Skip to the content
Pansofica
Talk to us

Evidence note and sources

The world runs on registers that do not talk to each other. People are dying in the gaps, and AI is now being trained to guess across them.

The long form behind the release of October 2, 2026: each claim in full, with the figures and the pages they were read from.

The release

Boxing: the first attempt, and why gatekeepers create the risk

Pansofica began with combat sports because the failure is visible and the dead have names. A boxer’s record, medical suspensions and license are held by whichever commission sanctioned the last bout. There is no single register, and a suspension issued in one jurisdiction does not reliably reach the next.

The cost is not hypothetical. On average 13 participants die each year inside the ring. In February 2025, John Cooney, 28, died a week after a title bout in Belfast. In August 2025, Shigetoshi Kotari and Hiromasa Urakawa, both 28, died of brain injuries suffered on the same card in Tokyo. On September 29, 2026, Arturo Torres, 26, died six days after a seventh-round knockout in Mexico.

The clearest case of the gap itself is Hugo Santillan. After his death in 2019 it emerged that his final bout should never have taken place: officials in Argentina had not been notified of a 45-day suspension issued elsewhere. The suspension existed. The register that held it did not speak to the register that licensed the fight.

This is what Pansofica means by gatekeepers creating risk. Each commission is a careful keeper of its own file. None is responsible for the whole fighter. The record that would have stopped the bout sat in a drawer in another jurisdiction. Pansofica’s first corpus assembled 22,556 fighter records with their sources, and the rulings registers of two state racing and boxing authorities are now public with a page and a citation for every ruling.

Safety recalls in a globalized economy

A recall is a register entry that has to reach every owner, every workshop and every buyer of a used part, across borders, before harm is done. Scattered registers do not reach them.

About 4.7 million vehicles in the United States still carry an unrepaired Takata airbag inflator twelve years after the first recall, out of 67 million inflators recalled; at least 28 people have been killed and 400 injured. In 2026 one manufacturer issued a stop-drive order for every vehicle still unrepaired. The recall exists. The owners cannot be found, because the vehicle register, the sales register and the repair register are three different files.

Medical devices show the same curve. The United States Food and Drug Administration reports that Class I recalls, the most serious category, increased by 232 percent between 2020 and 2025. The Consumer Product Safety Commission announced a record 422 recalls in 2025. Each notice is published once, on one agency’s site, in one country.

The aviation case shows what happens when the register of record is forged rather than scattered. A single British broker, AOG Technics, sold jet engine parts with falsified documents from 2019 to 2023; CFM traced the parts into 68 engines, and a wider accounting counted 126 engines worldwide and 95 falsified documents, on aircraft flown by major United States carriers. Prosecutors put the losses at 39.3 million pounds. A part’s paper trail was trusted because no one could check it against a register that did not exist.

Pansofica has made the United States device recall register, the consumer product recall register and the vehicle manufacturer register public as cited corpora, with the source file and its fingerprint on every record, and has cataloged 16 further manufacturer and product registers outside the United States for the same treatment.

Migration and national security: registers that do not talk

Borders are where scattered registers meet moving people, and the numbers are the largest in history. 117.8 million people were forcibly displaced at the end of 2025, and the United Nations refugee agency budgets for 136 million displaced and stateless people by the end of 2026. Every one of them is a person whose identity, status and history sit in registers that were never built to be read together.

The tools that exist are only as good as the registers feeding them. Interpol’s Stolen and Lost Travel Documents database holds around 138 million records, yet it depends on each country submitting and each border checking. In January 2026 a database error flagged valid United States passports as stolen and stranded travelers in Rome, Thailand and Qatar. A register that no one can check against its source fails in both directions: it misses the fraud and it convicts the innocent.

