Research-Backed Implementation
We do not just "build". We validate. Our delivery models are rooted in rigorous academic and industrial research methods to ensure efficacy and reduce waste.
Our Approach
"Research-backed implementation" means that every technical decision is preceded by evidence gathering. We apply mixed-methods research (quantitative data analysis + qualitative user research) to define the problem space before writing a single line of code. This aligns perfectly with the GDS Discovery phase but adds a layer of academic rigour to the validation process.
Upstream contributions
The same reading that produces the studies below produces defect reports against the standards themselves. We have reported faults in the IATA ONE Record data model, in FIBO, in the W3C JSON-LD specification, in DCAT-US at the General Services Administration and in the USDA National Agricultural Library thesaurus, and contributed SHACL shapes to GLEIF on the maintainers' invitation. Every one is a public thread you can read.
Ontology engineering
The largest single strand of the research below. Register integrity ontologies, standards crosswalks, verification research and language model grounding, indexed in one place with the definitions that go with them.
Tesseract Foundational Research
Our self-funded research programme: open standards, evidence bases and reference implementations built on public data, published in full for independent verification and reuse. Each project ships with a complete write-up covering challenge, intervention, assurance and reusable assets. Browse by topic below.
Defence, security & space5
The Challenge
The UK zero-emission-flight sector has strong policy signals and funding under the Jet Zero Strategy, ATI FlyZero and the DfT-funded, CPC-led Zero Emission Flight Infrastructure (ZEFI) programme, but the data on who is building what, who funds whom, and which technologies gate which pathway sits scattered across press releases, programme pages and reports. That fragmentation is the problem any coordination tool has to solve.
The Outcome
42 entities, 55 relationships and 272 RDF triples validated at zero SHACL violations, with an interactive network view and a machine-readable graph. The three primitives shown, a typed stakeholder graph, a controlled-vocabulary technology model with sourced maturity, and referential-integrity validation, are the data spine of any single-view coordination tool for a fragmented sector.
The Challenge
The UK Defence Investment Plan (June 2026) commits over £5bn to autonomous systems and £7.5bn to a Digital Backbone and Digital Targeting Web, all of which depend on heterogeneous systems and allies sharing data a machine can reason over. On the UK side that vocabulary is the Information Exchange Standard (IES); the built environment runs on HQDM and the National Digital Twin. Both are 4D and share a common heritage, yet no machine-readable crosswalk between them had ever been published.
The Outcome
A shared starting point the field lacked, released so suppliers building across defence and built-environment data can start from something concrete. The divergences record names the traps, including the ies:Event to hqdm:event false friend, that a naive label-based mapping would fall into.
The Challenge
PYRAMID (UK Defence Standard 00-134) is the MOD open reference architecture for avionics and mission systems. By design it has no shared data model: its Technical Standard states that components "do not share interface definitions", so a deployment "will use bridges to close the semantic gap", and the accompanying MOD assessment confirms "the PRA does not define a data architecture". The meaning of data across components is left to bridges that are, today, hand-built per deployment.
The Outcome
A shared meaning layer beneath an open avionics architecture, demonstrated on the public standard. No published work grounds PYRAMID, or the related FACE Shared Data Model, in any upper ontology; this occupies that gap with ontologies the UK government already owns.
The Challenge
The Information Exchange Standard (IES) is the 4D ontology behind UK defence and security data, but authoring IES-conformant RDF by hand is slow and demands scarce ontology expertise. General-purpose language models cannot help: asked to write IES Turtle, a strong 30B code model invents terms that do not exist in the ontology 94% of the time. That hallucination is worse than useless in a standards context, because invalid data that looks plausible is harder to catch than data that fails outright.
The Outcome
A working assistant for IES data authoring, released as a research prototype with its dataset and eval code, so the small community that works with IES has a concrete starting point rather than a closed demo. It pairs naturally with the IES-to-HQDM crosswalk and the open-ontologies validation engine: the model drafts, the symbolic layer verifies.
The Challenge
Policy on the security of connective products (IoT, operational technology, computing devices, networking equipment, and software and firmware) is set against headline incident numbers that cannot be reconciled: no shared denominator, non-additive definitions, and vendor telemetry that is never quantified. The field lacks an honest, machine-readable baseline.
The Outcome
A shared, honest baseline that maps real incidents to product class, vector and impact, and states its own limits. It gives policy work something concrete to build on instead of irreconcilable headline figures.
Government & public services4
The Challenge
Government is raising the evidential bar behind spending: the Evaluation Task Force is shifting from checking that evaluation happens to ensuring evidence shapes decisions, and the Victims and Prisoners Act 2024 and Victims' Funding Strategy push victim services towards outcome-based, accountable commissioning. Most of that evaluation rests on a Theory of Change, which as a prose diagram hides its own weakest links: a causal claim with no evidence behind it looks identical to one with strong evidence, so effort pours into the well-understood links while the untested load-bearing assumption sits unexamined.
The Outcome
A method that turns "name the weak links" from a hope into a mechanical result: the products a commissioner sees are still the familiar Theory of Change diagram, narrative and M&E framework, while the structured graph underneath makes them auditable, internally consistent, and reusable in a full evaluation rather than redrawn from scratch.
The Challenge
Government publishes a great deal of evaluation and asks suppliers to build on it, but the published record has never been mapped as a whole: what kinds of evaluation dominate, who commissions them, and how clearly they state their own methods. Grounding a bid in the existing evidence base requires knowing the shape of that base, which no one publishes.
The Outcome
1,770 classified publications spanning 1996 to 2026. Impact evaluations dominate (886), as the Magenta Book agenda predicts, but only 11% declare a recognisable method in their metadata, so the catalogue itself is largely method-blind: an automated reader cannot distinguish a robust study from a light-touch one without opening every file. The largest commissioners are DfID, DfE and DWP.
The Challenge
Government runs thousands of public consultations and publishes their outcomes, but the estate has never been assembled as a single, queryable corpus: who consults, on what, and whether they publish what they heard back. Consultation-response coding and automated summarisation both need a structured view of that estate to sample and benchmark against.
The Outcome
6,260 consultations with published outcomes, coded by policy area at 98% coverage. Every one has a published outcome page, yet only 77% attach response documents: nearly a quarter close with an outcome carrying no published response analysis. Defra, DfT and MHCLG are the most prolific consulting bodies.
The Challenge
The UK asks public bodies to publish how they use algorithmic and AI tools in decisions affecting people, through the Algorithmic Transparency Recording Standard, the nearest thing to a public register of government AI. It is published one record at a time, which makes each disclosure visible but the pattern across government invisible.
The Outcome
136 records spanning 73 distinct public bodies, led by DSIT, DWP, the Money and Pensions Service, DfE, DESNZ and DBT. The corpus makes answerable the questions no single record can: which parts of government are most transparent about their AI, which kinds of tools recur, and how disclosure is spreading.
Economy, finance & skills6
The Challenge
Skills England publishes the national occupational maps (1,269 occupational standards with their routes, pathways and clusters, SOC classifications, apprenticeship and technical education products, green-jobs themes and progression pathways) through a public API. The data is relational but delivered as documents, has no published schema for the connected object, and buries the SOC crosswalk (the bridge to official labour-market statistics) inside each record. Consumed the usual way, a flattened slice ends up in a spreadsheet and the relationships that make it valuable are lost.
The Outcome
An open, machine-readable ontology of the entire national occupational map (51,355 triples: 1,269 standards, 15 routes, 35 pathways, 172 clusters, a 278-code SOC 2020 crosswalk, 1,313 products and 2,717 progression edges), with a live explorer on the Tesseract Academy for the Public Sector. It is the operational companion to our open PIAAC research on skills and social mobility: PIAAC measures what adults can do, the occupational maps show where those skills lead, and SOC is the join.
The Challenge
Local authorities must repeatedly evidence their housing market for local plans, strategic housing market assessments, stock-condition programmes and viability studies. The raw material, HM Land Registry Price Paid Data, is open and national but arrives as fourteen unlabelled columns and roughly nine hundred thousand rows a year, with no comparable per-authority summary.
The Outcome
Indicators for all 318 local authorities in England and Wales from 760,607 standard 2025 transactions, with local median prices spanning £135,000 to £1.16 million. A comparable, defensible starting point for any local housing-market evidence base, built entirely from open data.
The Challenge
"Structured data beats PDF for AI-driven reporting" is received wisdom in regulatory and reporting circles, but the empirical premise is rarely measured: across the accounts companies actually file, what fraction expose each core figure as a structured fact a machine can read without OCR or layout guessing?
The Outcome
Across 8,856 filings and 899 distinct concepts (median 49 tagged facts per filing), the core balance sheet is genuinely structured: equity exposed in 97.2% of filings, net assets 86.6%, creditors 75.6%. The tail falls away fast and unevenly, with debtors present in only 28%, and that unevenness is the finding: systems must be built on the measured exposure profile, not an assumption of uniform availability.
The Challenge
ESCO is the reference occupation and skills classification across Europe; England's Skills England maintains a separate, nationally specific set of occupational standards behind apprenticeships and technical education. To read English skills data against the international vocabulary you need a crosswalk, and one that is honest about how far each link can be trusted.
The Outcome
114 exact, 281 close, 270 related and 604 unmatched matches (52% any-band, 31% strong). The 604 unmatched occupations are the headline: a measurement of how far England's occupational language sits from ESCO's, and precisely the surface a semantic crosswalk or national ESCO extension must cover.
The Challenge
The FCA warns the public about unauthorised and scam firms through its warning list, a useful safety tool with one structural blind spot: it shows only the firms flagged right now. There is no history, so the trend, the very thing that says whether the scam problem is growing, is invisible to anyone designing fraud-prevention services or detection tooling.
The Outcome
18,224 live warning pages, 15,437 dated, revealing an unambiguous signal: FCA warnings ran at a few hundred a year through the late 2010s, then climbed to 1,616 in 2022 and around 1,900 a year across 2023 to 2025, a near four-fold rise on 2019. Just under a fifth flag clone firms, among the hardest frauds for a consumer to spot.
The Challenge
Employability programmes, skills bootcamps and local growth plans must target the right places, and the question of where the labour market is tightest has a monthly answer in the claimant count. It is open through the ONS NOMIS service, but delivered through a query builder few outside statistics teams use, so it usually is not to hand.
The Outcome
A like-for-like reading for all 374 local authorities, refreshed monthly, with claimant rates spanning 1.1% to 10.1% against a 3.5% national mean, Birmingham highest at 10.1%. Exactly the spread a place-based programme needs before deciding where to act.
Environment & climate3
The Challenge
The UK is rolling out continuous water-quality monitoring on wastewater assets under the Environment Act 2021, and the Environment Agency is exploring whether that data can serve as a regulatory tool. The binding question is data quality: calibration drift, fouled probes, telemetry gaps and transcription errors all masquerade as real signals.
