A register is a page. The engine does the rest.
One page describes a source. The engine turns it into records with proof.
OIG List of Excluded Individuals/Entities (LEIE)U.S. Department of Health and Human Services, Office of Inspector General
UPDATED.csvoig.hhs.gov
Fingerprint, SHA-256
A source publishes a file. The engine reads it and takes its fingerprint, so anyone can check it is the same file.
Compile, link, resolve, infer, serve.
Two stages run today. Three are specified, and measured before they are claimed.
- 1Compile
A page becomes records bound to the keeper's bytes. Walls first, history kept, nothing deleted.
on the door · 25 pages - 2Link
Exact identifier joins across registers, each with its rule, its basis and both receipts.
on the door · on every record view - 3Resolve
The same subject across keepers, names in any script, under a measured bound. Below it, a doubt.
specified, measured before it is claimed - 4Infer
Contradictions, failed checks, patterns that resemble known shapes, signals with lead time.
specified, measured before it is claimed · first class live: contradictions - 5Serve
One door, one shape, to people, code and agents. HTTPS and the Model Context Protocol.
on the door
The record remembers. Links, corrections and history are kept beside the file, so the next reader starts where the last one stopped.
Measured.
Every number has a method and a date. Where there is none yet, it says so.
Numbers on this page: Oct 7, 2026 for the door's; Oct 3, 2026 for the measurement; Oct 5, 2026 for the pages. Not claimed: resolution precision, licensee savings, the reading-side figure. Each is published when it is measured.
Every measured figure, with its method
- Register pages written25
one file per register, counted by git date from the engine's repository and committed with its as-of (Oct 5, 2026); 2 · 5 · 3 · 2 · 13 by day from September 26.
- Public kinds on the door24
GET corpora, Oct 7, 2026.
- Records on the door71.3 million
GET corpora, Oct 7, 2026; current records only, history kept behind them.
- From page to door2 h 55 min
TED, October 4, 2026: page written about 15:20Z, dry run clean 15:54Z, loaded 17:52Z, public 18:15Z, including two engine fixes made on the way. Find a Tender and Contracts Finder followed the same afternoon.
- Parameters per page14
the union of top-level parameters across the register pages (attributes, connector_spec, dates, drop, holdings, identifiers, names, record, refresh, register, search, source, subject, walls); a page is 123 to 883 lines (123 to 883 on 2026-10-04, read from the files).
- Description against readingabout 14,300 tokens against about 475 million
seven registers, 173,130 records, 1.52 GB of source files: the seven pages total about 14,300 tokens; a model reading every source row once would read an estimated 475 million (o200k_base; the exclusions CSV tokenized whole, the rest measured on a 20 MB sample and scaled by bytes). A cut of 99.997 percent in what has to be read. Measured October 3, 2026.
- Proof of the compile685 of 690
the engine's own verify runs: row counts equal to the source on all seven registers; 685 of 690 sampled fields identical, one differing (CPSC).
- The reading sidenot claimed
an agent reading the door's record instead of the keeper's own payload was measured on 25 records across five kinds: 40 percent smaller overall, but providers and manufacturers are larger on the door because a record carries every field and its receipt. This variant failed and is not claimed.
- Resolution precisionmeasured and published here when it is
where several keepers hold entries about the same thing, the engine holds them under a bound it keeps and keeps a doubt rather than a guess; the bound's precision against labeled data has not been measured. No percentage is claimed for it.
- Cost saved for a licenseemeasured and published here when it is
one register owner's or one data company's own source through the engine against its pipeline, two numbers published: time to onboard and precision with doubts.
Methods: docs/features/CONNECTOR-TOKEN-CLAIM.md in the engine's repository; the door's own GET corpora and door/status.
How this compares
Each of these exists somewhere. We have not found them together.
| Who | What it does | What the engine adds beside it | Read |
|---|---|---|---|
| Senzing | an embeddable resolution engine licensed per record to data companies and governments | a public receipt per record bound to kept source bytes, with a public validator | senzing.com, 2026-10-03 |
| Quantexa | decision intelligence over entity networks for banks and governments | a declarative page per source and a public, dated contradictions log | quantexa.com, 2026-10-03 |
| OpenSanctions | a commons of sanctions and persons of interest, open crawlers and per-dataset metadata, with a business license | an engine a keeper or a data company licenses to run its own register | opensanctions.org, 2026-10-03 |
| Airbyte and Singer | open-source connectors, a declarative builder manifest per source | terms walls, the accounting rule, kept bytes and receipts in the declaration | docs.airbyte.com, 2026-10-03 |
Read on the date shown from each party's own site; named as what each does, not as a comparison of results.
What it never does.
- Guess.A fact it cannot source is not served; a doubt is kept as a doubt.
- Delete.A keeper's new file supersedes a record; the old one keeps its dates.
- Store what the terms forbid.The wall is in the page, before the first row is read.
- Resell a record without its receipt.Every record leaves the door with the register, the file, the fingerprint, the dates and the terms.
- Decide about a person.A record is what a keeper published; the correction rail is for the person it is about.
Run your register through it.
Two numbers from your own source, measured side by side with your pipeline: time to onboard and precision with doubts.