Europe has audited its own systems and said so. The European Court of Auditors found that data in the Schengen border systems is in some cases not included, incomplete, or not entered in a timely manner, which reduces the efficiency of border checks, and the European Commission’s own 2025 State of Schengen report records that the main weakness of the system is the sharing of information between member states, with national security legislation in some cases limiting the exchange. The remedy, an interoperability roadmap, now runs to 2027 and 2028.

Pansofica does not hold immigration or security records and does not intend to. The warning is structural: the same failure that killed Hugo Santillan, a valid record in a register that did not speak to the one that mattered, is the failure that Europe’s auditors describe at its borders. Any system built on scattered registers will have gaps exactly where the people moving through it are most vulnerable, and exactly where those who mean harm will look.

AI laboratories are inferring truth from scattered information

The most consequential readers of scattered registers are no longer people. They are the AI systems now answering questions about doctors, products, companies and travelers for hundreds of millions of users, and they infer. Where the record is scattered, incomplete or unreachable, the model fills the gap with a plausible sentence.

The measurements are public. On general tasks, the 2026 frontier models still hallucinate in roughly 4.6 to 6.1 percent of answers. On the questions that matter most the figures are far worse: Stanford researchers measured hallucination rates of 58 to 88 percent on legal queries across major models, and a 2026 study of the scientific literature counted thousands of invented citations every month entering preprint servers and medical archives. An invented citation is a scattered register problem in its purest form: the model could not reach the record, so it made one.

The people building these systems are now saying in public what the figures imply. In September 2026 a research lead at one laboratory put the probability that AI kills all humans within the decade at more than 10 percent, its chief executive published an essay arguing the frontier must be paced, and on September 24 Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act. Five days later the United States government launched America.gov, an AI front door that answers only from official sources, with links, across 29,000 federal websites.

Pansofica’s warning to the laboratories and to the governments deploying them is this. A link to an official page is not a record. A page changes, moves and disappears, and a model trained or grounded on it cannot show which version it read. If the world is going to let machines answer from the public record, the public record has to be consolidated into something a machine can cite and a human can check: one address per record, the source file and its fingerprint, the publish date, the publisher’s terms, and a way for the person the record is about to correct it. Confidence in AI answers will not come from larger models. It will come from better records.

What Pansofica has built, and what is live today

Pansofica is an evidence layer for the public record. It does not publish opinions about people or products. It consolidates what authoritative keepers have already published and makes each record checkable.

One engine, any register. A register is described in a one-page specification, not coded. The same engine has read a federal exclusion list, state racing and boxing rulings, product and device recalls, a vehicle manufacturer register, a bank charter register and a state parcel roll with no code written per source.

Evidence on every record. Every record carries the register, the publisher, the source file, the file’s SHA-256 fingerprint, the publish date, the date it was read and the publisher’s reuse terms, verbatim. Nothing is invented. The record says what the file says.

An address for every record. Each record has a public page with a citation line, and the same record is available to ChatGPT, Claude and any software through one interface.

History, never deletion. When a keeper republishes, superseded records are retired, not erased, so any answer given last month can still be checked today.

Correction by the subject. The person or organization a record is about can claim, flag or dispute it. The correction sits beside the record. The record itself is never rewritten.

As of October 2, 2026: 12 public corpora holding more than 87 million records, including 9.8 million licensed healthcare providers, 83,975 federal healthcare exclusions, 18.8 million land parcels, 173,130 recall, manufacturer, bank and ruling records, and 22,556 combat athletes; 182 registers cataloged with their obtainability and reuse terms, kept by governments, sports federations, exchanges, standards bodies and industry consortia; 16 manufacturer and product registers outside the United States queued for the same treatment. The catalog is published as a coverage map on this site.

The next step is scale. Pansofica is in discussions with a number of family offices on a $100 million seed round. The capital takes the engine from 12 public corpora to 25,000 registers and 5 billion records over 36 months, builds the team that operates it, and funds the legal and licensing work to open registers in every jurisdiction.

Sources

All figures as read on the date of issue.

The releaseAll releases