The Outcome
The data is complete and internally consistent (COD ≥ BOD holds on every one of 1,382 rows), yet the battery still surfaces multi-year baseline drift and dozens of statistical outliers per determinand. The SHACL rules reject every physically impossible record and pass every clean one, making the trust verdict auditable.
The Challenge
In January 2026 the UK Government published its National Security Assessment on global biodiversity loss, ecosystem collapse and national security, assessing with high confidence that six strategic ecosystems are on a pathway to collapse with security consequences. The assessment is a one-off written document: it ships no machine-readable dataset, no shared ontology and no standing monitoring mechanism, so its cascade chains cannot be queried, versioned or stress-tested.
The Outcome
A first, correctable occupation of a genuine whitespace: the framing is government-endorsed and the science exists (IPBES Nexus 2024), but no open, computation-ready ontology and evidence base of nature-to-security cascades existed. The three government functions map to concrete methods: assess (a nature-security exposure profile binding open indicators to transmission channels), monitor (early-warning statistics including critical slowing down), and mitigate (typed intervention points per cascade, mapped to National Risk Register, IPBES and Kunming-Montreal levers).
The Challenge
Stakeholder mapping for the environment sector is usually a slide: boxes and arrows, no provenance, out of date the day it is drawn. The relationships that matter, who sponsors whom, who advises government internationally, whose data flows into the national record, which duties are statutory, are exactly the ones a diagram cannot defend, and getting them wrong in a bid or plan is a real risk.
The Outcome
A queryable graph that records the accurate relationship, not the convenient one: it correctly captures that the Forestry Commission is a non-ministerial department not a sponsored body, that NatureScot, NRW and SEPA answer to the devolved governments, and that Local Nature Recovery Strategy authorities are designated under Environment Act 2021 section 104, each carried in a cited basis rather than smoothed over.
Culture, health & science6
The Challenge
Most of the UK's heritage is digitised but not computable. Aerial photography collections holding tens of millions of frames are scanned, catalogued and mapped, yet a researcher cannot query them at scale by space and time or cross-reference them to other archives. This is the gap the Towards a National Collection programme (AHRC / UKRI) and its N-RICH work set out to close.
The Outcome
Measured finding: the collection is far closer to computation-ready than assumed. 100 percent of records already carry a machine-readable footprint and an ISO-8601 date; the one genuine gap is machine-readable rights. The same standard was then applied unchanged to two further national collections (NAPL Canada and WHAIFinder USA), each lifting to the identical NAPH Baseline at zero SHACL violations, with each collection missing a different Baseline field. The records export cleanly to STAC 1.0, GeoJSON and IIIF, and NAPH publishes the previously unoccupied RiC-O to STAC crosswalk that binds archival provenance to spatiotemporal computability.
The Challenge
Agri-environment schemes are the largest source of government funding for the rural historic environment, yet the wider co-benefits of heritage actions, for nature and for people, are real but evidentially fragmented. As nature recovery is delivered faster and under tighter budgets, heritage actions risk being overlooked unless that value can be evidenced.
The Outcome
A published, replicable protocol, an evidence-map and gap framework, and a real spatial analysis showing 84% of monuments at risk are on farmland, where heritage actions act. Together they make the case that heritage actions deliver multi-objective value, for nature and for people, in a form that withstands scrutiny.
The Challenge
Scientific data is widely described as FAIR (Findable, Accessible, Interoperable, Reusable), but the claim is rarely tested at scale against a machine-checkable contract, and "AI-ready" is asserted more often than it is evidenced. Without a measurable baseline, funders and repositories cannot see where the real gap is.
The Outcome
The datasets are overwhelmingly findable and accessible, but 0% are interoperable or AI-ready: 100% lack a machine-readable schema, checksums and provenance. The finding names the real gap precisely, and the paired ontology models the AI-ready layer that closes it.
The Challenge
UK museums are opening their collections as data faster than the data can be made usable. The records come out flat and free-text: the same polymer written four different ways as unrelated strings, relationships between objects trapped inside caption prose, no links to anything. The gap between a collection being published and being computable is real, and until it is measured it stays invisible.
The Outcome
99.9% of 35,172 free-text material assertions resolved to a science-grounded concept (476 raw strings collapsed to 137); 100% of technique and use-domain assertions; 55 verified Getty AAT exact matches; a 485,000-triple graph over all 11,865 objects; and 289 object-to-object variant edges recovered from prose no keyword search could traverse.
The Challenge
When a council commissions a health-and-wellbeing survey it asks residents to rate their own health, the same question the Census already asks the whole population. A survey read on its own is a number in a vacuum; the full-population baseline for that authority exists but is not laid out ready to compare against.
The Outcome
A comparable baseline for all 331 local authorities. Self-reported bad or very-bad health runs from 2.7% to 9.6%, highest in Blaenau Gwent, Merthyr Tydfil and Blackpool, the geography of health inequality made measurable per authority alongside disability and unpaid-care prevalence.
The Challenge
DCMS publishes monthly visitor figures for every museum and gallery it sponsors, one of the cultural sector's best open data sets, yet it circulates as static spreadsheet tables: no baselines, no seasonality context, no way for a trustee or a marketing team to answer 'are we back, and compared to what?'
The Outcome
A working demonstration of the difference between reporting and insight on data every cultural organisation already knows: recovery is uneven across museum groups, and within a single group the site-level trajectories diverge further still, which is where the actionable signal lives.
Commissioned delivery case studies
Contracts commissioned by government and government-funded programmes, each with a full write-up: challenge, intervention, assurance and reusable assets.
Selected Publications & Talks
Grouped by topic. Open a group to see the studies, papers and talks inside it.
Safe & verifiable AI16
Which cells of a sociotechnical AI evaluation matrix can anyone check?
Open research, AI evaluation assurance, 2026
This is a partition of the Ada Lovelace Institute and NICE sociotechnical evaluation matrix into the 39 of 120 cells a published evaluation can settle, the 75 that rest on organisational records and the 6 that are judgements, run against 13 open-access evaluations of large language models for title and abstract screening. A study evidences a median of 20 of the 39 checkable cells. 11 of 13 publish the prompt, 2 of 13 report confidence intervals, 1 of 13 measures performance by language, and 0 of 13 pin an exact model version with a run date.
How Many Passes of HTML Unescaping Should a JSON-LD Parser Do? A Census of the Tranco Top 10,000
Open census, measured against Google's own parser, 2026
In August 2026 Google changed its JSON-LD extraction to apply a single pass of HTML unescaping rather than two, saying the change brought its parser up to JSON and other standards. Two specifications put the conformant number at zero: the HTML Standard never decodes character references inside a script element, and JSON-LD 1.1 states that escaped content remains escaped after processing, with a worked example showing exactly that. One pass is as unspecified as two. Rather than trust the announcement we measured Googlebot against Google's own validator, which confirmed exactly one pass, applied to string values after JSON structure is parsed, so character references cannot inject structure. We then censused the Tranco top 10,000 homepages. Of 2,969 domains publishing JSON-LD, 10 are affected by the change Google announced and 199 contain a value that Googlebot and every conformant parser read differently, a ratio of roughly twenty to one that held steady as the sample grew. Dell and Investopedia are visibly broken inside Google today, confirmed against their live markup, and two Google properties sit in the diverging set. A second pass past the homepage roughly doubles that, to about one domain in six. The operational point is not search ranking. It is that those publishers have structured data that is correct in Google and wrong in every other parser we measured, including PyLD and jsonld.js, the two JSON-LD 1.1 reference processors.
The Open-World Hole: Why SHACL Cannot Catch a Hallucinated Ontology Term
Open benchmark, 2026
SHACL, the default RDF validator, is open-world: it silently passes any ontology term it has no shape for, which is exactly the failure mode of an LLM authoring RDF. Measured on three real vocabularies (schema.org, IES4, and the OBO Foundry's PATO and RO), open-world SHACL validated as conformant every one of 300 data graphs carrying a fabricated term, across 418 fakes. A closed-world vocabulary gate caught all 300 with zero false positives on clean data. The correctness layer for AI-generated knowledge graphs, released reproducible.
Marginal Coverage Is Not Joint Coverage: Copula-Coupled Uncertainty Certificates
Open case study, 2026
Scientific models predict correlated vectors, and calibrating their uncertainty one output at a time gives a per-target guarantee that looks right and a joint guarantee that is quietly wrong. On 1,600 real OQMD materials, a "90%" independent conformal certificate delivers only 79% joint coverage; a coupled certificate restores 90% (22% tighter than Bonferroni) and a Gaussian-copula region matches it at less than half the size. Distribution-free, machine-checkable, reproducible.
Fluency Is Saturated, Correctness Is Not: An Un-Game-Able Scientific-LLM Benchmark
Open benchmark, 2026
Most LLM benchmarks grade on fluency or an LLM judge, so a model that invents a nonexistent gene or ontology term scores as correct. verifiabench grades with a deterministic closed-world oracle instead. Across nine models, including Claude Opus, Sonnet, Haiku and local open checkpoints, raw fluency saturates at 1.00 while verified capability ranges from 0.00 to 1.00 on the same tasks: a local Qwen3-Coder-30B ties Claude Opus at perfect verified correctness, while some equally fluent models invent roughly half their terms. Un-game-able, reproducible against any endpoint.
The Gate You Can Trust: A Proof-Carrying-Action Gatekeeper, Exhaustively Verified
Open reference implementation, 2026
A safeguarded AI permits an action only if it carries a certificate a small trusted component can check. This is a working reference gatekeeper for a bounded multi-agent system, verified exhaustively rather than tested: over 96 reachable states and 1,176 transitions it blocks 672 of 672 unsafe actions and admits 504 of 504 safe ones, at 0.36 microseconds per check with a 31-line trusted core. Without the gate, 57% of actions violate the safety spec.
The Privacy Fix That Quietly Breaks Fairness: One Report Card for Health AI
Open case study, 2026
Privacy and fairness are audited separately, but they interact. On the real UCI Diabetes readmission dataset, one report card runs membership inference and per-subgroup equity on the same clinical model across a differential-privacy sweep. At the tightest privacy, differential privacy costs minority subgroups 2.6x more accuracy than the majority (drops ordered by subgroup size), while the membership leakage it targets is near zero, almost all equity cost for almost no privacy gain, visible only when both planes sit on one card.
