DCAT-AP-UK: DCAT Application Profile for the UK
Help Site map

About

Design and governance

Updated 5 sections on this page are new or updated since 1.2.0-rc.1 (2026-11-06). Everything that changed

How the profile is governed and changed, the decisions behind it, what is known to be wrong or missing, how it is checked, and an honest self-assessment.

StatusUpdated

Proposed. Not endorsed. To become a standard it needs, in order: review by the Metadata Implementation Working Group, a decision by the owning team in the Government Data directorate of the Government Digital Service, and endorsement by the Data Standards Authority Steering Board, as the xGov MEM had. Until then the xGov MEM remains the endorsed model.

Checkpoint

Intent
Evolve the xGov MEM into DCAT-AP-UK, resolve its backlog, align to DCAT 3, DCAT-AP 3.0.1, the 2026 guidance and the future-state strategy, and present it in a portal that joins the estate.
Arc
  • Read the xGov MEM source, its 99 issues and pull requests, the 2026 guidance, the Data Asset Management Policy, the Data Marketplace schemas and the future data architecture.
  • Wrote the profile as one YAML source, with codelists, lineage and worked examples.
  • Generated SHACL, JSON Schema, JSON-LD context, LinkML, SKOS and CSV; built the browser validator; held all three validators to the same verdict on every example and deliberate break.
  • Built the portal and its knowledge layer, compared the four portals like for like, and joined the estate compass as its fourth bearing.
  • 1.1: scanned DCAT profiles in Europe, the US and sectors; added nine sector extensions, cross-border flows, a quality score and levels, AI-ready and quantum-ready features, an adoption toolkit, bulk and OpenAPI converters and an MCP server; linked the future pan-UK metadata architecture.
Verdict
Meets the brief as a proposal. Needs user research, specialist review of the K mappings and the governance steps before it can be a standard.

Versioning and releases

Changing the profile

  1. Ask: Anyone raises a change request with the form. It asks for the user need, the evidence and the impact on existing records.
  2. Assess: The maintainers label it, check it against DCAT, DCAT-AP and the backlog, and say whether it is a major, minor or patch change.
  3. Discuss: Major and minor changes go to the Metadata Implementation Working Group for review.
  4. Decide: The owning team decides. Major changes go to the Data Standards Authority for endorsement.
  5. Release: The change is made in profile/dcat-ap-uk.yaml with a test, passes the release gate and ships in the next release.

Request a change

Namespace and identifiersUpdated

Proposed persistent identifier (issue 13). The redirect was submitted to w3id.org as pull request 7100 on 6 November 2026; until it is merged, these identifiers do not resolve.

How it is checked

Checks on every change
CheckWhat it provesWhere
Profile integrityThe loader refuses a broken profile: unknown properties, bad obligations, cardinalities that contradict obligations, orphan properties, email addresses in text, broken backlog links.tests/test_profile.py ProfileIntegrity
Three validators agreeEvery example passes, and every deliberate break fails, in SHACL, JSON Schema and the browser validator.tests/test_profile.py ValidExamples, DeliberateBreaks
Every rule has a breakNo rule exists without a record that breaks it.tests/test_profile.py test_every_rule_has_a_break
CoverageEvery open xGov MEM issue is registered, every xGov MEM property is mapped, and every DCAT-AP mandatory property is mandatory here.tests/test_profile.py Coverage
PagesUnique ids, links that resolve, no email address anywhere, no external scripts or fonts, a Content Security Policy, captions on tables, colophon and self-describing metadata on every page, no deck or film files.tools/check_site.py, with its own deliberate breaks
ExtensionsEvery sector extension’s example passes and every break fails in all three validators; an extension that weakens the base is refused.tests/test_toolkit.py Extensions
ToolkitQuality score, levels, CSV and OpenAPI converters, exports and usage policies; browser Turtle and build RDF are isomorphic.tests/test_toolkit.py
MCP serverStarted and called over stdio: tools list, checks with and without extensions, errors.tests/test_toolkit.py MCPServer
In a real browseraxe-core with no serious or critical issues, no horizontal overflow from 320 to 1920 pixels, contrast in light and dark, keyboard, no console errors, and the builder, bulk checker, OpenAPI converter and checklist working.tools/check_browser.py

DecisionsUpdated

ADR 0001: A new repository, not a rename Accepted

Context
Issue 78 asked whether to rename the xGov MEM repository. Its discussion said a completely new solution suits a new repository.
Decision
Start a new repository (first cyberdudeuk/dcat-uk-ap, renamed cyberdudeuk/dcat-ap-uk on 6 November 2026; GitHub redirects the old name). Keep every xGov MEM slot in the crosswalk. Propose archiving the old repository with a pointer, once the owners agree.
Consequences
Old issue links keep working. Records written to the xGov MEM convert with the validator.

ADR 0002: Generate from one YAML file with our own script, not LinkML Accepted

Context
The xGov MEM used LinkML and MkDocs. Ten of its twelve open pull requests were dependency updates to that toolchain, and a LinkML limitation blocked issue 63.
Decision
One profile file; a small Python generator for SHACL, JSON Schema, JSON-LD context, SKOS and the portal. A LinkML file is still generated for continuity.
Consequences
Fewer dependencies to patch. CI checks the generated LinkML file with LinkML's own validator and linter, so it cannot drift from the other validators (D-029).

