An open ontology for UK public registers, tested against Companies House, the Charity Commission and the Global LEI System

What we measured

Four register surfaces, all keyless, all pinned to dated snapshots. The Companies House Free Company Data Product of 1 August 2026 carries 5,695,465 live companies; its DissolutionDate column is empty on every row, so the product contains live companies only, and absence from it never proves nonexistence. The Charity Commission full register download of 17 August 2026 carries the registered company number of 31,612 charitable companies. The Companies House PSC snapshot of the same day carries 14,923,476 records across 12.9 GB of JSON. The GLEIF golden copy published at 08:00 that morning carries 3,403,856 LEI records, of which 117,324 name UK Companies House as their registration authority, verified against GLEIF's own registration-authorities API, with Companies House Gibraltar excluded as the false friend it is.

Finding one: 1,268 charitable companies point at nothing

Of the 31,612 Registered charities that carry a company registration number, 1,268 cite a number absent from the live register. That is 4.01 per cent. Specimens: charity 211570, SPORT HORSE BREEDING OF GREAT BRITAIN, cites 00221470; charity 214020, THE W R V S TRUST, cites 00522824; charity 212735, THE FOOD EDUCATION SOCIETY, cites 00156201. None of these numbers exists on the live register.

The honesty caveat, stated plainly: the free Companies House product contains live companies only, so an absent number means the number is wrong or the company has dissolved while the charity remains Registered. The split between the two cannot be computed from open bulk data. But both readings are register-boundary failures, because a Registered charity citing a dissolved corporate shell is itself a finding. The operational consequence is direct: anyone joining the charity register to Companies House, for grant due diligence, procurement screening or beneficial-ownership analysis, silently loses 4 in every 100 charitable companies at the first join, and nothing in either register tells them.

Finding two: the placeholders that pass validation

Inside the dangling set, 18 charities give the placeholder 00000000 as their registered company number, 9 give 01234567, and 5 give 12345678. Every one of these 32 values passes the 8-digit format rule. Charity 217602, SOUTH CROXTON VILLAGE HALL, is registered with company number 00000000, and no format validator on earth will object.

This is the cleanest demonstration we have yet found of why identifier validation needs layers. Our SHACL model separates them explicitly: layer 1 checks the shape of the string against the scheme's declared rules, layer 2 checks the syntactic defect classes, and layer 3 checks that the identifier resolves in the target register and that the two registers agree about what it names. The placeholders sail through the first two and are caught only by the third. A register that runs format validation and stops has validated nothing.

Finding three: the company-number field is a lost-property box

17 charitable companies publish values in the company-number field that violate the register's format entirely, and most of them are not typos. They are identifiers from other schemes: HMRC charity tax references (XN89306 at charity 238015, XR38029 at THE DAKHLEH TRUST, XT17973, XT12514), a mutuals register number (IP24891R), six CIO-scheme codes (CE00020, CE00571, CE00683, CEO10917, CEO11932, CEO13168), a Sea Cadet unit code (MSSC298), and a malformed royal charter number (RC00834, seven characters where the register form is RC followed by six digits).

The pattern is that regulators file whatever identifier they hold, because the field declares no scheme and validates against none. The fix is a scheme registry: a machine-readable declaration of each identifier scheme's shapes, canonicalisation and authority, so that a value can be checked against the rules its claimed scheme declares and recognised when it belongs to a different scheme entirely. Our ontology ships one as SKOS, with every scheme above in it. The register does not have one.

Finding four: 1,082 charities disagree with Companies House about their own name

Of the 30,327 charitable companies that join to a live Companies House record, 1,082 carry a name that diverges from the Companies House name after conservative normalisation that ignores case, punctuation, ampersands and suffixes such as Limited. These are real divergences, not normaliser noise. Charity 200124, ABINGER CONSOLIDATED CHARITIES, cites company 01648018, which Companies House names ABINGER HAMMER VILLAGE SCHOOL TRUST. Charity 200359, MEATH EPILEPSY CHARITY, cites 05822835, named THE MEATH TRUSTEE COMPANY LIMITED at Companies House. Charity 200264, LINGFIELD UNITED CHARITIES, cites 08296570, named LINGFIELD UNITED TRUST.

A further 662 of the joined companies are in a distress state at Companies House while the charity remains Registered: 410 with an active proposal to strike off, 223 in liquidation, 25 in administration, and 4 in other insolvency states. The charity register asserts Registered; the company register asserts the corporate shell is dying; nothing reconciles the two claims.

Finding five: the LEI system versus the live register

117,324 GLEIF golden-copy records name UK Companies House as their registration authority. 16,627 of them cite company numbers absent from the live file, and the honest reading of that number matters: most of those are RETIRED or LAPSED LEIs of dissolved companies, which is the system working exactly as designed. The candidate defect set is the 127 records that are entity ACTIVE with registration ISSUED yet cite a number absent from the live register. Decomposed: 82 cite BR-prefixed UK establishment numbers, a genuine Companies House series that the free bulk product excludes by design, so they are unverifiable at the open register boundary rather than provably wrong; 9 cite numbers above the highest number in the 1 August file and are almost certainly companies incorporated in the 16-day window between snapshots; and 36 cite format-valid, in-range numbers that simply do not resolve. MENHADEN RESOURCE EFFICIENCY PLC, a live issued LEI, cites 09242421. WEBB TRADERS UK LTD cites 10777356. These are currently certified identifiers pointing at companies that are not on the live register.