Proving the Utility of Large Language Models in Cybersecurity Simulations
Peer research paper with The Alan Turing Institute, arXiv 2608.16422, 2026
Language models write the network itself. The pipeline synthesises YAML topologies for cybersecurity simulation, checks each candidate against the emulator before it is kept, and uses the resulting environments to train and test reinforcement-learning defence agents. Across multiple synthetic topologies, LLM-instantiated Python attackers reached a 94.5 per cent compromise rate at 0.02 to 0.06 seconds per assessment, four to five orders of magnitude faster than the Double Q-learning with prioritised experience replay baseline they are measured against. Co-authored by Stylianos Kampakis and Fabio Rovai with Marcos Charalambides, Theodosis Mourouzis and Chris Hicks of The Alan Turing Institute.
Read the paper on arXiv (opens in new tab)Silence Is Not Assent: What Catches a Wrong AI Classification in Orbit
Open knowledge graph and measurement, 2026
A public 833,403-triple knowledge graph of all 70,122 catalogued space objects, aligned to the Space Situational Awareness Ontology. The ontology declares exactly one exclusion axiom, between two coordinate formats, so 351 of its 353 classes can never reject a wrong AI classification. The instance data can: LogMap 4.0 on this pair reports zero repair conflicts while the catalogue refutes nine of its candidates, catches a category error in its final output and rescues one it wrongly discarded. The threshold sweep also caught a circular measurement in our own earlier version, which is reported rather than buried. Includes the first openly published language model for a space ontology: hallucinated ontology terms fall from 13.81 to 0.06 per output.
Grounded, Not Retrieved: An Ontology-Validated Biomedical Knowledge Graph
Open case study, 2026
A gene-disease knowledge graph built from 40 real Open Targets associations, every edge typed with the Biolink Model and checked by a closed-world vocabulary gate before it enters the graph. The grounded graph has zero SHACL and zero vocabulary violations across 284 triples; an ungrounded twin with one fabricated Biolink predicate passes SHACL but is caught by the gate. Ships with a provenance-carrying hypothesis triage and a one-click reproduction.
Construction Data Standards Cannot Check Your AI: IFC, COBie, Uniclass and BOT Measured
Open crosswalks and measurement, 2026
AI tools now assign Uniclass codes and fill COBie workbooks automatically, and the target standards assert zero disjointness: measured falsifiability 0.00%, so no wrong mapping into them can ever be machine-rejected. Open, argued crosswalks across IFC4, COBie, Uniclass 2015 and the W3C Building Topology Ontology: 49 correspondences, 7 recorded refusals including the bot:Zone false friend and the framed-structures part-whole trap, all 96 identifiers verified against pinned sources, every Uniclass code confirmed against the live NBS service. BOT reaches 80.95% checkability from 9 axioms; IFC4 reaches 11.45% from 2,443.
Is the Semantic Web Counterfactual-Ready? A Tractability Census of Public Ontologies (opens in new tab)
Open dataset and preprint (forthcoming), 2026
The first tractability map of the public semantic web. A compiler certifies whether each of 528 public ontologies can, as written, support a counterfactual ("what if") query. The finding: 56% declare no constraints and are counterfactual-blind, including both of the UK's 4D data standards, IES4 and HQDM, alongside schema.org, CIDOC-CRM, SAREF and GS1. On the 233 that can, a certified counterfactual is computed for each (median 0.19 ms), confirming that the cost is governed by ontology structure, not by the degree-based hardness that would suggest infeasibility. Open, reproducible, and released with a "counterfactual-ready standards" proposal.
Open Governance: Open-Source AI Governance Server (opens in new tab)
Open-Source Tool, Ongoing
An open-source AI governance platform that helps organisations discover, assess, and monitor AI systems against EU AI Act, NIST AI RMF, and ISO 42001 frameworks. Provides automated risk classification, compliance matrices, bias and hallucination monitoring, policy enforcement gates, and audit-ready reporting through 48 governance tools.
The closest thing to a public register of government AI, as a corpus
Open research, 2026
The full set of 136 published UK Algorithmic Transparency Recording Standard records, structured as an open corpus: which of 73 public bodies have disclosed which algorithmic and AI tools. Published on GitHub and Hugging Face under the Open Government Licence.
Teaching an open-source LLM a government information standard: from 94% hallucination to 1%
Open research, 2026
Tesseract Academy fine-tuned Qwen3-Coder-30B on IES4, the UK government Information Exchange Standard for defence and national security data sharing. Term confabulation fell from 93.7% to 1.0%, term conformance rose from 0% to 88.6%, every claim machine-verified against the published ontology. Published openly on Hugging Face.
FAIR Dataset Contracts for Scientific Data
Open research, 2026
An open study of 1,738 real public biomedical datasets across three repositories (EMBL-EBI BioStudies, Dryad, PRIDE): overwhelmingly findable and accessible, but 0% interoperable or AI-ready (100% lack a machine-readable schema, checksums and provenance). Paired with an open, tool-certified OWL ontology that models the AI-ready dataset layer. Open source and reproducible.
Environment & climate5
Can a Parametric Climate Insurance Product Prove It Paid?
Open standard and measurement, 2026
An open vocabulary and twelve machine-checkable rules that audit a parametric climate insurance promise at the three points where it fails: whether the promise is specifiable at all, whether a registered household is payable before a storm, and whether the payout arrived after one. The replay produces a documented false negative from public records. Severe Tropical Storm Nalgae came ashore at 110 km/h, eight short of the 118 km/h that defines a typhoon, and destroyed roughly 67,000 tonnes of mostly rice worth about PHP 1.3 billion. A trigger set at typhoon strength pays nothing for that. Every one of the twelve rules ships with a case built to trip it, and the build report lists what we could not obtain.
Half the Detail Dies on the Way to the Return: What UK Waste Reporting Throws Away
Open crosswalks, measurement and reporting engine, 2026
AI analytics classify waste at the belt in over a hundred categories and the industry sells that granularity as regulatory reporting. Every UK channel that consumes composition data accepts between 7 and 47 values. Of 5.907 bits of composition detail, the List of Waste retains 53.8%, pEPR 52.1%, RAM 2027 49.5% and Simpler Recycling 38.4%. The crosswalk cannot be a function in either direction: it collapses up to 23 classes onto one value, 7 classes fix no regulatory value at all, and the packaging schemes cannot represent 19 to 22 of them. Of ten operational questions, two are answerable in every channel and four in none, including the aluminium against steel split that sets the pEPR fee and for which the List of Waste offers a single code. Ships a reporting engine that emits returns as intervals rather than point estimates, and a release gate that caught a language-tag defect hiding the most-used municipal code in the upstream ontology.
From raw museum records to a knowledge graph
Open research, 2026
A reproducible pipeline turning the Museum of Design in Plastics' raw open catalogue (11,865 objects, CC BY 4.0 via the Museum Data Service) into a standards-based knowledge graph: a SKOS materials taxonomy grounded in polymer science, a CIDOC-CRM instance graph, verified Getty AAT alignments and a variant graph recovered from description prose. 99.9% of 35,172 material tags resolved; zero SHACL violations. Published in Open Ontologies under CC BY 4.0.
The Value of Agri-Environment Heritage Actions
Open research, 2026
A self-initiated, open-data demonstration of how we evidence the wider value of agri-environment scheme heritage actions for nature and for people: a scoping-review framework on Natural England\'s NEER001 method, with a real spatial analysis of the 2,206 Scheduled Monuments on the Historic England Heritage at Risk Register 2024.
Mapping the UK Zero-Emission Flight Ecosystem
Open research, 2026
An open, SHACL-validated knowledge graph of the UK hydrogen and electric aviation ecosystem: 42 real entities and 55 relationships across organisations, airports, programmes, projects, funders and technologies, validated at zero SHACL violations with enforced referential integrity. Open source via Open Ontologies.
Industrial and freight standards assurance4
An overview of ontology quality issues in air cargo data standards, tested against IATA ONE Record
IATA ONE Record audit, dual-engine verified, English and Korean, 2026
An independent, dual-engine audit of the IATA ONE Record data model ontology across six releases. Every one of the 496 properties in the 2022-12 release declares an rdfs:domain. None of the 522 properties in 2023-12 do, and none of the 534 in 2024-12 do. Classes with no rdfs:label rose from 1 to 58 over the same span, and total lint issues rose from 2 to 650. Verified two independent ways, with rdflib and with the open-source Open Ontologies engine, which agree exactly. Written for Korean air cargo operators building on ONE Record, in English and Korean, with a reproducible check over the MIT-licensed public repository.
An overview of ontology quality issues in air and ocean freight data standards, tested against IATA ONE Record and DCSA
IATA ONE Record and DCSA audit, English and Traditional Chinese, 2026
An independent, dual-engine audit of the IATA ONE Record data model ontology across six releases, written for Taiwan's air and ocean freight sector. Every one of the 496 properties in the 2022-12 release declares an rdfs:domain. None of the 522 properties in 2023-12 do, and none of the 534 in 2024-12 do. Classes with no rdfs:label rose from 1 to 58 and lint issues from 2 to 650. Includes a contrast with the DCSA ocean container OpenAPI specifications, 174 files under Apache 2.0. Verified two independent ways with rdflib and the Open Ontologies engine. In English and Traditional Chinese.
An overview of ontology and schema quality issues in machinery data standards, tested against MTConnect and the Asset Administration Shell
EU Machinery Regulation readiness, MTConnect and AAS audit, English and Korean, 2026
A readiness assessment for Korean machine tool and equipment exporters against EU Machinery Regulation 2023/1230, which applies from 14 January 2027, verified against the EUR-Lex primary text at Article 54 because several vendor summaries give the wrong date. Audits the two semantic carriers the sector relies on. MTConnect publishes each version as both XSD and JSON Schema, and the JSON AlarmStateEnum admits INSTANT while the XSD AlarmStateType does not, identically in 2.0, 2.1 and 2.2. Includes a retracted finding: an initial count of 111 divergences collapsed to one after 105 proved to be a deliberate design difference and three were our own matching artefact. Also tests and refutes the common claim that AAS submodel compliance requires a paid ECLASS subscription. In English and Korean.
An overview of ontology and schema quality issues in machine tool data standards, tested against MTConnect and the Asset Administration Shell
EU Machinery Regulation readiness, MTConnect and AAS audit, English and Traditional Chinese, 2026
A readiness assessment for Taiwan's machine tool cluster against EU Machinery Regulation 2023/1230, which applies from 14 January 2027, verified against the EUR-Lex primary text at Article 54. Audits MTConnect and the Asset Administration Shell. The MTConnect JSON Schema AlarmStateEnum admits INSTANT while the XSD AlarmStateType does not, identically across 2.0, 2.1 and 2.2. Includes a retracted finding, an initial 111 divergences reduced to one verified defect. Counts 1,529 IDTA IRI references against 18 ECLASS and 9 IEC CDD, refuting the claim that compliance requires a paid dictionary subscription. In English and Traditional Chinese.
