

one entity

Robert J. Smith

Robert Smith Jr., same address: a household, not a match

Taxa

Quercus robur L.

Q. robur

Quercus pedunculata Ehrh.

one entity

Quercus robur L. (accepted, WFO)

Quercus ilex Lour., a later homonym of Q. ilex L.

Measured on real data

Real UK companies, identifiers hidden

GLEIF and Companies House records, with supplier-file variants generated from the true company. 62% of live UK companies share a registered address; the hard cases are real.

.9992

precision

3,507 records, no model

.912

recall

one false-merge cluster

.938

recall with a model

gpt-6-luna deciding contested pairs, $0.62

46,565

records at scale

precision .9965, recall .934

How it is measured

Start

A library, a command line and a Workbench

Python

from thisness import Thisness
from thisness.profiles.organization import organization

engine = Thisness("identity.duckdb", profiles=[organization()])
engine.ingest([engine.observe("organization", r, src) for r, src in rows])

result = engine.resolve_many("organization")  # fit, then resolve
outcome = engine.resolve("organization", {"name": "IBM Corp."}, src)
engine.explain(outcome.decision_id)  # the decision and its evidence

Command line

pip install 'thisness[ui]'
thisness ingest vendors.csv --resolve
thisness ui   # the Workbench

# a model decides contested pairs
thisness resolve --provider openai \
  --model gpt-6-luna --budget 2

Know what your records refer to