ADR 0003: Contact form preferred over email in contact points Accepted

Context
Issue 71 found the contact point was defined as an email address. Email addresses go stale when people move, and can be personal data.
Decision
A contact point gives a contact form, a team mailbox or both, and never a named person. The form is preferred. The portal publishes no email address at all.
Consequences
Existing records with team mailboxes still conform. Records with personal addresses should be changed.

ADR 0004: Obligations for critical data assets as conditionals Accepted

Context
The April 2026 guidance asks for more for critical data assets than for other data.
Decision
A critical data asset flag switches on extra mandatory properties, in the same record, rather than a second profile.
Consequences
One record serves the catalogue and the return. SHACL and JSON Schema both express the rule.

ADR 0005: EU frequency list and data.gov.uk themes Accepted

Context
DCAT-AP uses the EU frequency vocabulary; the xGov MEM used Dublin Core terms. Issue 49 proposed the data.gov.uk topics for themes.
Decision
Use the EU frequency list, converting Dublin Core values. Use the 14 data.gov.uk topics for themes, mapped to EU data themes.
Consequences
UK records harvest into European tools. The mappings are the author's and need review (open register O-04).

ADR 0007: Sector extensions that only add or raise Accepted

Context
Sectors need more than the base (health, geospatial, statistics, transport, AI, data spaces, energy). DCAT-AP solves this with separate extension profiles; the future pan-UK architecture plans domain profiles too.
Decision
Extensions in one file, merged onto the base at build time. An extension may raise an obligation, narrow a cardinality, add a pattern, a property or a class; the build refuses one that weakens the base. Each gets its own shapes and schema.
Consequences
A record that conforms to an extension conforms to the base. Sector terms from other namespaces are cited, not copied; terms this profile adds stay provisional (O-13).

ADR 0008: Cross-border flows as two properties and a rule Accepted

Context
Data moves to the Crown Dependencies, Overseas Territories, Commonwealth and partners. UK GDPR restricts personal data leaving the UK; the ICO updated its guidance on 15 January 2026.
Decision
Record every place the data is held or sent, and the transfer safeguard. Rule R-08 requires the safeguard when personal data leaves the UK. The jurisdiction list is open, in ISO 3166 form.
Consequences
The record states the basis; it is not legal advice and not a transfer risk assessment. The adequacy column must be kept current (O-15).

ADR 0009: One validator implementation, used everywhere Accepted

Context
The quality score, levels, converters and exports could be written twice, in Python for the build and JavaScript for the browser, and drift.
Decision
They live once, in site_src/validator.js. The build, the MCP server and the browser all run it; Python checks only what SHACL and JSON Schema check.
Consequences
Building needs Node. Tests compare the browser Turtle with the build RDF so the two paths cannot drift.

ADR 0010: Quantum-ready means ready to migrate Accepted

Context
Post-quantum migration is a long programme (NCSC: discovery by 2028, completion by 2035). Metadata cannot make data quantum-safe.
Decision
Record algorithms as identifiers, warn on weak checksums, and record connection security on data services so the catalogue helps the inventory.
Consequences
The pages say plainly what the profile does not do.

ADR 0015: A requirements stream for a family of profiles, with controlled layers kept out of the public repository Proposed

Context
The same needs keep coming up in engagement with catalogue owners, data asset management, discovery services, defence and security users, policing and geospatial teams: a common core, specialist layers, separate handling for records and data, explicit missing values, strict machine validation and honest mappings. Some of those needs can only be specified with material that is not public, and publishing who asked what would aggregate information about protected communities.
Decision
Keep one public requirements catalogue that names kinds of community only, labels each requirement with its evidence (engagement, verified public source or operating rule) and traces it to properties, rules and tests. Build the open core and its extension points here; list requirements needing controlled material with status Controlled and specify them elsewhere. Add record-level classification and explicit missing values to the core now.
Consequences
The requirements page counts how many are met, partly met, planned or controlled. C-22 fails the build on any protective marking or caveat, and a test limits the catalogue to kinds of community. No conformance claim is made for any allied or controlled specification.

ADR 0016: Share the build, change-flag and self-check patterns of the connected solutions Accepted

Context
DCAT-AP-UK links to and is linked from the Data Quality Commons, the DAM capability playbooks, the GitHub Repos Asset Catalogue and the Knowledge Web. They share one way of building a portal (generated pages, a retractable drawer, display options, a colophon, checks with deliberate breaks) but had drifted: only the Data Quality Commons flagged what changed, only the DAM playbooks published a status page and only the catalogue ran a repository self-assessment. A reader moving between them met different conventions.
Decision
Adopt the three missing patterns. Change flags as the Data Quality Commons does them: a section baseline recorded at each release (tools/section_baseline.py), with New and Updated flags on drawer items and groups, on section headings, under page titles and in the global bar, and a list on What's new. A Status page as the DAM playbooks build it, fed by a health run on every build. The five self-* routines (verification, assessment, staleness, reporting, repair) as the catalogue reports them. Not adopted: GitHub Discussions voting and the ask-a-question assistant, which need the owning team to decide on moderation and hosting first.
Consequences
C-24 fails the build if a flag disagrees with downloads/changes-since.json; status.json and changes-since.json are machine-readable. Someone runs section_baseline.py at each release so flags reset; until then they accumulate. Flags count words only, so restyling never flags a section.