Security, defence & space7
Consistency Is Not Coherence: Orientation Search for Certified Alignments Between 4D Defence Upper Ontologies
Long paper accepted at the Ontology Matching workshop (OM 2026), co-located with the International Semantic Web Conference 2026, Bari. arXiv 2608.21914, 2026
Three upper ontologies sit under UK and NATO defence data infrastructure: the Information Exchange Standard (IES), the Higher Quality Data Model (HQDM) that underpins the National Digital Twin, and Basic Formal Ontology (BFO). No public alignment between IES and HQDM existed. Promoting a hand-curated seventeen-correspondence crosswalk to OWL and reasoning over the complete merged ontologies with HermiT gives three results. Published HQDM is natively incoherent, with 39 classes that can never have a member before any mapping is applied. Reading the crosswalk as strict equivalences leaves the merge consistent, so the usual check passes, while adding 100 further empty classes and 218 conservativity violations: consistency is the property that is tested and coherence is the property that matters. The repair is not deletion but orientation, treating the direction of each correspondence as the variable the reasoner solves for, which yields a 21-axiom certified bridge and settles the IES-to-BFO question in favour of bfo:history rather than material entity. Measured against LogMap, LogMapLt and its repair debugger, and against a language-model oracle in base and IES fine-tuned form: both models fall for the same ies:Event to hqdm:event false friend.
An Open Ontology for Space Object Catalogues, Tested Against CelesTrak and the General Catalog of Artificial Space Objects
Open ontology, two-catalogue measurement, 2026
An open OWL 2, SKOS and SHACL ontology and reproducible audit of the boundary between the two open catalogues of objects in Earth orbit, built from keyless downloads of CelesTrak's SATCAT of 70,292 objects and Jonathan McDowell's General Catalog of Artificial Space Objects across all four of its catalogues, both pinned to 18 August 2026. GCAT marks 22 entries as corresponding to no real object, giving its reasons in its own words including radar error, cataloging error and duplicate, and CelesTrak carries all 22; one of them, catalog number 11006, has no decay date and therefore counts as a tracked object still in orbit. 1,094 objects GCAT records as no longer tracked carry no CelesTrak data status code, which is a finding rather than a scope difference because CelesTrak maintains that field and applies it to 1,292 other objects. 20,198 of 34,814 on-orbit objects, 58.0 per cent, have no published radar cross section, so no size or mass can be derived from the open record. CelesTrak's owner code collapses the Soviet Union and the Russian Federation into a single value where GCAT separates 16,142 objects from 9,016, and 157 objects are attributed to the United States by one catalogue and to New Zealand by the other. All 70,292 COSPAR designators are well formed with no collisions, a hypothesis that died and is reported as such. Every headline computed two independent ways.
Space debris metrics that cannot be added up: an open composition checker
Open research, 2026
The two premier orbital debris engineering models disagree by a factor of 2.3 to 3.0 at the 1 cm collision-risk threshold, and by almost two orders of magnitude in the sub-millimetre regime, because they partition the same physical population along orthogonal axes: MASTER-8 by generative source, ORDEM 3.1 by material density. Every indicator built on either model silently inherits that partition and never declares it. An open, MIT-licensed composition checker built on QUDT, SOSA/SSN and PROV-O that refuses invalid combinations with a reason and states what may be done instead.
IES to HQDM: an open 4D ontology crosswalk for defence data
Open research, 2026
The first public crosswalk between the UK Information Exchange Standard (IES) and HQDM, two 4D upper ontologies. Open SSSOM and RDF correspondences, a curated divergences record, SHACL validation, and a worked SAPIENT-node safety case grounding autonomy in IES-typed world states. Supports the Defence Investment Plan interoperability and autonomy-assurance agenda.
Grounding a PYRAMID avionics bridge in IES and HQDM
Open research, 2026
PYRAMID (Def Stan 00-134) is an open avionics reference architecture with no shared data model; it pushes interoperability into inter-component bridges. A worked, open, SHACL-validated example grounds those bridges in the IES and HQDM 4D ontologies, so three PRA components (Geography, Tactical Objects, Data Fusion) that model the same object resolve to one referent. No prior work grounds PYRAMID in an upper ontology.
Nature-Related Security Risk: an open evidence base and systems ontology
Open research, 2026
A self-initiated, open, machine-readable evidence base and systems ontology that operationalises the UK National Security Assessment on biodiversity loss and ecosystem collapse (2026): 14 documented nature-to-security cascades graded with the assessment\'s own confidence ratings, an OWL ontology (NSRO), a SKOS taxonomy cross-walked to IPBES and SHACL shapes, published on GitHub under CC-BY-4.0.
Cyber Incidents affecting Connective Products: an open evidence base
Open research, 2026
A self-initiated, open, machine-readable evidence base of cyber incidents affecting connective products (IoT, operational technology, computing devices, networking equipment and software/firmware): 16 documented incidents (2022 to 2025), a SKOS taxonomy and a source-quality rubric aligned to the six Government Data Quality Framework dimensions, published on GitHub under CC-BY-4.0.
Register assurance21
An open ontology for biodiversity registers, tested against the Biodiversity Heritage Library, ZooBank, and the Natural History Museum Data Portal
Open ontology, register and rights census, 2026
A measured census of the identifier and rights layer beneath biodiversity AI, run against complete harvests of the Biodiversity Heritage Library bulk export, the NHM Data Portal, and ZooBank with its public mirrors, with WoRMS and IPNI as same-day controls. On the day of census the Official Register of Zoological Nomenclature had no working machine interface: 50 of 50 sampled canonical act URLs returned 404, its declared publication endpoint is dead, and its freshest public mirrors are 518 and 1,327 days stale and disagree with each other by 79,420 records. In the BHL export, 86.0 per cent of 329,129 distinct digitised items carry no machine-actionable licence, 84 of them defeated by a single invisible Unicode character, ten titles carry the literal string Array as their registered DOI, and the item table's primary key is not unique, a defect our verification gate exposed when its two independent computations refused to reconcile. The NHM Data Portal mints a resolving DOI for every one of its 294 datasets, best in class among the catalogues this programme has measured, while eight datasets clear a five-property product-readiness bar. Every headline computed twice, set-based and by SPARQL, with a gate that fails on disagreement.
An open ontology for UK trade remedy measures, tested against the TRA case register, the UK Trade Tariff and legislation.gov.uk
Open ontology, full-chain census, 2026
A complete census of the UK trade remedy chain, from the Trade Remedies Authority case register to the duty charged at the border. All 21,008 commodity codes in the UK Trade Tariff were visited: 405 carry 9,385 distinct measures citing 49 legal instruments, with 990 additional codes assigning duty rates to named exporters. The tariff truncates the legal instrument locator at 200 characters, and all thirteen locators sitting exactly at the cap return 404 while every shorter locator resolves, which leaves 1,477 measures, 15.7 per cent of the live book, publishing a broken link to the instrument that gives them force. The case register separately drops leading zeros from chapter 3 commodity codes, so twelve published codes resolve to nothing until the zero is restored, each then landing on a trout product consistent with its case. Both defect classes are repaired in the repository, with each repair accepted only where it strictly extends what was published. Of 982 named exporters, 955 do not resolve to exactly one entity by name, and both EU regulations still cited as the legal basis of live UK measures are no longer in force in the EU. Every headline computed twice, set-based and by SPARQL, with a gate that fails on disagreement, and the build report records the claims that died under checking alongside the ones that survived.
The Quantum Pipeline That Checks Its Own Homework Twice: An Open Benchmark for Collateral Optimisation
Open ontology, quantum optimisation benchmark, 2026
A hybrid quantum optimisation pipeline encodes its business constraints twice, once as penalty terms in the QUBO the quantum routine consumes and once in the classical checker that validates the results, so when both are written by hand from the same prose rule book they can share a misunderstanding and the check still passes. Qollateral removes that failure mode by construction. The collateral rule book is data, written in Turtle under an OWL 2 and SKOS ontology in which every rule declares whether it admits an exact quadratic encoding and which algebraic form that encoding takes, and one compiler produces the CP-SAT integer model while a second produces the QUBO from the same source. A concentration bound of one unit per issuer compiles exactly to a pairwise penalty and a bound of two is refused at load time rather than approximated. Every solver run is then certified twice, once by a Python certifier working from raw attributes and once by SHACL shapes re-deriving feasibility from reified allocation assertions, with eligibility itself derived three ways. On a seeded grid of eight instances between 8 and 20 binary variables, CP-SAT proved every optimum in at most 0.0034 seconds while QAOA at depths one to three on a noiseless simulator returned a certified feasible best sample in 20 of 24 runs and the exact optimum in 11, with sampled feasibility mass falling from 0.19 at 8 variables to 0.0002 at 19 as routed depth grew from 43 to 663. An agreement battery of 120 cells across two rule books, four sizes and fifteen seeds found zero disagreements on all 98 feasible instances. The classical routes dominate at these sizes, which is reported plainly: the contribution is the frozen protocol and the resource accounting that any later hardware claim must clear.
EU Law Cites More Than Its Own Citation Graph Records
Open research, EU regulatory dependency graph, 2026
EUR-Lex embeds a machine-readable citation relation, WORK_CITES_WORK, inside every instrument's Common Data Model notice, carrying CELEX, ELI and Official Journal identifiers. It is the only structured source of the dependency graph between EU legal instruments, which makes it what regulatory compliance tooling is built on. Measured against the instruments actually named in the operative text of GDPR, NIS2, DORA, MiFID II and the AI Act, it records 63.4 per cent: 78 of 213 citations appear in the text and not in the graph. The shortfall is concentrated rather than general, with GDPR at 93.8 per cent, DORA 85.3 and MiFID II 80.9, against the AI Act at 51.5 and NIS2 at 38.0. NIS2 matters most because of where its gap sits: the Directive defines its own scope through Annexes I and II by pointing at other legislation, and those pointers are what is missing, so a tool built on the structured metadata rather than the text may not surface the legislation a company must read to establish whether it is in scope at all. Because EU legal documents are reusable commercially under Commission Decision 2011/833/EU, this dataset holds full text rather than addresses, unlike our ISO work. Three corrections are published in full, each of which produced a more dramatic and wrong answer first, and all three are covered by regression tests.