Name reconciliation finds 2,666 disagreements among the 100,663 joined records, including stale post-rename records on ISSUED LEIs: ALBION CAPITAL GROUP LLP is now ALBIONVC LLP at Companies House, ESSENSYS PLC is now ESSENSYS GROUP LIMITED, RIVER GLOBAL SERVICES LIMITED is now LIONTRUST SERVICES LIMITED. The LEI annual renewal did not catch the divergence in any of them. And 34 records publish junk in the registeredAs slot itself: ORBIT HOUSING ASSOCIATION LIMITED, an ISSUED LEI, records the mutuals number IP27802R as its Companies House number. The status distribution completes the picture: 61,832 ISSUED, 38,544 LAPSED, 16,742 RETIRED. One in three UK LEIs has lapsed certification.

Finding six: the register's own numbers break the register's own rules

Companies House's bulk file contains 5,912 live companies whose company numbers violate the documented format of 8 digits or a 2-letter prefix plus 6 digits: 5,243 of shape IP10067R, 528 of shape RS0001FI, 93 of shape R0000019, 40 of shape SP015CUS, and 8 stranger shapes including SP3COLLS and IPIP2960. These are genuine live companies, mostly industrial and provident legacy registrations. Any downstream validator that implements only the documented shapes rejects all 5,912.

The PSC snapshot adds the sharpest specimen in the study. The enum value no-individual-or-entity-with-signficant-control, with the second i of significant missing, is the literal production value in 621,705 statement records. Any consumer who validates against the correctly spelled string silently matches nothing. Behind the misspelling sits a re-measurable baseline: 617,557 distinct companies explicitly declare that no individual or entity with significant control exists, and 345,627 of them resolve on the live register, which is 6.07 per cent of all live companies. That is a narrower metric than Open Ownership's published figure of roughly 20 per cent, which counts companies without a named beneficial owner on a broader basis including unidentified-PSC and unfiled cases; ours is the explicit declaration rate alone, re-measurable monthly from the same keyless products.

The findings, graded

All figures are computed from register snapshots pinned to their dates: the Companies House bulk file of 1 August 2026, and the Charity Commission extract, PSC snapshot and GLEIF golden copy of 17 August 2026. The 16-day skew between the bulk file and the other sources is quantified where it matters. All are living systems, so a later run produces different totals while the method reproduces exactly.

The method, transferable

None of this required privileged access or a single API key. The method is the same one we have now run against US bank registers, US securities registers, EU insurance registers, scholarly records, learning standards and enterprise knowledge bases.

First, model identity honestly. An identifier in a register is not a property of an entity; it is a dated claim by a named register. Our OWL model reifies every published value as an IdentifierAssertion carrying the value exactly as published, the scheme the field claims, whether the value conforms to the rules that scheme declares for itself, and whether it resolves in the target register. Each observed failure mode is a class: DanglingIdentifier, PlaceholderIdentifier, CrossSchemeIdentifier, FormatInvalidIdentifier, SeriesOutOfScope, NameDisagreement.

Second, make the schemes self-describing. A SKOS registry declares each scheme's shapes and canonicalisation as data, including the two documented Companies House shapes, the BR branch series the bulk product excludes, the legacy shapes the register itself contains, and the schemes whose values were found squatting in the company-number field.

Third, make defects first-class. Each defect class is one SHACL shape, so the validation report is the findings table, and the layering is the argument: format checks pass placeholders, referential checks catch them, and only cross-register reconciliation catches a name divergence between two internally consistent registers.

Fourth, never trust one computation. Every headline was computed by two independent implementations with different parsers and engines, and they agreed exactly only after the dual computation caught a real bug: the Charity Commission extract is unquoted tab-separated text, 172 charity names contain literal double quotes, and default CSV quoting silently dropped 234 charitable companies. The governance gate recomputes every number set-based and via SPARQL and SHACL over the graph, asserts a known-answer case before computing anything, and fails the build on any disagreement.

What was not measured, and why

The FCA Financial Services Register requires a free API key we did not hold, and we scraped nothing; a keyed run would join firm reference numbers the same way, including for the 1,652 charities that list the FCA as a regulator. The Land Registry's overseas and corporate ownership datasets sit behind a bespoke non-OGL licence and were not downloaded; a licensed run would produce dangling and name-mismatch rates for land-title proprietors, the same shape as the findings above. And the missing-versus-dissolved split inside every dangling count is not computable from the free Companies House product, which excludes dissolved companies by design. The repository states each gap with the same precision as the findings.

What a register operator or a charity should do with this

If you operate a register: declare the scheme of every identifier field you publish, validate against the declared scheme at ingestion, run the referential check against the target register on a cadence, and treat the placeholders and the cross-scheme values above as the test cases they are. If you are a charity trustee or a due-diligence team: the join that loses 4 per cent of charitable companies is the join your screening process is almost certainly running today, and every number above is checkable from your side with open data and the shipped pipeline.

Common questions

Open, reproducible, and free to use

The repository, with the ontology, the SKOS scheme registry, the three SHACL layers, the harvest and finding pipelines in both independent implementations, the query library and the full findings report, is public: code under MIT, ontology and documentation under CC BY 4.0. The findings report lists what could not be obtained as carefully as what could, and everything regenerates from roughly 25 GB of keyless open register data with the scripts as shipped.

Related work: the same register-boundary discipline found every FDIC LEI truncated in the US bank register, the missing CIK-to-LEI crosswalk in the US securities register, and the category both belong to is set out in Register assurance: why every public register fails at its boundary.