ADR 0017: Routines do the upkeep; a person keeps control Accepted

Context
Keeping the portal right took a person for chores: resetting the change flags at each release, refreshing the organisation list after machinery of government changes, watching sources, chasing issues and reading many separate signals. Chores that rely on someone remembering get missed, as D-031 and D-032 showed. The same is true across the connected portals.
Decision
Every workflow is declared in ops/automation.yaml with a tier: 0 observe, 1 maintain (changes derived or reference data on its own, only after every check passes, in one commit naming the routine), 2 propose (a pull request or issue for a person) and 3 person only (releases, settings, deletions, anything sent outside the repository, claims). tools/routine.py runs every tier 1 routine the same way. The change flags reset straight after each release; upkeep runs weekly; triage labels waiting issues; a weekly digest issue is the owner's one place to look. A repository variable, AUTOMATION=off, stops every tier 1 change. Across the estate, a weekly conductor reads each repository's digest and status, writes an estate digest and opens pull requests for a person; it never merges.
Consequences
A person presses one button to release and reads one digest a week. tests/test_automation.py fails if a workflow is undeclared, a tier 1 routine ignores the kill switch, an action is not pinned to a commit or permissions are not declared. Routine commits are signed off with a Routine trailer, so the digest lists them and each is one revert to undo.

ADR 0018: Run the estate as flight operations: test flights with an envelope, near misses reported without blame Accepted

Context
ADR 0017 made routines do the upkeep, but said nothing about trying new things safely or learning from what nearly went wrong. A conductor metaphor describes coordination, not safety or experiment. The owner asked for a model that trims waste, prevents incidents and pre-empts threats while still allowing frontier work inside guardrails with kill switches.
Decision
The estate is run as flight operations, after aviation safety practice: routines are flights inside an envelope (the tiers), the digest and Status page are the instruments, a weekly AI agent is operations control (dispatch and watch, never fly), and a person is pilot in command. Threats are named before they bite, after the Threat and Error Management framework (threats, errors, undesired states). New ideas fly as trials in ops/automation.yaml: each states a hypothesis, an envelope, a measure, an abort rule and an end date at most 90 days away, may only observe or propose, and has its own kill switch (TRIAL_ plus its name) as well as AUTOMATION. Near misses are reported without blame as occurrence issues, after the just culture principle, and each closes with a check or a recorded reason for none.
Consequences
The build fails if a trial is malformed, touches anything kept for a person, runs over 90 days or passes its end date, so no experiment is left running unowned (tests/test_automation.py Trials, with negative tests for each guardrail). The digest lists occurrences and trials. Nothing new can be merged, released or switched on without a person. The metaphor is a naming aid only: every term maps to a mechanism, and a term without one is dropped.

ADR 0019: A plainer look, and model maps for how things connect Accepted

Context
The portal was rich but busy: a dark masthead with many controls, boxed and coloured headings, bordered cards, loud change flags and long tables on every class page. Readers who are not technical found it harder to use than the xGov MEM site, whose plain pages, left navigation and one diagram per class are easy to follow. The owner asked for that simplicity, every feature kept, and a friendlier way to see how parts of the model connect, after the MEM class diagrams and the Morgan Map pattern.
Decision
Keep the architecture, every page, tool, check and pattern (drawer, tabs, change flags, glossary, display options), and change the look: a white masthead with search up front, quiet navigation, plain headings, light cards and tables, quieter flags that still say New and Updated. Show how classes connect as model maps after Morgan Maps: cards with a type chip, curved links named by the property that makes them, and a details panel with the card's facts and connections. Class pages follow the MEM order: the map first, then the properties as plain tables by obligation, then the detail. The links come from the profile data, plus the structural links DCAT expresses as IRIs, so the maps cannot drift from the model.
Consequences
Every map works without JavaScript (each card is a link and a list repeats every link), carries no inline style (the CSP holds), and is checked: tests/test_integrations.py requires a card for every class a class links to and well-formed panel data, and tools/check_browser.py requires that selecting a card opens its details and Escape closes them, with a deliberate break (map). Busy maps show line labels for the card being looked at. Contrast is checked in light and dark on every page, as before. The old UML-style class diagram is replaced; the property tables carry the same facts as text.

ADR 0021: Vendor connectors from published models only, with each vendor's AI offer beside them Accepted

Context
The owner asked for connectors, add-ons and crosswalks for the products in the Gartner Magic Quadrants and Forrester Waves on data catalogues, metadata management and data governance, linked to each vendor's own repositories, with each vendor's AI agent skills and Model Context Protocol (MCP) offer. The reports are paywalled; live connectors need a tenant and credentials, which a static site must not hold.
Decision
content/vendors.yaml registers each report (positions taken only from vendor announcements and abstracts, labelled S, or secondary sources, labelled K), each product, its public repositories at a named commit, its MCP server, tools and skills, and a crosswalk only where a public file defines its metadata model (labelled V). A product with no published model gets no mapping, with the reason, rather than a guessed one. Each mapped product gets an add-on kit in its own extension format and a weaver model; every property without a native field goes in that extension place, so a record survives the round trip. The portal reads vendor repositories and never posts to them.
Consequences
The connectors page and downloads/connectors/ carry twelve products, seven kits and the register. The weaver gains Collibra, OpenMetadata, DataHub, Unity Catalog, Dataplex and data.world, and writes Apache Atlas (Purview, Atlan). tests/test_vendors.py checks the register (with deliberate breaks), lossless round trips of every worked example, that people fields are never carried, and that kits and weaver agree; check_site C-25 checks the page and kits. Nine ranked vendors with no public model are listed for the weekly run. Kits are not tested against live tenants. Extended on 7 November 2026 to cloud, SaaS, on-premises and open-source products (31 in all), each with where it runs, labelled K until checked.