The Only Public Crosswalk to ISO 42001 Names Zero of Its Controls
Open research, AI management system assurance, 2026
A measured audit of what a buyer can actually establish about an ISO/IEC 42001 certificate from public evidence, built with the same address-not-text discipline as our ISO/IEC 27001 work so that no ISO text is reproduced anywhere. NIST publishes the only public domain crosswalk into the standard, and it holds 281 pairs across 72 AI RMF subcategories and 66 distinct ISO addresses, of which 202 point at Annex B implementation guidance and 79 at mandatory clauses. Annex A control references: zero. Annex A is the normative control catalogue and the Statement of Applicability is a declaration about it, so any tool ingesting that crosswalk to build an SoA is inferring the critical hop without showing its work. The same document is titled ISO/IEC FDIS 42001, targeting the Final Draft rather than the published 42001:2023, which is the second NIST crosswalk into an ISO management system standard we have found addressing something other than the current edition. Separately, ISO/IEC 42001 published in December 2023 while ISO/IEC 42006, which sets out what a competent certification body must demonstrate, published on 7 July 2025, leaving roughly nineteen months in which certificates were bought and sold under no published scheme for certifying the certifiers. The complete UKAS public register contains zero occurrences of 42001 across 9,278 records, nine months after BSI announced the world first UKAS accreditation for exactly that standard. One per objective breakdown was deliberately not published because secondary sources disagree and do not reconcile to 38.
Nobody Can Check an ISO 27001 Certificate, and the Registers Are Built That Way
Open ontology, certification register measurement, 2026
An open OWL 2, SKOS and SHACL vocabulary for management system certification, built so that it reproduces no ISO text at all: standards are modelled as reference addresses rather than as text, and a SHACL layer rejects any graph in which a standard element carries a label, definition, note, comment or any literal longer than 24 characters, so the absence of protected expression is demonstrated rather than asserted. The measurement behind it is that every link in the chain a buyer would use to check a certification claim is broken. IAF CertSearch, the global register of roughly two million accredited certificates, states that no user may view or download a list of certificates issued by a certification body and that all verification must be of known entities, so a certificate can be confirmed and never discovered and absence cannot be observed. UKAS publishes 2,921 accredited organisations openly and separately publishes a taxonomy of standards, but 0 of the 2,921 records carry a standard term and ISO/IEC 27001 does not appear in that taxonomy at all, because it holds conformity assessment standards that govern the assessor. 21.6 per cent of the UKAS scope vocabulary is duplication, with 3,336 terms carrying 2,617 distinct names and sampling existing as 15 separate terms. The public domain NIST SP 800-53 to ISO/IEC 27001 mapping returns 404 and last addressed the superseded 2013 edition, while the maintained CSF 2.0 references cover 70 of the 93 Annex A controls and carry five identifier defects, two of which name clauses that do not exist in the standard. The licences that actually constrained the build belonged to the Secure Controls Framework and CIS Controls rather than to ISO. Every headline computed twice, with five dead hypotheses and three of our own bugs recorded in the build report.
USDC Has a Different Upgrade Key on Every Chain, and Its Own owner() Names None of Them
Open ontology, cross-chain control census, 2026
An open OWL 2, SKOS and SHACL ontology for on-chain control, in which control is never a property of a contract but a dated assertion carrying an explicit evidence class, measured across 6,316 contracts on seven EVM chains plus a Solana leg. Of the 656 upgradeable contracts whose upgrade authority resolves, 390 return a different account from their own owner() function than the account that can replace their code, because on a transparent proxy that call is delegated to the implementation. For 556 of the 656 the proxy admin slot is not the controller either, since the path continues past it: only 100 resolve in one hop, 536 need two, 18 need three and 2 need four. Walking every path to its terminal node gives 335 multisigs, 132 single private keys, 109 timelocks and 7 contracts that renounced control, and eight of those multisigs have a threshold of one, which puts effective single-signature control at 140 of 656. Thirty-seven contracts are upgradeable through storage slots outside EIP-1967 and are reported as immutable by tooling that checks only that slot. Native USDC carries six different bare-key upgrade authorities across six EVM chains with no signer reuse, while the same issuer runs two separate 2-of-4 on-chain multisigs with disjoint signers for mint and freeze on Solana. The vocabulary marks where chain state stops answering, and a SHACL rule refuses any link from an account to a legal entity that is not backed by an off-chain attestation. Every headline computed twice, with a build report listing four defects found in our own code.
The Register That Publishes Its Own Contradiction: An Open Ontology for Mass Spectrometry Reference Libraries
Open ontology, three-register measurement, 2026
An open OWL 2, SKOS and SHACL ontology and reproducible audit of the public mass spectrometry reference libraries, built from keyless downloads of 2,091,754 GNPS2 library spectra across 98 libraries, 20,052 MassIVE datasets and 1,065,831 ReDU sample rows, all taken on 21 August 2026. GNPS is unusual in deriving chemical structure twice per record, once from the depositor's SMILES and once from their InChI, and publishing both, which makes a register's disagreement with itself countable rather than merely suspected. On 1,485,480 records carrying both keys they diverge 14,012 times: 13,112 at the stereochemistry layer and 662 at connectivity. Independent recomputation with RDKit cuts that connectivity figure to 229 and reclassifies the other 418 as failures in the register's own derivation step, a different defect that we report separately rather than folding into a larger number. The corpus rate of 0.94 per cent conceals a forty-twofold spread, from 1.41 per cent in MONA to 59.44 per cent in BILELIB19, so the per-library figure is the only one that describes anybody's actual situation. 164,137 InChI strings, 9.81 per cent of those populated, lack the mandatory prefix and are rejected outright by standard chemistry toolkits; 6,710 CAS Registry Numbers fail the check digit CAS itself defines; and not one of the 2,091,754 records carries a valid SPLASH. 145 of the 4,418 ProteomeXchange accessions MassIVE publishes resolve on neither of ProteomeXchange's two surfaces. ReDU is the counter-example that matters: 100.00 per cent conformant across 977,015 taxonomy values with zero non-conformant ontology indices, which shows controlled identity in metabolomics is being achieved one layer away from the libraries that lack it. Six hypotheses died during the build, five of them our own errors, and all six are reported. Every headline computed three independent ways.
An Open Ontology for Italian Public Registers, Tested Against IPA, ANAC, OpenCUP and ISTAT
Open ontology, five-register measurement and upstream SHACL contribution, 2026
An open OWL 2, SKOS and SHACL ontology and reproducible audit of Italy's public register fabric, built from hands-on downloads of the IPA register of 23,735 public administrations, the ANAC contracts register, the OpenCUP register of 11.9 million public investment projects, both flavours of the ISTAT register of administrative units and the national semantic vocabulary, taken on 17 August 2026. The national cities vocabulary and the ISTAT CSV export both predate the 2026 Sardinian territorial reform: 380 dead municipality codes are published as current and 378 current codes are missing, so joins against the semantic layer silently drop every Sardinian municipality, while the operational IPA register already uses the new codes with zero pair mismatches across all 23,735 entities. All 23,735 IPA codici fiscali pass their checksum, a clean null result stated as a finding. In a seeded two-month sample of 2025 contracts, 140 well-formed CUPs referenced by 255 contracts are absent from the published OpenCUP exports, and 45 values in the CUP field fail the CUP grammar, including ESENTE and ATT000NON000CUP. The findings are delivered upstream as a contribute-first SHACL package answering the vocabulary maintainers' own request in issue #225, including the currency test that would have caught the drift in CI. Every headline computed two independent ways, reproducible from public data.
Measuring Peat on a Scottish Landholding, Tested Against the National Cadastre and the National Habitat Map
Open pipeline, 8.1 billion pixel comparison, 2026
The Land Reform (Scotland) Act 2025 requires a land management plan addressing carbon for every holding of 1,000 hectares or more, and in Scotland carbon mostly means peat. Sizing it is an overlay of the national cadastre on the national habitat map, and three independent defects compound against the answer. Registers of Scotland publishes no areas at all, the INSPIRE areavalue attribute carrying the literal string UNPOPULATED on every one of 1,564,345 parcels. Area computed from the published geometry depends on the method chosen, summation exceeding the geometric union by 1,867,044 hectares or 28.37 per cent because overlapping polygons are valid in the data model. And NatureScot's habitat map declares a NoData value of 0 that neither GeoTIFF actually encodes, so 7,205,895,872 of 8,100,000,000 pixels, 88.96 per cent of the national grid, present themselves as a valid class and unclassified ground inside a boundary is silently counted rather than excluded. Against that, the habitat map tests well: comparing all 8.1 billion pixel pairs of the EUNIS Level 1 and Level 2 maps, 98.6761 per cent roll up correctly and the 1.3239 per cent that do not, 118,374 hectares, are every one of them the same marine and coastal confusion across only four class pairs, with the peat class Q1 clean, no undocumented codes and perfect co-registration. Scotland's raised and blanket bog covers 1,330,616 hectares, 14.88 per cent of classified extent. Every cadastral headline computed two independent ways.
Nordic Health Dataset Catalogues Measured Against HealthDCAT-AP Release 7
Open ontology, five countries, eleven catalogues, 2026
The European Health Data Space obliges health data access bodies to publish a machine readable national dataset catalogue, and the European Commission publishes HealthDCAT-AP as the profile those descriptions should follow. On 25 August 2026 the eleven Nordic national catalogues harvested by data.europa.eu held 2,811 dataset descriptions carrying the EU health theme, and not one satisfies all eight properties Release 7 makes mandatory on dcat:Dataset. Three of the eight, the identifier of the health data access body, the health category and the structured data flag, are present on exactly zero across all five countries, so the gap is a system change question rather than a data entry one. The generic layer is meanwhile in good order, with title and description on all 2,811 and publisher on 2,785. Finland is a case of its own: it contributes 1,146 themed descriptions and not one uses the European data theme authority vocabulary, so a European health filter returns zero Finnish datasets while avoindata.fi's own search returns 57. Sitra's February 2026 proposal for a Finnish Health Data Space names that compatibility gap in prose and its text mentions DCAT zero times. Separately, 36 IRIs are in use inside the EU data theme authority namespace and the authority defines 14, with 1,238 datasets carrying one of the other 22, including 806 on the literal string undefined; the authority host answers HTTP 200 with an empty graph for terms it never defined, so a status code check cannot see any of it. Findata's catalogue holds 89,368 variable descriptions, the most detailed health metadata in the region, of which 84.74 per cent carry no English label. Four of our own claims were wrong before they were right and all four corrections are published. Every headline computed two independent ways, with SHACL as a third path.
