Search 14 public corpora
Every public record, with its receipt.
Pansofica is public-record infrastructure for AI agents: the grounding layer that turns the world's public registers into citable records, each with its receipt. Machines can cite it. People can check it and correct it.
Pansofica Corpus Labs builds the public record for machines.
Pansofica Corpus Labs is the research and data company of PXL8 Technology, Inc., Miami. The engine was built in production and it runs the same way on parcels, providers, recalls, exclusions, banks and manufacturers, and the method is published as a standard anyone can check. The name comes from Comenius, who in the 1600s called the ordering of all knowledge so that anyone could use it pansophia. Ours is narrower and checkable.
- Every record with its receipt, including our own claims about ourselves.
- An open standard and an open correction rail.
- The data stays public, the method is ours.
From 186 registers cataloged to 25,000 opened, over 36 months.
The engine already opens a register from a one-page description. The plan is to write the pages at scale, in every jurisdiction and every register family where a fact about a person, a product or a company is public, and to keep every one of them refreshed on its keeper's clock. Alongside the corpus, four things grow with it.
| Workstream | What it delivers | Public milestone |
|---|---|---|
| The corpus | Registers opened across the United States first, then the EU and the Commonwealth; non-English sources; keepers who publish only on paper | 25,000 registers; every public fact about a person, product or company reachable through one door |
| The door | A laboratory-grade API: snapshots, license manifests per record, metering, full history, published evaluations | Agents from the major laboratories grounding on it under contract |
| The standard | The Receipt Specification adopted beyond our own records: a validator, a mark, and keepers publishing to it directly | A receipt on a record from any source means the same thing everywhere |
| The correction rail | One door for a person or an organization to see their record and correct it, with the keeper in the loop | Every subject in the corpus can claim and correct their record |
| Refresh | Every register re-read on its own clock, with the change history kept | A record is never older than its keeper's last file |
The round sizes and the financial plan are in the investor memo, available by request; the page states the work, not the sums.
The world runs on records nobody can check.
- The world runs on records nobody can check.
- Machines now read more of the web than people do.
- Generated text is leaking into the record.
- The gatekeepers add risk, not trust.
- Pansofica's answer is arithmetic anyone can check.
The world runs on records nobody can check. A license, a recall, a ruling, a sanction, a parcel, a bank charter: each sits in a register kept by someone, published in a form of the keeper's choosing, moved, renamed and withdrawn without notice. A fact copied out of a register loses its origin the moment it is copied. Most of what circulates about people and organizations is that kind of copy.
Machines now read more of the web than people do. Cloudflare reports that for the first time more than half of Internet traffic is not human, that daily requests from AI agents on its network grew by more than 1,700 percent in a year, and that by June 2026 more than half of crawler requests were for AI training. An agent that acts on a copy acts on something no one can trace.
Generated text is leaking into the record. Models trained on their own output lose the tails of what they knew within a few generations (Shumailov et al., Nature 631, 755 to 759, 2024). The same leak runs through the data agents read: summaries of summaries, with the source gone. The only defense is a record that carries proof it came from the keeper.
The gatekeepers add risk, not trust. Data resold as a commodity arrives without origin, date or terms. The buyer inherits the seller's errors and the seller's exposure. A business built on selling copies has no place in a world where every reader, human or machine, can demand the source.
Pansofica's answer is arithmetic anyone can check. Every record on the door carries a receipt: the keeper, the file, its hash, its date. The bytes are kept, so the hash can be recomputed by anyone, at any time. Where several keepers hold entries about the same subject, they resolve into one record, and the join is measured against labeled data rather than asserted. Coverage is counted: which registers exist, which are open, which refuse, what has been loaded. Nothing on the door asks to be trusted.
Two things changed the timing. Regulators now require an agent to say what it relied on: the EU's transparency duties apply from August 2026, and the high-risk obligations follow in December 2027. And the agents themselves moved from answering questions to taking actions, where a wrong fact about a person, a product or a company is no longer an error in a chat window but a payment, a hire, a recall missed.
- Models reason. Registers know. Pansofica connects them.
- The web is what the model read. The record is what the world did.
- Every answer about the real world should come with a receipt. Soon it will have to.
Four things a lab does with a public record that carries its receipt.
One door, every register, the same shape. Each use below is a different shape of the same engine; the first two run on the door today, the third is how an agent is held to the same standard, the fourth is the standard itself.
Ground an answer in the record, and show the receipt.
An assistant that answers from the door cites the record it read, and the citation carries the file, its fingerprint and the date. The reader can check it; so can the next model.
Count a subject once, and never guess.
Entries about one subject from different keepers resolve into one record, with a receipt on every join. Where two candidates remain, the engine holds a doubt instead of choosing: an evaluation set stays clean, a screen never fires on a namesake.
Govern an agent that writes.
Every write an agent makes is kept with a snapshot of what it saw, a receipt and a way back. A new version replays past decisions before it ships, its error rate is measured against a stated bound, and a write it cannot make inside that bound is withheld.