ADR 0020: Say where every fact comes from, and use real records wherever they can be shown Accepted

Context
The portal used invented records and samples to show its features, and some facts were labelled background knowledge (K) with no check behind them. Some invented examples named real organisations. The owner asked for genuine, authoritative content wherever it exists, synthetic content only where it must be, and routines that keep it checked.
Decision
tools/provenance.py builds a register of every labelled fact, worked example, extension example and tool sample on every build, shown on the Status page and in downloads/provenance.json. A fact needs action when it is K or U, or when its V check is undated or more than 180 days old (sources watched by the weekly health run count as re-checked). Worked examples and tool samples use real public records, each with its source and the date checked; a sample written in another metadata model says so. Synthetic content stays only where nothing genuine can be shown in public (an unpublished critical data asset, personal data shared across a border, a conformance claim no real catalogue has made, candidate extensions no publisher uses yet), states why, and never names a real organisation. The upkeep routine keeps one provenance issue listing what needs a person; the weekly operations-control agent verifies up to ten K or U facts a week through a pull request, never relabelling without a source.
Consequences
tests/test_provenance.py fails on unexplained synthetic content, on synthetic content under a real publisher and on an undeclared example address in the tools, with negative tests for each. The census examples and five of the seven tool samples are now real (the FSA hygiene ratings API and CKAN package, the Environment Agency river levels, The National Archives CAB 360, the ONS Living Costs and Food Survey). The DCAT-US sample is still invented and listed for replacement; the Purview sample stays invented because data map entities are not published.

ADR 0014: A global bar that fits a laptop screen Accepted

Context
With eleven items the bar folded About, Downloads and Estate into More below about 1,400 pixels, so most laptop users needed an extra click and could not see where About was.
Decision
What has changed and Strategy become one Context menu. Help leaves the bar: the masthead Help button leads to the user guide, which links the tutorials and roles, and it is marked when you are on those pages. Help stays in the site map and llms.txt.
Consequences
Nine items fit at 1,280 pixels with room to spare, checked in the browser at 1,280, 1,440 and 1,920 pixels (break: fold).

ADR 0011: Name the profile DCAT-AP-UK Accepted

Context
National profiles of DCAT-AP are named DCAT-AP- and a country code (DCAT-AP-NO, DCAT-AP-SE, DCAT-AP-NL, DCAT-AP-DE). The draft was called DCAT-UK AP.
Decision
Call it DCAT-AP-UK, the DCAT Application Profile for the UK, in every page, file and namespace. Keep the dcatuk: prefix for terms.
Consequences
The w3id namespace becomes https://w3id.org/dcat-ap-uk/. The repository was renamed dcat-ap-uk on 6 November 2026, and the site moved with it.

ADR 0012: Integrations through a conformance kit and one reference implementation Accepted

Context
Suppliers need to show conformance without reading the whole profile, and customers need to check claims.
Decision
Publish test records with expected verdicts, level badges, a command line, a local HTTP API and a GitHub Action, all running validator.js.
Consequences
A product conforms when it gives the expected verdict on every kit record. Product entries are background knowledge until each supplier confirms them (O-19).

ADR 0013: Model contract clauses for open standards and exit Proposed

Context
Lock-in comes from proprietary formats, closed interfaces and missing exit terms. The Open Standards Principles and the Technology Code of Practice ask buyers to prevent it.
Decision
Offer clauses in seven groups (open standards, evidence, ownership, exit, change and debt, security, accessibility) as a downloadable file.
Consequences
They are a starting point, not legal advice. Commercial and legal review comes first (O-20).

ADR 0006: A proposed w3id.org namespace Proposed

Context
Issue 13 asked for a persistent identifier.
Decision
Use https://w3id.org/dcat-ap-uk/ for the profile, its terms and codelists.
Consequences
Identifiers do not resolve until the redirect is merged at w3id.org (submitted as pull request 7100 on 6 November 2026).