UK Health Data Linkage Registers, Measured at the Discovery Layer
Open ontology, two national registers, 2026
The UK government and Wellcome are investing up to 600 million pounds in a Health Data Research Service on the premise that value comes from linking health datasets together. On 28 August 2026 we measured the machine readable layer that premise stands on, across the HDR UK Gateway and the NHS England Data Uses Register. The Gateway Data Model field that ties a dataset to its data uses is empty on all 1,278 active datasets, while the platform's own relational layer links 1,704 of 2,565 data use rows to datasets, so the linkage is knowable and the standard's field for it carries none of it. 101 of 324 published dataset-to-dataset links resolve to nothing, including 35 descriptors with no identifier, no URL and no title, every one of which validates. All 59 pid values inside linkage entries are dataset names rather than identifiers, because the schema types the linkage pid slot as a 150 character string while a dataset's own identifier is typed Uuidv4. The two registers of UK health data uses share an identifier for 68 of 2,565 rows, leaving 96.96 per cent of NHS England's 1,942 agreement families with no identifier-joinable Gateway entry, and four agreements sit active on the Gateway while absent from both current NHS publications. The NHS register is itself internally clean, with 100 per cent scheme-conformant references and zero dangling cross-sheet references across 103,706 release rows, which is the point: the failures live between registers, and 1,277 of 1,278 active records validate against the schema that cannot see any of this. Every headline computed two independent ways, with SHACL as a third path.
Medicinal Product Registers, Measured for Joinability: EMA, openFDA and GLEIF
Open ontology, three registers, two jurisdictions, 2026
ISO IDMP exists so that a medicine, its substance and its holder can be identified the same way everywhere, and industry is staffing data standards roles with posted ranges up to 305,000 dollars to comply. On 28 August 2026 we measured what the public can actually join across the three open surfaces: the EMA medicines report, the openFDA NDC directory and the GLEIF legal entity register. The EMA register publishes its 392 marketing authorisation holders as free text with no organisation identifier; joined to GLEIF by exact legal name, 102 resolve and 290 do not, and the mechanism is spelling, with AbbVie Deutschland GmbH & Co. KG failing solely because GLEIF stores GmbH And Co. KG. Of the joins that succeed, 14 of 102 land on an LEI that is not in good standing, AbbVie Limited among the eleven lapsed. The EU and US registers join by holder name for 8.2 per cent. Internal hygiene is high on both sides and its defects are precise: one malformed product number in 2,732, one truncated ATC code, four prose strings sitting in a code column, zero NDC scheme violations in 137,521 records against 1,904 NDCs listed twice and 1,103 labeler-name variant groups. EMA SPOR, the master data service that should carry the join, answers 401 to anonymous requests, so the measurable public surface is the XLSX shadow. Every headline computed two independent ways, with SHACL as a third path, and the error we nearly made is published as the mechanism finding.
Fifty-Nine Federal Agencies Are Validated Against the Schema That Exempts Them
Open ontology, complete federal catalogue estate, 2026
An open OWL 2, SKOS and SHACL ontology and reproducible audit of the US federal open data catalogue estate, built on 21 August 2026 from data.gov's own keyless harvest API. Of 130 DCAT-US harvest sources, 76 covering 59 federal agencies are registered under the dcatus1.1 non-federal validation schema; zero federal-typed organizations are registered under the federal one. The two schemas differ in exactly two respects, both confined to bureauCode and programCode, which the non-federal schema drops from required and additionally permits as null. The affected list includes the Office of Management and Budget, which mandates the bureau code, and the General Services Administration, which operates data.gov. Rebinding alone would not help: the federal schema's pattern is published unanchored, so it accepts xx024:01yy, and the property's own normative example 015:010 contradicts the pattern printed beside it, which is exactly the form the Department of Homeland Security publishes on 1,168 datasets. 997 Department of Energy datasets carry agency code 018, which belongs to the Department of Education. 669 bureau codes name nothing in the OMB register, including the placeholder 000:00 on 401 of the Department of Labor's 404 datasets. 179 of 793 harvest sources completed their most recent job having errored every record they saw while still reporting status complete. Three of our own claims were wrong before they were right and all three corrections are published. Every headline computed two independent ways, with SHACL as a third path.
We Probed Every .gov Domain for DKAN and Found Three Portals, All Publishing the Same Wrong Identity
Open platform profile, complete national census, 2026
A complete census of the DKAN open data portal across the US federal estate, taken on 21 August 2026, and a platform profile that separates defects caused by a platform's shipped defaults from defects authored by publishers, because that distinction decides who can fix them. We probed 20,135 hostnames: all 16,535 registrable .gov domains from CISA dotgov-data, plus 3,600 DNS-resolving open data subdomains. DKAN 2.x answers on exactly three, data.medicaid.gov and data.healthcare.gov at the Centers for Medicare and Medicaid Services and data.sba.gov at the Small Business Administration, and all three publish the catalogue identity http://dkan/data.json, root-caused to line 3 of schema/collections/catalog.json in GetDKAN/dkan. Because a data.json is JSON-LD, that value is the subject every statement hangs from, so merging any two of these catalogues merges their datasets onto one node. A further 368 occurrences of dcat:references and 7 of dcat:_update are minted into the W3C DCAT namespace, which defines neither, by the federal context's @vocab fallback; that one is the context's doing and not DKAN's. Our first census frame returned a false zero because the CISA registry lists registrable domains only and never subdomains, and both frames are published. On the checks their operators control the two CMS portals are clean, with 886 of 886 bureau codes valid.
An Open Ontology for Land Register Integrity, Tested Against All 33 Scottish Registration Counties
Open ontology, complete national cadastral census, 2026
An open OWL 2, SKOS and SHACL ontology and reproducible audit of the Registers of Scotland INSPIRE Cadastral Parcels dataset, built from a complete keyless download of all 33 registration counties, 1,564,345 parcels, retrieved on 21 August 2026 under the Open Government Licence. The Land Reform (Scotland) Act 2025 attaches duties to landholdings of 1,000 hectares or more, and the register publishes no areas: the INSPIRE areavalue attribute, defined as the quantification of the parcel's area, carries the literal string UNPOPULATED on every one of the 1,564,345 parcels, as do referencepoint, beginlifespanversion, endlifespanversion, validfrom and validto, six mandated attributes with one distinct value between them. Because the area must therefore be computed from the published geometry, and because overlapping and stacked polygons are a valid part of the Registers of Scotland data model with hole polygons left in, the answer depends on the method: summing parcels gives 6,581,717 hectares and the geometric union of the same polygons gives 4,714,673, a difference of 1,867,044 hectares or 28.37 per cent, ranging from 1.33 per cent in Nairn to 95.40 per cent in West Lothian, so no single correction factor exists. Every parcel passes the identifier rules Registers of Scotland states in its own specification, zero violations and zero duplicates, a null result stated as a finding. Every headline computed two independent ways, reproducible from public data.
An Open Ontology for UK Public Registers, Tested Against Companies House, the Charity Commission and the Global LEI System
Open ontology, four-register measurement, 2026
An open OWL 2, SKOS and SHACL ontology and reproducible audit of the boundary between the UK's public registers, built from complete keyless downloads of the Companies House bulk file of 5,695,465 live companies, the full Charity Commission register, the 12.9 GB PSC snapshot and the GLEIF LEI golden copy, pinned to snapshots of 1 and 17 August 2026. 1,268 of 31,612 Registered charitable companies (4.01 per cent) cite a company number that resolves to nothing on the live register, including 18 charities registered under the placeholder 00000000, 9 under 01234567 and 5 under 12345678, all of which pass the format rule and none of which resolve. 1,082 charities disagree with Companies House about their own company's name, and 662 charitable companies are in a distress state at Companies House, including 223 in liquidation, while the charity remains Registered. Of 117,324 GLEIF records naming UK Companies House as registration authority, 36 live issued LEIs cite format-valid, in-range company numbers absent from the live register, 2,666 disagree with Companies House on the name, and one in three has lapsed certification. Companies House's own bulk file carries 5,912 live companies whose numbers violate the documented format, and its PSC snapshot misspells its own enum value in 621,705 production records. Every headline computed two independent ways, reproducible from public data.
An Open Census of UK Operational Surveillance Reporting, and What Its Identifiers Actually Denote
Open ontology, complete publisher census, 2026
A companion to our census of the FSA alert register, asking the prior question: can the report that carries a surveillance finding be cited at all? Every publication held by APHA, UKHSA and Cefas on GOV.UK was censused on 18 August 2026 through the open Search and Content APIs, giving 6,897 publications indexed and 2,704 report pages carrying 12,769 published files. Two of our own hypotheses died and are reported: every page does carry a persistent content_id, and the report-level unique_reference field is populated on 38.62 per cent of files rather than never. What survives is that 10,582 files, 82.9 per cent, share their only persistent identifier with every other file on the same page, with up to 165 under one identifier, so a specific edition cannot be durably cited. Of 8,758 dated editions, 5,212 carry no report-level identifier and 242 recurring series carry none on any edition. The governance finding is that the UKHSA weekly national flu and COVID-19 surveillance series ran at 99.2 per cent identifier coverage in 2023 to 2024 and 0.7 per cent in 2025 to 2026, a range of 98.5 points across consecutive seasons of one series, which makes the practice a habit rather than a rule. A field named unique_reference is non-unique on 204 values after excluding 362 legitimate format variants. The dual-computation gate caught a 14-page duplication in our own harvest before publication.
The Missing Pathogen Crosswalk: An Open Ontology for One Health Biosurveillance Registers
Open ontology, complete register census, 2026
One Health is an integration claim, and it can be tested by asking whether a record in one surveillance register can actually be joined to a record in another. Measured across the Food Standards Agency's complete open food alert register, all 1,348 alerts published from 9 January 2018 to 10 August 2026 under the Open Government Licence, the answer for most pathogen alerts is no. The register maintains 27 allergen concepts, granular enough to distinguish macadamia from pecan and gluten-free oats from oats, and 5 pathogen concepts whose scheme declares its last modification as 4 September 2017. Of 234 alerts naming a pathogen, 57 name it in prose with no concept to join on. Pathogen coding was zero across all 188 alerts issued in 2018, including the register's first record about Salmonella in pulled pork, and has never exceeded 19.7 per cent in any year, while allergen coding has run between 47.7 and 64.6 per cent every year without exception. Four organisms appear with no concept available to code them: Bacillus cereus, hepatitis A, Cronobacter sakazakii and norovirus. No pathogen concept carries any external mapping, and all five denote at genus or species level while genomic food chain surveillance identifies isolates at serovar and sequence type, so a record coded "Salmonella spp" cannot be joined to one about Salmonella Typhimurium ST34. The study separates three failure modes with three different remedies, ships the NCBI Taxonomy crosswalk the register lacks with every taxon resolved two independent ways and zero disagreements, and computes every headline twice with a build that fails if the two paths differ.
