Mimir
01How it works

Deterministic identity. Inferred control, scored and shown.

Two different things happen inside Mimir, and conflating them is how this industry gets to sound more certain than it is. Keeping them separate is the whole design.

02The distinction

Resolution is a function. Control is an inference.

Resolving one canonical id per building and per person is a pure function of the source records. Same inputs, same id, every time, with no machine learning anywhere in the identity spine. That is what lets an answer be re-derived rather than trusted.

Working out who controls whom cannot be a function, because no filing states it. It is inferred from behaviour, so every control edge carries the evidence behind it and a confidence value — shown on the surface, not buried in a single score.

441 Vanderpool Street, Brooklyn, NY 11215synthetic
Ownership chain
Block 1042 · Lot 17 · 12 units · built 1928
  1. Owner of record
    VANDERPOOL HOLDINGS LLC
    Deed recorded 2019-04-11nyc_acris
  2. Controlled by
    HARROW STREET CAPITAL LLC
    Shared registered agent + mailing addressny_dos_corps 0.82
  3. Ultimate controller
    A. Reyes
    DOS process address across 6 entitiesny_dos_corps 0.74
The receipt
ACRIS 2019041100428001
Source
nyc_acris
Document type
DEED
Recorded
2019-04-11
Grantor
ESTATE OF M. OKONKWO
Grantee
VANDERPOOL HOLDINGS LLC
Consideration
$2,150,000

Every value on the left opens a record like this. Nothing is asserted without one.

03The pipeline

Six layers, each rebuildable from scratch.

01HarvestRaw public records

Government dumps pulled whole and never filtered at collection — county recorders, tax rolls, corporate registries, code enforcement, parcel geometry. Filtering at intake destroys answers you cannot get back.

02FingerprintPer-source tokens

Each source produces normalised tokens for the things it describes: a parcel identifier, an address, a name. Normalisation is a fixed function, so the same input always yields the same token.

03ResolveOne canonical id

Overlapping identifiers collapse to one id per real building and one per real person, through a deterministic ladder of match rules. No machine learning, no probabilistic record linkage, no clustering that drifts between runs.

04MaterialiseDomain tables

Resolved ids become the queryable record: properties, owners, sales, mortgages, liens, distress. Every field keeps a pointer to the source row that produced it.

05LinkThe control graph

Behavioural signals connect properties to entities and entities to people. This layer is inference, so each edge carries evidence and a confidence value rather than being asserted flat.

06ServeThe terminal

Search, dossiers, portfolios, stacks and alerts read the resolved record. Any value on screen opens back to the government document underneath it.

The whole chain runs self-hosted on commodity hardware and can be rebuilt from the raw dumps. Nothing in the answer depends on a vendor feed we cannot reproduce.

Bring a building you already researched by hand.

The fastest way to judge the system is to watch it re-derive something that took you a day, with a citation on every step.