How the record is built
Fundprint is not a hand-kept list. It is a deterministic pipeline: the same public inputs run through the same code produce the same record, and every published claim carries the document that proves it.
One machine, start to finish
- IAcquire649 + 255 records in
The federal NPPES provider registry (649) plus owners' own machine-readable location directories (255). Every page is fetched and content-hashed on the way in.
- IIStorepgvector + snapshots
Nothing advances without a stored, hashed copy of the document behind it. Names are embedded as vectors in Postgres so the resolver can reason over them.
- IIIResolveclinic → owner → firm
A deterministic name matcher links clinics to owner entities and owner entities to their parent financial firm. Chain confidence is the weakest link along the chain.
- IVValidate95% hand-checked gate
Confidence floors and a 95% hand-validation gate. Contradicted or under-confidence claims are quarantined, not published. Over-matched brand names are rejected.
- VPublishverified chains out
Only claims that clear validation are written to the frozen snapshot this site reads. Every published claim carries the source that proves it.
Where the clinics come from
One source is NPPES, the federal provider registry. Anyone can download it and check us. But it enumerates registrations, not centers: a chain that runs forty centers under three identifiers appears as three. And nothing obliges a company to withdraw a registration when a center closes, so the registry also reports places that are no longer there.
The other is the operator’s own public list of its centers, read from the structured data its site publishes for search engines, never scraped from rendered prose. Where that list is complete, it wins. It is what the company tells parents today, and it corrects the registry in both directions at once: it names centers the registry never saw, and it declines to name the ones that closed.
Both are de-duplicated on one key, the street address and ZIP within an owner, so a center appearing in both is counted once.
A center matching no tracked owner is left unlinked. A registration an operator’s own complete directory does not list is quarantined, not deleted.
What we refuse to do
- Never assert without a source
Every published ownership claim traces to a content-hashed snapshot of a public document. A claim missing its source does not publish, at any confidence.
- Verify the current owner
Ownership turns over. We cite the owner today, not a stale deal, and drop owners we cannot currently substantiate.
- Reject over-matches
A brand name is only a string, and two companies can share one. “Hope Bridge” matches Hope Bridge Counseling, which is a different business. Names too generic to tell apart are refused, and an operator’s own directory is what catches the rest.
- Label honestly
Pension funds and family offices are named as what they are, not folded into private equity.
Pinned, versioned, open
The site ships a frozen snapshot, not a live feed: what you read is a reproducible state, dated and version-pinned, so a screenshot dates itself. The methodology is versioned separately and the whole pipeline is open source.