The Missing CIK-to-LEI Crosswalk: An Open Ontology for US Securities Entity Registers
Open ontology, four-register measurement, 2026
An open OWL 2, SKOS and SHACL ontology and reproducible audit of the boundary between the SEC's EDGAR entity register and the Global LEI System, built from complete downloads of the SEC's entire entity register and every open register that embeds or is embedded by it, all taken on 17 August 2026. The lei field the register carries on every one of its 981,355 records is populated in 773 of them, and only 667 of those are valid LEIs: the remainder includes telephone numbers, IRS employer identification numbers, entity names typed into the identifier field and thirty-five literal "N/A" strings. In a single quarter 1,973 fund registrants stated their LEI to the SEC on Forms N-PORT and N-CEN, and the entity register surfaces 12 of them, while for CIK 892538 it carries the investment adviser's identity on the fund, the same defect class found at the FDIC a week earlier. GLEIF publishes the reverse crosswalk for 27,718 EDGAR-registered LEIs daily under CC0; the SEC publishes no mapping in either direction. Reconciling the two halves across 12,604 fund series finds 100 carrying two different LEIs on the two sides of the register boundary, including a checksum-failing character substitution and three ETF series holding each other's LEIs in a cycle, and 56.6 per cent of all US LEI records are LAPSED. An open OWL 2 ontology, SKOS scheme registry and SHACL governance layer, 526,098 triples, every headline computed two independent ways, reproducible from public data.
Register Assurance: Why Every Public Register Fails at Its Boundary
Research programme pillar, 2026
Public registers increasingly assure their own records, and GLEIF is the standing proof that it can be done: 99.99 average data-quality score across 3.39 million LEI records. But assurance stops at the register boundary. Nothing checks conformance when one register embeds another's identifiers, nothing verifies that embedded identifiers still resolve, and nothing reconciles what two registers claim about the same entity. Measured across six domains with one open method, the boundary fails in the same four ways every time: embedded-scheme non-conformance, resolution failure, cross-register disagreement, and missing governance metadata. All 2,252 FDIC LEIs are truncated and invalid, 19.5 per cent of active EEA insurers carry no LEI, the largest index fund is missing from the open identifier map, the registers of retraction agree on 72.42 per cent, all 67,141 dereferenced US academic-standards identifiers return 404, and zero of 300 sampled GOV.UK documents carry any maintenance commitment. The pillar names the discipline, indexes the six open ontologies as one program, and announces the Register Integrity Index.
Economy, finance & skills20
What 13 Million Triples Reveal About the Quality of US Federal Vocabularies
Open register, reproducible assessment, 2026
Consultancies sell ontology assessment, and almost none has published an assessment of a single named real-world ontology. This register does it in public against 28 vocabularies from 10 US federal publishers, retrieved and hashed on 16 August 2026 and put through 26 checks. The estate holds up better than a gotcha audit would claim: every asset retrieved, HTTPS compliance was complete, and nothing was logically inconsistent. What makes the ten real failures actionable is that every check names the authority it derives from and declares whether failing it violates a specification or departs from a convention the publisher never agreed to. The Library of Congress MADS/RDF ontology does not parse, using rdf:resource on a node element 23 times where the RDF/XML grammar forbids it, and the Thesaurus for Graphic Materials types its 7,782 concepts against it. All 507 ISO 639-2 concepts violate SKOS integrity condition S14, because the vocabulary of language codes carries three preferred labels each and no language tag on any of them. The strain in the estate is governance rather than logic: 21 of 28 assets declare no licence, 20 name no publisher, 14 carry no version. Findings are W3C EARL assertions anchored to dated hashed snapshots, publishers have a documented right of reply, and two flaws in the register's own method were caught and discarded before publication.
Checking the Knowledge Graph Your Pipeline Just Built
Open method, reproducible verification, 2026
A pipeline that turns documents into RDF is making a claim in a formal language, and it has no way to check its own claim. We ran that check against Semantica, an MIT-licensed Python pipeline with about 9,300 stars, at versions 0.6.5 and 0.6.6, by reading every output back through two independent RDF engines and comparing what each one saw. Seventeen defects came out of it, four fixes are merged upstream and six pull requests are open. The failures share a shape worth knowing before you buy one of these systems: the export succeeds, a lenient reader accepts it, and the graph is wrong or largely absent anyway. A two-entity JSON-LD export parsed as 2 triples through a plain graph reader and 21 quads through a dataset reader, with no error on either path, and the 19 statements in the gap were the content. Generated SHACL shapes targeted a namespace the generated data never used, so no shape selected a single node and pySHACL reported conformance on data that violated every constraint in the file. Timestamps written without a UTC offset make a timezone-qualified SPARQL filter indeterminate, which SPARQL raises as an error rather than a false, so Oxigraph drops every affected row and the audit query still returns. The negative result matters too: the one module that serialises through a real RDF library rather than by hand had nothing wrong with it at all.
An Open Ontology for US Bank Registers, Tested Against the FDIC, the Federal Reserve, and the Global LEI System
Open ontology, full-register measurement, 2026
An open OWL 2, SKOS and SHACL ontology for the entity and regulatory-reporting fabric of US banking, measured against the FDIC BankFind register, the Federal Reserve's MDRM dictionary and the complete GLEIF golden copy of 3,403,760 records. Every LEI value in the FDIC BankFind register is exactly 16 characters, truncated from the 20 that ISO 17442 requires, discarding both check digits, so no value can be validated and none matches the Global LEI System as published. Resolving all of them against the golden copy shows why arithmetic cannot repair it: 16 characters puts 6.37 per cent of the entire global LEI population into a collision, and nine FDIC values denote up to six unrelated companies each, across four countries. Two values, confirmed by GLEIF's own consolidation records, identify the wrong legal entity: Associated Bank carries the LEI of Associated Banc-Corp, its parent holding company, while the bank's own active LEI appears nowhere in the register. The Federal Reserve's MDRM dictionary shows the mirror-image defect in the concept fabric: item names and definitions are pure functions of the item code across forms with different consolidation bases, and 27,445 of 47,305 item codes carry no definition at all. The artefact comprises 1,123,634 triples, with every headline computed twice, reproducible from public data.
The Corpus Readiness Index: Is Your Knowledge Base Fit to Feed an AI?
Open ontology, corpus measurement, 2026
An open OWL ontology defining authority as a maintenance commitment, ten SKOS schemes, a SHACL publish gate that exits non-zero in CI, and the Corpus Readiness Index, a seven-dimension score computed from a corpus's own metadata, measured against 69,306 documents across three public estates. GOV.UK guidance scores 78.2: zero of 300 sampled documents carry any field saying who maintains them or when anyone last checked, roughly 55,000 withdrawn pages are still served to any crawler that asks, 4,810 documents are owned only by organisations that no longer exist, and 29.7 per cent carry under 500 characters of indexable text because the answer sits inside an attachment. On the one estate that does publish verification dates, the median document was declared verified 9.38 years ago and edited 1.94 years ago, and 38.3 per cent look current while stale by their publisher's own date. Every number regenerates from committed code.
Science Has No Shepard's: Measuring How Far the Registers of Retraction Disagree
Open ontology, four-register measurement, 2026
Legal research has known since 1873 whether an authority still stands, because Frank Shepard began printing citation labels marking which decisions had been overruled. Science has registers of retraction instead, and they do not agree. We built an open OWL ontology, SKOS status vocabulary and three-layer SHACL suite for the integrity status of scholarly works, then measured it against four public registers. Crossref and Retraction Watch, both published by Crossref since it acquired the database in September 2023, agree on only 72.42 per cent of the retracted DOIs between them. OpenAlex flags 94.5 per cent of retraction notices as retracted research, a category error confirmed independently at 95.95 per cent against Europe PMC's MEDLINE publication types, while Crossref does this for 0.91 per cent of the same notices and Europe PMC for 0.32 per cent, which shows the distinction is achievable rather than hard. The central metadata field is declared a closed 12-value enumeration on a required attribute, yet the live index holds 34 values, 22 of them invalid, including two different misspellings of retraction, a bare integer and a test string reading this_is_some_update_23. Only 19.24 per cent of a 137,243 DOI union is agreed by all four registers. 43,683 citations to the most-cited retracted works post-date their retraction, including 1,171 to the Wakefield paper retracted in 2010 for falsification, none carrying any machine-readable warning. 3.19 million triples, reproducible from public data.
The Content Registry Television Measurement Runs On Covers 1.84 Per Cent of Television
Open ontology, full-corpus identifier measurement, 2026
EIDR is the audiovisual industry's own content registry, and it models properly the distinction measurement depends on: a title is not a cut, and a cut is not an encoding, because nobody watches an abstraction and duration is what completion and frequency are computed from. In the open graph it reaches 53.44 per cent of films and 1.84 per cent of television series. We harvested 224,710 EIDR identifier assertions across 224,182 works and resolved 1,179 of them against EIDR's public registry to learn what each actually denotes. Of the 522 works carrying more than one EIDR identifier, 285 hold identifiers at more than one level of the abstraction hierarchy at once, pairing a title-level record with a specific cut as though they were interchangeable, and a further 133 identifiers are each claimed by two distinct works. There is a trap for implementers underneath all of it: EIDR runs ISO 7064 MOD 37,36 but maps the supplementary value to the digit 0 where the standard convention is an asterisk, so a validator built from the standard alone rejects 6,252 perfectly valid identifiers, one in thirty-six, as corrupt, while across all 224,577 distinct identifiers exactly zero genuinely fail their arithmetic. IMDb mints no season-level identifier at all and covers 0.78 per cent of seasons. The IAB Tech Lab taxonomies, the advertising industry's shared vocabularies for content, audience and ad product, contain no RDF, SKOS or OWL of any kind, and are published here as SKOS for the first time, 2,845 concepts, with the 10 structural defects found in the source reported rather than repaired. schema.org carries 2,987 labelled terms and not one expresses viewership, impressions, exposure or reach. 2,702,154 triples, findings computed twice and agreeing exactly, reproducible from public data.
Every Identifier Underneath American Academic Standards Is Dead, and the Redirect Still Works
Open ontology, three-population census, 2026
The Achievement Standards Network minted the persistent identifiers that United States K-12 academic standards were published under, and that the LRMI educationalAlignment pattern was designed around, so they sit inside learning-resource metadata across the open education web. We dereferenced 67,141 of them across three populations, two of which are complete censuses rather than samples. Every single one returns HTTP 404. The vocabulary that describes them still returns 200, served from a static object store, so a system checking whether the scheme is alive gets a reassuring and false answer. The dependency reaches the standards bodies: 1EdTech, publisher of the successor specification CASE, ships a QTI v3 example package whose curriculum references are dead ASN URIs, and so do DCMI's own worked LRMI examples. Alongside the census we measured what the surviving artefacts can reject. The CEDS Ontology v14 from the US Department of Education declares zero owl:ObjectProperty and zero rdfs:domain across its 2,336 properties, types 965 terms as both an owl:Class and a skos:ConceptScheme, and publishes 19,546 concepts with no broader relations at all; given six specific mis-statements it detects none, five of them being inexpressible in it, while the abandoned 465-triple ASN schema detects one. The corpus fails in both directions at once: 433 identifiers name materially different standards, one attaching four statements about controlled investigations and one about cellular respiration to the same name, while the single most replicated statement text carries 718 distinct identifiers. In the live Georgia CASE package, 2,453 of 2,483 associations are document structure and cross-framework associations number zero. 1,931,913 standard statements across 771 jurisdictions, 21,404,069 triples, reproducible from public data.
