A correct fact is of no use to a machine if it cannot be found, read, and checked. A Data Fidelity Register makes all three true of the same fact, on the organisation’s own domain, verifiable by anyone.
Everything an organisation has published about itself is a fact a machine might need to find: its figures, its filings, its credentials, its dates and registrations. Take one, say an organisation’s turnover for the year.
The figure existsIt is defined, audited, and published. It appears in a financial report and again in a press release. On the organisation’s website it sits among a great many other figures, and again inside a PDF offered for download on the same page. Everywhere it appears, it is correct.
The judgement is goneBut the same figure now sits in several places at once, and picking the authoritative one was always a small skill. A person knew to open the annual report and to distrust the aggregator. That knowing is what has gone missing. The reader now is usually a machine, working at a scale where nobody opens the report, and it cannot tell what is true from what is merely there.
The gap is filledSo the machine goes looking. It fetches what it can reach and reads what it can parse. If the figure is buried in a PDF it does not open, or lost among the other numbers on the page, it does not stop and say so. It falls back to what it absorbed in training, which may be years stale, or it averages the several figures it did find, or it just produces something plausible. A filled gap looks exactly like a fact, and it comes back with the same confidence as one.
The error spreadsFrom there it travels. Machines read what other machines wrote. A figure that was never quite right, or right once and now superseded, is copied forward, cited, and repeated until the wrong number is the one that appears everywhere and the correct one is the outlier. The original, sitting correctly in a report nobody’s machine could read cleanly, loses to the copy.
And this is one figure. An organisation publishes thousands.
Revenues, headcounts, dates, registrations, ratings, targets, holdings. Most of them appear in more than one place, often in more than one version, and every one of them can go the way the turnover figure went. The trouble was never that a single number is hard to pin down. It is that everything an organisation has said about itself is spread across formats a machine reads unevenly, and wherever the real figure is hard to reach, a made-up one gets in.
The problem is not that the figure was never published. It is that it was not published in a form a machine could find, read plainly, and confirm. That is what a Data Fidelity Register provides, and it rests on three things being true of the same figure at once.
ReachableThe figure is placed in one known location on the organisation’s own domain, as plain data rather than buried in a document or lost among others on a page, so a machine reads it directly instead of guessing at it or giving up.
StructuredThe figure is recorded as what it is: a named value, for a stated period, of a stated kind, so a machine knows it has found a turnover figure for a given year rather than an unlabelled number it has to interpret from the surrounding text.
VerifiableThe figure is bound to the public source it came from and sealed, so anyone, anywhere, can confirm it is unchanged from that source, at the level of the single figure rather than a whole document, and without taking Apparat’s word for any of it.
This serves any system that takes a fact and runs with it, no person checking each one: a retrieval model answering a question today, a corpus a model trains on, an internal tool that drops a figure into a report, and, more and more, systems that act on the figure and not just quote it. The register assumes nothing about which of them is reading. It puts the fact in a form all of them can use and any of them can check.
Each of these is a solved problem on its own. What no one does is all three, on the same figure, in the same place.
A website can be made readable to machines, but it carries no structure and nothing to check against. Structured data gets published and then sits buried, backed by nothing. Documents can be timestamped and proven unaltered, though only whole documents, not the single figure inside them, and with none of the rest. The pieces exist on their own and never come together, which is how a correct figure ends up unreachable, or unreadable, or impossible to confirm. A register holds the three in one place.
Building one takes five steps, from deciding what belongs in it to leaving it open for anyone to check.
01
Curation
What belongs in the register, and what is kept out.
A register begins with a decision about what it is of. We take a defined set of an organisation’s own public documents, its results releases, its regulatory filings, its official registrations, and treat those as the sources. What falls outside that set is not in the register, and the register says so. Scope is not a limitation to hide; it is the first thing a reader needs to know, because it determines what the register can answer.
Within that set, a fact is admitted only if its source supports it plainly. A value read from a comparative table without its column headers, a figure whose surrounding text does not establish what it measures, a claim that has quietly gone stale: these are left out, even when they are probably correct. The register attests that a value is faithful to the source cited beside it. A source that does not establish the value cannot be sealed, so it is not.
02
Extraction and structuring
Each fact recorded as a unit, with its kind and subject.
Each admitted fact is drawn out and recorded as a small, self-contained unit: a label, a value, the period it applies to, the exact quote it came from, and the address of the source document. A figure never travels without its unit or its date, because a number without them is where most machine-read errors begin.
Every fact also carries what kind of assertion it is and who it is about. A registration number that belongs to a parent company, a target the organisation set rather than a result it reported, a fact about the sector rather than the organisation itself: these are marked as such rather than silently promoted into the organisation’s own record. A model does the reading; a separately configured model checks it. Two passes catch more than one, though neither is a claim that the underlying fact is true, only that it faithfully reflects its source.
03
Sealing
A fingerprint binding each fact to its source.
Each fact is sealed: a cryptographic fingerprint is computed over the six fields that make it up, so that any later change to the value, the unit, the period, the quote or the source is detectable. The seal binds the value to the source cited beside it. That is the whole of what it certifies, and it is worth being exact about the boundary: the seal proves fidelity to the source, not the truth of the source. Apparat verifies; it does not judge. If a public document is itself wrong, the register reflects that faithfully, and the seal lets anyone confirm the register did not add an error of its own.
04
Publication
Released on the domain, committed to a public log.
The sealed facts are gathered into a release, and the release itself is committed to: the individual seals are combined into a single fingerprint, which is entered into a public transparency log that no one, including us, can alter after the fact. The register is then published on the organisation’s own domain, as plain files a person can read and a machine can parse: the same facts as a page, as structured data, and as a table, each fact at its own stable address.
Publishing on the organisation’s own domain matters, because that is the strongest available signal that the organisation stands behind the record. Publishing the commitment to a public log matters just as much, because it means a reader does not have to take our word for anything. The two together are what let the register be trusted without trusting us.
05
Verification and maintenance
Checkable by anyone, kept current over time.
Anyone can check a register, and checking it needs nothing from us. A reader recomputes the seals from the published facts, confirms they combine to the committed fingerprint, and finds that fingerprint in the public log. If every step agrees, the register is exactly what was published, unchanged. We publish the method and a reference tool that does this, so the check is a procedure rather than a promise.
A register is kept, not left. When a fact changes, it is re-sourced and a new release is published, and because each fact keeps a stable address, a citation made against an earlier release still resolves after the correction. Every release is recorded, including those never deployed, so the history is walkable and its gaps are visible rather than hidden. A register is only as useful as it is current, and being current is part of what is maintained.
None of this depends on trusting Apparat, by design. The format is an open specification, published under CC BY 4.0 with a citable DOI. The tool that checks a register is open-source under the MIT licence. So the way a register is built and the means to verify one are both public: anyone can read the format, run the checker, and confirm a register for themselves, whether or not Apparat is around to ask. What is ours is the engine that produces a register; what is open is everything needed to trust the result without us.
Apparat keeps a register of its own, built to the same standard, at dfr.derappar.at. It is the shortest way to see what the rest of this page describes.
Apparat builds and operates Data Fidelity Registers for organisations. If getting your public facts right matters, start with a short conversation about what your register could cover.