- createdby an operator
- writtenby the agent, with what it saw
- editedby an operator, kept apart
- withheldoutside the stated bound
An illustration of the shape.
Adopt the standard.
The Receipt Specification says what a citable record must carry. A validator anyone can run checks any record, from any source, ours included; a lab can publish to it, or require it of its suppliers.
POST https://pansofica.com/api/standard/validate
{ "record": { ... "receipt": { "register", "file", "sha256", "readAt" } } }
200 { "valid": true, "version": "1", "findings": [], "checked": { "schema": true, "hashRecomputed": true } }The exclusion check: 9.8 million providers, one list of 83,975 exclusions.
Federal law bars payment to an excluded provider, so every payer screens its providers against the list. The list names an identifier for one exclusion in ten. The other nine must be found by name, among 9.8 million providers. 7,724 excluded people and companies still hold an active identifier in the provider registry: the list does not switch it off.
Exclusions that carry an identifier
- carry an identifier8,880 11%
- found in the provider registry8,700 10%
- still active in the registry7,724 9%
Exclusions that carry a name only
- name only, no identifier75,095 89%
- match a provider by name and state9,441 11%
- match exactly one provider5,468 7%
- match two or more: the join decides3,973 5%
A wrong match denies a clinician a living; a missed match is a federal penalty for the payer. Both registers are on the door with their receipts, and the join between them is measured, not asserted.
One pair, as the door serves it
Public is not open.
Of 186 registers cataloged, 56 are open. The rest are public on paper.
- 56open 30%
- 6non-commercial 3%
- 119terms not yet read 64%
- 5refused by their terms 3%
A register that can only be searched one name at a time, a license that forbids reuse, a PDF scanned from a printout, a list published without a date. Each is technically public and practically closed. We count them.
Eleven register families, one door.
Two families carry the mass. The rest are hairlines on this scale, and each is a register someone needed. The scale is linear.
- Parcels18.8M
- Providers9.8M
- FMCSA Motor Carrier Census — One record per subject, as published by U.S. Department of Transportation, Federal Motor Carrier Safety Administration.4.5M
- Exclusions83,975
- Device recalls59,290
- Consolidated Screening List — One record per subject, as published by U.S. Department of Commerce, International Trade Administration.26,184
- Manufacturers23,068
- Combat sports results21,110
- Banks4,231
- Product recalls2,276
- Racing rulings290
Every public fact about a person, a product or a company, reachable through one door: from 186 registers cataloged to 25,000 opened, over 36 months.
The engine opens a register from a one-page description. The plan is to write the pages at scale, in every jurisdiction and every register family where a fact about a person, a product or a company is public, and to keep every one refreshed on its keeper's clock. Four things grow with the corpus.
- The corpusRegisters opened across the United States first, then the EU and the Commonwealth; non-English sources; keepers who publish only on paper25,000 registers; every public fact about a person, product or company reachable through one door
- The doorA laboratory-grade API: snapshots, license manifests per record, metering, full history, published evaluationsAgents from the major laboratories grounding on it under contract
- The standardThe Receipt Specification adopted beyond our own records: a validator, a mark, and keepers publishing to it directlyA receipt on a record from any source means the same thing everywhere
- The correction railOne door for a person or an organization to see their record and correct it, with the keeper in the loopEvery subject in the corpus can claim and correct their record
- RefreshEvery register re-read on its own clock, with the change history keptA record is never older than its keeper's last file
Progress prints on the strip above as it lands.
The receipt, live.
One record as the door serves it: the file it came from, the fingerprint of that file and the day it was read. Check it yourself with one command.
- Recall number
- 24299
- Title
- Claire's and Icing Stores Recalls Halloween Witch Hats Due to Burn Hazard and Violation of Federal Flammability Requirement
- Recall date
- 2024-07-11T00:00:00
- Manufacturers
- []
- Made in
- [{"Country":"China"}]
- Importers
- []
- Distributors
- [{"CompanyID":"","Name":"Claire's and Icing Stores, of Hoffman Estates, Illinois"}]
- Sold at
- [{"CompanyID":"","Name":"Claire's and Icing stores nationwide from July 2023 through December 2023 for about $15."}]
- Register
- CPSC Consumer Product Recalls
- Publisher
- U.S. Consumer Product Safety Commission
- File
- pages.json · 7,128,028 bytes
- Published
- 2026-09-24
- Read
- 2026-09-29
- Terms
- The API provides machine readable access to publicly available recall information visible on cpsc.gov. (read 2026-09-28)
- Source file
- not retained
The facts are already public. We are attaching the receipts.
Pansofica is raising. Investors read the memo first, then meet the door.
Talk to us.
Say who you are and why you are writing. A code goes to your work address, and nothing reaches us before you type it.
Investors read the memo first. Ask for it here. It goes to a verified address at your firm.
- Who you are
- Why you are writing
- Your work email
- The code we mailed you
We keep what you send only to reply. Nothing is added to a list. Personal webmail and disposable addresses are not accepted. Privacy policy