An Open Ontology for Insurance and Reinsurance, Tested Against Every Insurer in Europe
Open ontology, full-register measurement, 2026
There was no open, usable ontology for insurance and reinsurance entity data, so we built one and proved it on the hardest available evidence. An open OWL 2 ontology, two SKOS registries and a three-layer SHACL governance suite, run against the complete EIOPA Register of Insurance Undertakings (33,924 rows), joined to a same-day GLEIF harvest of all 3,630 identifiers it files, then cross-checked against BaFin's German register. 643 of 3,304 active insurers and reinsurers carry no LEI at all. Four filed values are arithmetically impossible, including a letter O where the real identifier carries a zero, which quietly disconnects a Danish insurer from the global system. 118 identifiers have lapsed, 42 name entities GLEIF says no longer exist, 283 cross-border passports outlive the authorisation they depend on, and one identifier is filed for both SCOR Global Reinsurance France and SCOR Global Reinsurance Ireland, two distinct reinsurance legal entities. The German cross-register test recovers 14 valid identifiers that the national regulator publishes and the European register does not, while confirming that where both populate the field they agree on all 344 matched undertakings: a coverage problem, not a contradiction problem. Reproducible from public data with five commands, no licensed sources.
The Largest Index Fund Is Missing From the Open Identifier Map: the US Fund Register as a Governance Graph
Open ontology, full-universe measurement, 2026
Fund data is an identifier problem before it is anything else. An open OWL 2 ontology, twenty-scheme SKOS identifier registry and SHACL governance layer, built and run against the whole public US fund universe: 2,316 registrants, 14,841 funds and 43,344 share classes joined to four quarters of Form N-CEN and all 9,119,948 GLEIF ISIN-LEI pairs, every one check-digit validated with zero failures. SEC filings are not as clean: 19 of 14,960 self-reported LEIs (0.13%) fail the same check digit. Only 12.3% of self-reported ETF fund LEIs (497 of 4,053) carry any ISIN in the open mapping, including the Vanguard 500 Index Fund, which has a valid ISIN in commercial data that GLEIF's open file does not carry, and 96.8% of register quotations (29,258 of 30,238) carry no venue field, mostly because the field is ETF-only in the source schema. Class-level ISIN resolution is enclosed behind licensed CUSIP data: 259 of 19,803 funds (1.3%) resolve from public data alone.
Provenance Beats Plausibility: Catching Wrong Financial Answers Without a Gold Key
Open measurement and benchmark audit, 2026
The failure that matters when a model reads a filing is a confidently wrong number wearing plausible provenance. FinanceBench ships 2,400 model answers carrying human correctness labels, an unused supervision set for verification. A deterministic check that never sees the gold answer recovers 57.4% of labelled errors and lifts served accuracy from 68.8% to 78.0% at 60.5% coverage, with positive recall in all sixteen configurations. Two negative results matter more: the same check against the whole filing recovers 3.6%, because a filing carries a median of 1,270 numbers against 69 on the cited page, and excusing answers reconstructible by arithmetic excuses 98.6% of them while the excused are more often wrong than the rest. The audit also finds evidence page numbers are 0-indexed and that the two shipped gold files disagree on 15 of 51 numeric cases, every one by exactly 100x.
FCA Consultation: Stablecoins and UK Crypto Regulation (opens in new tab)
Financial Conduct Authority Regulatory Consultation, 2025
Contributed expert analysis to the FCA's consultation on stablecoin regulation and the future of crypto asset oversight in the UK. Provided evidence-based commentary on regulatory frameworks, consumer protection mechanisms, and systemic risk considerations for digital assets.
Is Blockchain Part of the Future of Art? (opens in new tab)
Journal of the British Blockchain Association (JBBA), Peer-Reviewed
Peer-reviewed research exploring the intersection of distributed ledger technology and the creative industries. Examined provenance tracking, digital ownership, and the implications of blockchain for cultural asset management and intellectual property governance.
Welsh Government Land Valuation Research Report (opens in new tab)
Commissioned by Welsh Government, 2025–2026
Independent research into the feasibility of land value tax models for Wales. Combined statistical analysis of land registry data with international comparator evidence and stakeholder interviews across Welsh local authorities. Findings presented to Welsh Government officials and cited in Senedd committee proceedings.
What Adult Skills Reveal About Social Mobility That Qualifications Hide (opens in new tab)
Open research, OECD Survey of Adult Skills (PIAAC) public-use data, 2026
A reproducible analysis of educational and skills mobility in England using open OECD PIAAC data. Among adults with the same qualification, those from a higher-educated background score about 37 points higher in numeracy, so qualifications understate the advantage of social origin. The data harmonisation is published as an open, machine-readable scheme.
AI Skills for the UK Workforce - Skills England (opens in new tab)
Skills England / UK Government Publication, 2025
Tesseract Academy is cited as an AI training provider and consultancy in Skills England's official research into AI skills for the UK workforce. The publication's methodology included stakeholder workshops with 43 organisations, with Tesseract Academy contributing alongside institutions including The Alan Turing Institute and the Surrey AI Centre.
A live list with no memory: reconstructing the FCA scam-warning signal
Open research, 2026
The FCA warning list of unauthorised firms shows only the current set with no history. This open dataset reconstructs it as a monthly time series from 18,224 live warning pages: FCA warnings rose nearly four-fold from around 500 a year in 2019 to roughly 1,900 across 2022 to 2025, 18 percent flagging clone firms. Published on GitHub and Hugging Face.
Does structured data actually help AI read company accounts? A controlled pilot
Open research, 2026
A self-funded controlled pilot comparing how well an open-weights language model reads the same UK annual reports as structured XBRL facts, HTML text and PDF text. Three FY2026 filed reports, 21 tasks, 189 scored model calls. Financial extraction accuracy: XBRL 88.9%, PDF 86.7%, HTML 80.0%; only XBRL ever retrieved a bracketed negative equity figure with the correct sign, at 2.9x the token cost.
Every government consultation that reached an outcome, as one corpus
Open research, 2026
An open, structured corpus of 6,260 UK government consultations with a published outcome, harvested from the GOV.UK APIs and coded by policy area. 98 percent carry policy-area coding; only 77 percent attach response documents to their outcome. Published on GitHub and Hugging Face under the Open Government Licence.
What government evaluates, and how openly it says so
Open research, 2026
An open atlas of 1,770 UK government evaluation publications, harvested from the GOV.UK Search API and classified by evaluation type and declared method. Impact evaluations dominate, but only 11 percent declare a recognisable method in their metadata. Published on GitHub and Hugging Face under the Open Government Licence.
The most timely local labour-market signal, tidied for every authority
Open research, 2026
An open dataset of the latest monthly claimant count for all 374 local authorities in Great Britain, from the ONS via NOMIS. Claimant rate ranges from 1.1 to 10.1 percent against a 3.5 percent mean. Published on GitHub and Hugging Face under the Open Government Licence.
The baseline a local health survey should be read against
Open research, 2026
An open small-area health profile for all 331 local authorities in England and Wales, combining Census 2021 self-reported general health, disability and unpaid care. Self-reported bad health ranges from 2.7 to 9.6 percent. Published on GitHub and Hugging Face under the Open Government Licence.
Talks & community4
Conformance Over Nothing: Validating a National Metadata Profile Against Its Own Published Artifacts
Accepted for the Practitioners Track, SEMANTiCS 2026, the 22nd International Conference on Semantic Systems, Music Center de Bijloke, Ghent, 17 September 2026
DCAT-US 3.0 is the metadata standard for United States federal open data, and its live repository publishes no RDF binding. A JSON-LD context for the profile survives only in a separate repository that is now deprecated, and that context declares a namespace host which does not resolve. The 115 published example files expand to 76 triples, one predicate and no DCAT at all. The binding is already inside the standard: 26 of 26 classes and 231 of 270 properties still record the RDF term they stand for, retained when the SHACL definition was retired. Generating a JSON-LD context and shapes from those records, and editing nothing in the corpus, the same 115 files expand to 1,510 triples, 123 predicates and 228 DCAT triples. What that does to validation is the demonstration. Run against the corpus as published, the standard's own retired shapes select no focus nodes, report no violations and return conformance. Run against the same corpus once it carries a binding, those same shapes select 228 focus nodes and report 287 violations. The shapes were never wrong, they were pointed at a graph that did not exist. Every figure is regenerated by the build and pinned by the test suite, and every RDF figure is computed twice, by rdflib and by the Open Ontologies engine, with disagreement reported rather than resolved in favour of the convenient answer.
UK Government Business Academy - AI Webinar Series
Department for Business and Trade, Business Academy, 2025 and 2026
Delivered a series of three official UK Government Business Academy webinars on AI adoption for growing businesses, led by Dr Stylianos Kampakis. Topics covered: designing AI roadmaps, choosing the right AI tools using the OCT (Objectives-Capabilities-Tools) methodology, and building internal AI capability including skills-gap analysis and organisational models for long-term success.
The Business Academy recommissioned the series for autumn 2026, taking the programme to six official webinars. The three 2026 sessions run free and online at 11:00 on 24 August, 24 September and 6 October 2026, covering roadmap design, tool-agnostic automation of existing workflows, and what has to change inside a team for AI adoption to hold once the pilot ends.
2025 series
AI Roadmap (opens in new tab)|Choosing AI Tools (opens in new tab)|Building AI Capability (opens in new tab)2026 series, booking open
London Data Week 2026: AI Tools for Everyone, Advancing Disability Inclusion
London Data Week, Shaw Library, LSE, 8 July 2026
Hosted at the London School of Economics on Wednesday 8 July 2026, Tesseract Academy convened a public session bringing together speakers from technology, disability, research and policy to explore how AI can advance accessibility and support for people with diverse disabilities. Held in the Shaw Library, LSE, and hosted by the LSE Data Science Institute with the King's Institute for Artificial Intelligence, with speakers from Vision Ability CIC, Imperial College London, King's College London and UCL.
London Data Week 2025: AI Tools for the Visually Impaired (opens in new tab)
London Data Week, co-organised with Vision Ability CIC, 2025
Co-organised a public workshop and demonstration on making AI tools accessible to people with visual impairments. Delivered at Chabad Islington Community Centre as part of London Data Week 2025, in partnership with Vision Ability CIC.