DefectsUpdated

Defect register
DefectFoundWhatRoot causeFixCheck that stops it coming back
D-0012026-11-06SHACL accepted badly formed values in recommended and optional properties.Property shapes took their severity from the obligation, so value errors in non-mandatory properties were warnings, which conform.Every value constraint is a violation; the obligation only decides whether absence is allowed.tests/test_profile.py DeliberateBreaks (bad-date, summary-too-long, period-end-before-start)
D-0022026-11-06JSON Schema cannot check that an end date is not before a start date.JSON Schema has no way to compare two values.Documented. SHACL (sh:lessThanOrEquals) and the browser validator enforce rule R-07.The fixture is flagged; the test fails if JSON Schema starts catching it, so the flag is removed.
D-0032026-11-06The build library could not be imported in isolated mode and shadowed a standard Python module.It was named profile.py, the name of a standard library module.Renamed ukap.py; scripts add their own folder to the path explicitly.CI runs every script with python3 -I.
D-0042026-11-06The duplicate-title check missed a real duplicate.It read every <title> on the page, including the titles inside SVG images, so two pages never matched.Only the document title is read.tools/check_site.py C-13 and its deliberate break
D-0052026-11-06The table check reported header cells without scope on every page.Its pattern matched <thead> as if it were <th>.The pattern matches the th tag name exactly.tools/check_site.py C-06
D-0062026-11-06The caption check accepted a broken caption tag.It counted the text "<caption" as a prefix, so <caption-x> passed.Exact tag names are counted.tools/check_site.py C-06 deliberate break
D-0072026-11-06The strategy and organisations sections showed raw template placeholders (a variable name in braces) instead of content.A doubled apostrophe (YAML style) inside a Python f-string split it into two strings, the second not formatted.Typographic apostrophes in the build script.tools/check_site.py C-16 (no unrendered placeholders) and its deliberate break
D-0082026-11-06Small download buttons failed contrast in dark mode.The .small text style (muted grey) also applied to buttons with the small size.Small buttons keep the button text colour.tools/check_browser.py axe in dark mode, and its "contrast" break
D-0092026-11-06Loading the critical data asset example into the record builder lost its licence.The example used the compact form of the licence value; the licence list holds full addresses, so no option matched.Lists match compact and full forms of the same value.tools/check_browser.py: the critical data asset example must conform in the builder
D-0102026-11-06Escape did not close search when text had been typed.In a search box the browser uses the first Escape to clear the text.Escape in the search box closes the dialog.tools/check_browser.py palette checks
D-0112026-11-06The lineage said 20 xGov MEM issues were open; there were 22.Miscounted by hand from the API listing.Counted from the backlog register itself.tests/test_profile.py ReadmeCounts: totals must equal the register, and the README counts must match it
D-0122026-11-06The overflow check failed on random pages after a browser update.It measured the width straight after resizing, before the global bar had folded its items into More.The check waits two animation frames for resize handlers before measuring.tools/check_browser.py overflow, with its "overflow" break still caught
D-0132026-11-06The build stopped with a type error while checking codelists.YAML reads the bare code no (Norway) as false, so the code was not text.The code is quoted, and the loader refuses any code that is not text with a message that says why.tests/test_toolkit.py test_codelist_code_must_be_text, and Norway must still be on the list
D-0142026-11-06An unknown property name with a space crashed the Turtle export instead of being reported.The name was used in an address without being encoded.Unknown names are percent-encoded, so the record is kept and validation reports it.tests/test_toolkit.py test_unknown_key_with_space_serialises
D-0152026-11-06The record builder’s Turtle dropped every property of a distribution that had its own address.A nested node with an address was written as the address alone.Such nodes are written as their own statements.tests/test_toolkit.py test_turtle_matches_python: browser Turtle and build RDF must be isomorphic for every example
D-0162026-11-06The adoption page had two elements with the id checklist.The section heading and the form were given the same id.The form is checklist-form.tools/check_site.py C-04 caught it before release
D-0172026-11-06Validating inside the metadata weaver crashed on nested records.A weaver helper was named finish(), the same name as the validator’s own finish(), and replaced it.The helper is finishRec().tests/test_integrations.py Weaver: no two top-level functions in validator.js share a name
D-0182026-11-06The barcode, the knowledge graph and several new pages used style attributes, which the content security policy blocks, so they lost their styling.The browser check turned the policy off (bypass_csp), so it could not see what the policy blocked.Presentation attributes and classes only.tools/check_site.py C-19, and a browser pass with the policy on (break: csp)
D-0192026-11-06Every page scrolled sideways by 4 pixels on phones and tablets.The new page heading band used a negative side margin wider than the page padding at narrow widths.No negative margin.tools/check_browser.py overflow at 320, 390 and 768 pixels (break: overflow)
D-0202026-11-06The standards map hid its links from assistive technology, and some link names did not match their visible text.The map was marked as one image (role img) although it holds links, and long names were cut short on screen but not in the label.The map is a group; each standard has a short visible label that begins its accessible name.tools/check_browser.py axe nested-interactive and label-content-name-mismatch (break: nested)
D-0212026-11-06The quality score bars in the builder and checker were empty.Their widths were written as style attributes in generated HTML, which the content security policy blocks.Widths are set through the style object after the HTML is added.tools/check_browser.py CSP pass (break: meter)
D-0222026-11-06The contrast check could no longer see the text in page heading bands, so poor contrast there would pass.The bands had a gradient background, which axe-core cannot measure, so it marks the text as unchecked rather than failing it.A solid tinted background with a coloured edge, lightened so links in the band meet 4.5 to 1: the blind spot had hidden low-contrast links on 13 pages.tools/check_browser.py contrast (break: contrast, which targets heading band text and was not caught until the fix)
D-0232026-11-06The labels workflow failed on its first run.A label colour made only of digits (942514) was read by YAML as a number.Colours are quoted, and the workflow converts every value to text.tests/test_integrations.py RepoAsCode, with a deliberate numeric colour
D-0242026-11-06Menus in the global bar did not open once the bar had folded them into More, which hid About, Help and others on most laptop and phone screens. On phones, and for About on wide screens, open menus also ran off the edge of the screen.Opening a menu closed every other open menu, including More, its own parent. Menus were anchored to their own item, whatever the room beside it.Opening a menu keeps its parents open; on narrow screens menus span the bar; a menu that would run off the right edge opens leftwards.tools/check_browser.py menus at six widths: every group opens, its first item can be clicked and it stays on screen (breaks: menus, menuedge)
D-0252026-11-06Pressing Enter in a form with one text field, such as Describe an API from OpenAPI, froze the page.The browser tried to submit the form; the content security policy (form-action none) blocked it.No form on the portal submits; the page handles every submit itself.tools/check_browser.py submit: Enter in three forms, no error and no hang (break: submit)
D-0262026-11-06On About pages, and others whose section was folded into More, the bar did not show which section you were in.The current-section marker stayed on the folded item, out of sight inside More.More carries the marker when the current section is inside it.tools/check_browser.py current section at six widths (break: current)
D-0272026-11-06On a phone with the largest text, opening About inside More showed nothing: its links were below the part of More in view.More scrolls when it is taller than the screen, and opening a group inside it did not bring the group into view. Found when the menu check was extended to the largest text, because the shorter bar (ADR 0014) no longer exercised the right-edge rule at standard sizes.Opening a group inside More scrolls its links into view.tools/check_browser.py menus at six widths, standard and largest text (breaks: moreview, menuedge)
