Family Office Intelligence

Nine classes, never collapsed: confirmed single family office, probable SFO, confirmed multi-family office, outsourced family office provider, family investment company, family foundation (related), family investment vehicle, candidate, and not a family office. Every classification carries a confidence, a basis in words, the evidence class (fact, inference, estimate), the source, its date, a quote where there is one, and a verification date. A candidate is not counted as coverage. An inference is never silently promoted to a fact.

Form ADV cannot define the universe: the family office exclusion keeps most single family offices off it. So discovery is multi-source. The SEC adviser roster gives registered MFOs and the registered minority of SFOs, with three admission rules that are recorded on each entity (a name that says family office; an Item 5.D client mix that reads as one family with institutional assets; a client or services description that says family office). Form 13F gives every manager over $100M in 13(f) securities including unregistered SFOs, admitted by name, by resolution to a known office, or as a candidate when the filer is on no adviser roster and is not institutional or corporate by name. Form D gives investment vehicles whose issuer or related person is an office. Cited research adds the offices that file nothing.

An entity resolves by an official identifier first (CRD, CIK, file number), then by an exact normalised name within the same entity type and state, and otherwise is created. Two offices merge only on a shared identifier, an official other-name list (Form ADV Item 1.B as served by IAPD) or a citation, never on a similar name. A merge is reversible: the folded row stays with what moved recorded on it. An LLC with a surname in it is not the family's.

One substantiated investment: office, vehicle where known, target, date, kind, asset class, sector, lead or participant, control or minority, co-investors, sponsor, lender, exit. The amount is the number the source disclosed and nothing else; a research seed that puts a number in the amount must quote the sentence that states it. If six investors shared a $100M round, no row says each wrote $16.7M.

Public holdings are modelled separately, one row per filer per quarter per security, and the quarter-on-quarter diff produces opened, increased, reduced and exited events. A 13F covers long positions in 13(f) securities only. It is never presented as the portfolio, and its table value is never called AUM.

The mandate has three panels. Stated is the office's own words, cited. Observed is counted from the investment record. Inferred is what a regulatory shape or a team suggests, labelled as inference. Observed behaviour drives ranking and matching; a stated sector with no deal behind it is kept, shown, and weighted below one with six.

Twelve separate dimensions, each 0 to 100 from listed components: identity confidence, data confidence, freshness, active deployment, direct investment propensity, external manager propensity, private markets propensity, independent sponsor propensity, real estate propensity, co-investment propensity, contactability, mandate confidence. There is no single family office score. The components are on every profile.

Rights are data. Each source carries its access mechanism, cadence, redistribution rights and the operative terms quoted; unknown means not published. The product holds professional roles, investment activity, public transactions and business contact points with their verification status. No home addresses, no personal phones, nothing about minors, no speculative private-family information. A Form ADV main office flagged as a private residence is not published as an address.

Call every investment firm a family office. Infer office AUM from family net worth. Infer wealth from 13F holdings. Infer check sizes from round sizes. Attribute property from surname similarity. Merge on similar names. Label foundations as offices. Confuse grants with investments. Show a departed person as current. Fill a missing value with a plausible guess. Publish a claim without provenance. Ingest a source whose terms prohibit the use. Copy a competitor's dataset.