D-0282026-11-06The standards map could not draw any standard in the products and connectors group: adding the first one, Apache Atlas, broke the build.The map placed groups from a hand-written list that left that group out. The group had no standards until now, so nothing showed it.The map takes its groups from the data.tests/test_integrations.py Standards: every standard is drawn on the map
D-0292026-11-06LinkML's own validator rejected every worked example, so the LinkML file disagreed with the SHACL, JSON Schema and portal validators.The generated schema had no id slot, treated every link as a bare reference, and listed codes where records use compact or full IRIs. Nothing ran LinkML tools on it.The generator now writes an id slot that is not an identifier, accepts nested objects or IRIs, and keys closed codelists by the forms records use.tools/check_linkml.py, with its --break, and linkml-lint in CI
D-0302026-11-06A person's name, their team and the old name of their directorate were in the citation file, the profile attribution and the governance text, against the rule of roles, not names, and contact by form only.The attribution was written before the rule and nothing checked for it; the directorate has since been renamed.The maintainers are named as a group with the contact form; the governance text names the Government Data directorate in GDS.tools/check_site.py C-23 (personal details matched by hash, so the check publishes nothing), with its deliberate break
D-0312026-11-06The organisation list still showed the Department for Science, Innovation and Technology as live after the July 2026 machinery of government change, and the Information Commissioner's Office under its old name.The list was copied from a snapshot (31 October) that predated GOV.UK recording the change, and nothing checked its age.import_orgs.py reads the GOV.UK organisations API directly with --govuk; the list was refreshed (DSIT closed, the Information Commission added).tests/test_integrations.py OrganisationsFresh: read no more than 120 days before the release, and DSIT closed
D-0322026-11-07Four of the thirteen verified sources were never fingerprinted, so the weekly health run could not tell if they changed. The new Status page showed it.Three are GitHub pages, which the run could not fetch and which change on every load anyway; one is a site that refuses automated checking. Nothing checked that every verified source was watched.GitHub sources are watched through their raw files (watch: in lineage.yaml) and were fingerprinted; the other is marked watch: manual and is re-read by a person at each review.tests/test_integrations.py VerifiedSourcesWatched, and the staleness routine on the Status page
D-0332026-11-07Invented example data was shown under real names: the synthetic cross-border example gave the DVLA as its publisher, the synthetic base record for the sector extensions gave GDS, and the identifier guidance used company and charity numbers that could belong to real bodies.Examples were written to look realistic, and nothing checked that synthetic content stayed clear of real organisations. Two examples also described real things (the census) under invented addresses.Synthetic records now use an invented department and placeholder numbers made of zeros; the census examples now describe the real ONS table TS007a and its series; every synthetic example says why it cannot be real; tools/provenance.py registers every fact, example and sample.tests/test_provenance.py (no synthetic record under a real publisher, no unexplained synthetic content, every sample declared), shown failing with the DVLA put back
D-0342026-11-07Commit IDs cited for vendor repositories were padded from seven characters to twelve with invented characters, and a bulk edit then emptied four citations.Short commit IDs were lengthened by hand to fit a twelve-character rule, and nothing checked a cited commit against GitHub.Citations are full 40-character commit IDs read from the clone, and tools/check_vendor_repos.py fetches every one from GitHub by ID.tools/check_vendor_repos.py in CI (a padded ID fails it), the 40-character rule in the vendor register checks and check_site C-25, and tests/test_vendors.py
D-0352026-11-07The OpenLineage documentation facet written by the weaver named schema version 1-0-1; the specification at the cited commit is 1-1-0.The facet address was written from memory, and weaver output was checked only for round trips, never against the vendors' own schemas.The address now matches the specification, and tools/check_vendor_shapes.py validates every worked example in five formats, and three kits, against each vendor's schema at the cited commit.tools/check_vendor_shapes.py in CI, and tests/test_vendors.py VendorSchemas with ten deliberate breaks
D-0362026-11-14Where each product runs was written from background knowledge for all 31 products, and Secoda was shown as SaaS only, though Secoda documents a self-hosted, on-premises install.The deploy lists were added with one register-level "check later" item, not a source per product, so nothing tied each claim to evidence.Every product now carries its own deploy source with a check date and a note, checked at the vendor's documentation or the repository it pins; Secoda now lists on premises.Site.vendor_problems (deploy list needs provenance; V must cite a pinned repository; open source needs a repository), and tests/test_vendors.py Register with five deliberate breaks plus a Secoda regression test
D-0372026-11-15The first draft of the namespace register linked every namespace, though several do not resolve: the StatDCAT-AP namespace contains a literal "(xyz)" and w3id.org/dcat-ap-uk is not registered yet (O-17).A namespace was treated as a web address. It is an identifier, and only some are published as pages.Namespaces are shown as code, never as links, with a sentence saying so. Caught in review before release.tools/check_site.py C-26 (a namespace cell must not be a link), with a deliberate break
D-0382026-11-15The Adopt page's Words for contracts box showed the build's unfilled template placeholder (the E of pc expression in braces) instead of the procurement clauses, on the live site, and Copy the clauses copied that text. Found when a Python linter flagged the clauses as built and never used. The same lint pass found 66 other findings: dead code (an old Display options form, unused counts), loop variables not bound in a closure, and test comparisons that would accept a validator returning fewer results than expected.The line was not an f-string. Check C-16 removes everything inside pre before looking for placeholders, so the placeholder in the clauses box was never seen. No linter ran in CI.The line is an f-string; the clauses show. C-16 now flags the build's own signatures inside pre as well. Ruff and ESLint run in CI and fail on any finding; the validator comparisons in the tests use zip(strict=True).tests/test_check_site.py plants that placeholder inside pre and fails against the old C-16; tests/test_lint.py keeps both linters in the workflow
D-0392026-11-15Two defects were numbered D-037, and the build on main failed: two pull requests prepared in parallel each took the next free number, and merging them kept two copies of the knowledge page's Taxonomy heading and tab strip (one the linter had changed), so the page had duplicate ids (C-04) and the linter failed.Defect numbers were chosen by hand with nothing to catch a repeat; a pull request was merged while its lint step was failing.The Adopt page defect is renumbered D-038; the duplicate Taxonomy heading is removed, leaving Thesaurus, Namespaces, Taxonomy in the order the namespace change intended.tests/test_check_site.py checks that every defect, open item and closed item in content/portal.yaml has a unique id
D-0402026-11-15Code scanning found five weaknesses. The reference HTTP service returned internal error text to callers (posting null to /weave answered with a JavaScript TypeError message). The Copy buttons put back their label by re-parsing HTML kept in an attribute. A test treated any address containing github.com as GitHub. Two of the build's own HTML-stripping patterns missed upper-case or spaced closing tags.Error text was passed straight through; HTML was stored as text to restore later; host checks by substring; strict closing-tag patterns.The service returns only the validator's own messages about the request (unknown extension or model) and a generic message otherwise, logging the detail on the server; Copy buttons keep and restore their own nodes; the test compares the parsed host; the stripping patterns accept any closing tag in any case. The sixth alert, XML parsed with DOMParser in the DDI reader, is a false positive: nothing parsed is inserted as HTML.tests/test_integrations.py ReferenceAPI asserts a known message is returned and an internal failure is generic; it fails against the previous service

Open registerUpdated

Open register: what design cannot close
ItemWhatWho must act
O-01Test the portal and the profile with at least five people from four of the personas, including disabled people.User researcher
O-02Review by the Metadata Implementation Working Group.Owning team
O-03Endorsement by the Data Standards Authority Steering Board.Owning team
O-04Check every mapping marked K against its source. On 6 November 2026 the EU access-right values and data-theme codes were found at source and relabelled S; the frequency and schema.org mappings are still K.Metadata specialist
O-06A manual accessibility audit with assistive technologies.Accessibility specialist
O-07A Welsh language version of the portal.Owning team, with translation
O-08Research the levels for "who can see this record" (issue 5) and the reference data property (issue 3) with catalogue owners and information assurance. Candidate levels to test, from a concept board retired from Figma on 15 November 2026 (an idea to test, not a finding): your own team; your business area; your business unit; a restricted vault; your organisation; your parent group; partners and suppliers; joint-venture data co-ops; your sector's data ecosystem; an innovation sandbox.User researcher, information assurance
O-11Review the first UK GEMINI crosswalk (adoption toolkit) and the geospatial extension with the geospatial community.Metadata specialist
O-13Check every extension term and namespace marked K with its sector body (health, geospatial, statistics, transport, data spaces, energy, organisation identifiers).Metadata specialist, with each sector
O-14Reconcile this profile with the future pan-UK metadata architecture (co-cddo, private): base DCAT-AP version, namespace, extension names and validator, so there is one DCAT-AP-UK, not two.Owning team
O-15Check the UK adequacy (data bridge) column and the transfer safeguards against the ICO list (triangulated against secondary sources on 6 November 2026; the ICO page itself could not be fetched), and the Crown Dependency and Overseas Territory entries with their regulators.Data protection specialist
O-17Register https://w3id.org/dcat-ap-uk/ (the namespace changed with the rename, ADR 0011; this also covers issue 13) and retire the dcat-uk-ap spelling. The pull request is open at w3id.org (perma-id/w3id.org pull request 7100), waiting for its maintainers; the rules it adds are kept in docs/w3id/dcat-ap-uk/.Owning team
O-19Check the 22 product entries on the integrations page with each supplier before anyone relies on them.Owning team, with suppliers
O-20Have commercial and legal specialists review the model contract clauses before use in any procurement.Commercial and legal specialists
O-21Agree a governance route and an accountable owner for the requirements stream, including who may validate controlled requirements (ADR 0015).Owning team, with data and architecture leadership
O-22Run the community workshops in the proposed work programme and record which requirements each one confirmed, changed or rejected.Owning team
O-23Decide whether data product becomes a class of its own or a stated mapping (RQ-10).Owning team, with data asset management leads
O-25Compare obligations with the statistics producers' application profile and decide every difference (RQ-27). Its public repository was not found on 6 November 2026, so the comparison needs the current version from its owners.Owning team, with official statistics producers
O-26Review the candidate geodesy module with geodetic specialists, including which frames, epochs and uncertainty rules to enforce (RQ-23). The EPSG codes it uses were checked at the EPSG registry on 6 November 2026.Owning team, with geodesy specialists
O-28Decide whether main should require the pages and CodeQL checks to pass. That would stop direct pushes to main, so every change, including automated ones, would need a pull request (GDS Way). The ruleset already blocks deletion and force pushes and requires linear history.Repository owner

The case against, answered

RAID

Risks, assumptions, issues and dependencies
TypeWhat
RiskDepartments see another standard and wait. Mitigation: everything in the xGov MEM still maps, and conversion is automatic.
RiskThe profile is mistaken for an endorsed standard. Mitigation: the status banner on every page and in every file.
AssumptionDCAT-AP alignment is wanted. The European profile is widely implemented by catalogue software, which lowers cost for suppliers.
IssueIdentifiers do not resolve until the namespace is registered.
DependencyThe cross-government taxonomy for themes, from the master and reference data working group.

Theory of change

Problem
Metadata about UK public sector data is recorded differently in every catalogue, so data is hard to find, judge and share, and returns are retyped.
Activities
  • A profile with clear obligations and codelists
  • Shapes and schemas tools can check against
  • A portal that explains it to every role
  • A change process that keeps it current
Outputs
  • Records that conform
  • Catalogues and tools that declare and test conformance
  • Converted xGov MEM records
Outcomes
  • Data found across catalogues with one search
  • Fitness judged before asking
  • Returns built from catalogue records
Impact
Public sector data is found, trusted and reused, supporting a pan-UK public data infrastructure and the National Data Library.
Assumptions
  • Catalogue owners will change exports if the cost is low and the benefit is clear
  • A mandate or endorsement follows a credible proposal
We would be wrong if
  • If fewer than half of the catalogue managers in research can map their fields using the crosswalk, the profile or the crosswalk is too hard
  • If no catalogue declares conformance within two releases of endorsement, adoption needs a different route

Research plan

Measures of success:

Architecture

Architecture
LayerWhat
Sourceprofile/dcat-ap-uk.yaml, profile/codelists.yaml, content/*.yaml, examples/examples.yaml. Hand edited, reviewed in pull requests.
Generatortools/ukap.py loads and checks the source; tools/generate.py writes the machine-readable files; tools/build.py writes the portal.
Checkstests/test_profile.py (three validators, deliberate breaks, integrity); tools/check_site.py (pages); tools/check_browser.py (Playwright and axe-core).
PublicationStatic files on GitHub Pages, from the gh-pages branch, written by the workflow after every check passes.
RuntimeStatic HTML, one stylesheet, one script and data scripts from the same host. No server, no cookies, no third-party requests.

Service Standard self-assessment

Service Standard, self-assessed (not an assessment)
PointStandardStatusEvidence
1Understand users and their needspartialNeeds drawn from 99 backlog items and policy; no research with users yet (O-01).
2Solve a whole problem for usersmetOne record serves catalogue, return and discovery.
3Provide a joined-up experience across all channelspartialJoined to the estate compass; offline use is partial.
4Make the service simple to usepartialPlain English and role routes; untested.
5Make sure everyone can use the servicepartialAutomated WCAG 2.2 AA checks pass; manual audit open (O-06); English only (O-07).
6Have a multidisciplinary teamnot metOne maintainer. An owning team is needed.
7Use agile ways of workingmetSmall releases, open backlog, change requests.
8Iterate and improve frequentlymetRelease gate and weekly health run.
9Create a secure service which protects users' privacymetNo personal data, a Content Security Policy, no third-party code.
10Define what success looks like and publish performance datapartialMeasures named below; nothing is tracked, by design.
11Choose the right tools and technologymetOpen standards, static hosting, few dependencies (ADR 0002).
12Make new source code openmetMIT and Open Government Licence.
13Use and contribute to open standards, common components and patternsmetDCAT, DCAT-AP, SKOS, SHACL; GOV.UK Design System patterns.
14Operate a reliable servicepartialStatic pages on GitHub Pages; one maintainer.

Data protection and algorithmic transparency

Data protection
Screening only. The portal processes no personal data: no analytics, no cookies, no forms of its own. Display choices stay in the browser. The contact form is a GitHub issue form and is public, so it says not to include personal data. Records built in the browser never leave it. No DPIA is needed for the portal; adopters must assess their own catalogues.
Algorithmic Transparency Recording Standard
Not needed. The portal and profile make no algorithmic decisions about people. The validator applies published rules and shows its reasons.

Artefact set

The usual artefact set, with an honest status
ArtefactWhereStatus
Theory of changegovernance.html#tocDrafted
User research bundlegovernance.html#researchPlan only; research not done
Service blueprint and workflowsroles.htmlRole routes drafted; no blueprint tested
Product requirementschanges.html#backlogThe backlog is the requirements; 22 items traced
Functionalprofile.htmlThe profile is the specification
Architecture (SAD, HLD, LLD)governance.html#architectureSummary; one page is proportionate to a static site
Glossaryknowledge.html#glossaryDone
Thesaurusknowledge.html#thesaurusDone (alternative terms)
Data dictionarydownloads.htmlDone
Taxonomies and ontologyknowledge.html#taxonomyDone (classes, codelists, added terms)
Knowledge graphknowledge.html#graphDone
Governance records (RAID, ADRs, DPIA screening, ATRS decision)governance.htmlDone
Accessibility statementaccessibility.htmlDraft until a manual audit
Showcase deck and filmKept outside the repositoryNot made

Search the portal

Keyboard shortcuts

/ or Ctrl K
Search
?
These shortcuts
Esc
Close menus, tooltips and dialogs
Arrow keys
Move between tabs and search results