# Changelog

## Facet touch targets and explicit reference destinations — 27 September 2026

- Keep the compact facet bar proportional to declared counts while moving colour selection into a folded set of full-size, keyboard-operable controls. Preserve highlighted and unknown states, every labelled value in the expanded facet, and the existing multicolour release gate.
- Accept an optional nested, byte-bound reference target with exact destination document, passage, occurrence and leaf identity. Do not infer a destination from the printed row's own IDs or matching numbers.
- Test a skewed distribution with nine one-record values at desktop and medium widths, including the WCAG 2.2 target-size rule, and retain the prior build-bound Heritage evidence before refresh.
- Increase extraction-gap PDF link targets and make scrollable evidence passage text keyboard-focusable after public Reader accessibility checks.

## Reading-help corpus consumer — 27 September 2026

- Add a catalogue-led reading-help view that selects one document and passage, verifies bound gzip leaves and ordered continuations against exact frozen page spans, and keeps candidate explanations separate from source wording.
- Link related body and reference-row occurrences in the existing bounded reading aid only through their shared card occurrence IDs. Preserve its v1 manifest route and source checks.
- Let readers find source-declared paragraph labels, topics and proposed abbreviations within the selected document; expose exact but unresolved printed reference rows, candidate citation labels and browser-session fetch/cache counts.
- Add optional, byte-bound workbench case sidecars that link exact selected record IDs to corpus passages without fetching the corpus until the reader follows the link. Register Chrome, Firefox and WebKit coverage for the new routes and keep diverted CLML responses rejected across browsers.
- Show exact source PDF links for extraction-blocked pages and accept complete source-verified abbreviation tables within a tested 512-literal bound.
- Bind declared sibling child descriptors to the fetched federation location, retaining local and pinned raw commits and stopping when a bound child is missing.
- Serve nested application routes from their own inventoried files in the legislation runtime harness, reject unknown and traversal paths, and retain the failing journey phase in bounded diagnostics.
- Match the legislation acceptance facet checks to the visible coloured bar buttons used by the current Reader, preserving the multicolour and layout thresholds.
- Archive the preceding Heritage receipt and both browser results byte for byte, then refresh the current-build 100-question and three-journey browser evidence.

## Legislation passage Reader repair — 26 September 2026

- Register the focused CLML browser suite in the impacted browser test plan so Reader and source changes exercise it in CI.
- Keep complete normalised CLML passage text, including nested numbering, table and formula wording, repeated wording and endings beyond the 2,400-character excerpt. Add a keyboard-operable control to show the full text.
- Place dated, enacted and made versions after the selected provision in official links; prefer a safe source-native `DocumentURI` and retain source IDs, hierarchy and version markers.
- Bound on-demand CLML download and parsing, reject redirected or credentialled requests, retain the exact source bytes and hash for download, and direct readers to the official source when a limit is reached.

## Curated bounded evidence package research — 26 September 2026

- Add the September BEP briefs, reports and reconstructed Research A package as
  byte-preserved research records, with an exact curation inventory and one
  reader index. Retain original package manifests and archival formats; omit the
  duplicate ZIP and incidental macOS metadata. Preserve original line endings
  in Git so imported checksum manifests remain valid after checkout. Keep original HTML in Git and
  serve the same-stem Markdown reading pages at its Site routes.
- Keep package Markdown outside the default OKF semantic bundle. Label the
  reports as research proposals with unverified citation and execution limits;
  defer Jev-Mem and the historical WebMCP task prompt.
- Record a source-linked code audit of what the current BEP-related engine and
  service implement, and which research schema, storage and assurance proposals
  remain unadopted.

## Occurrence-scoped reading help — 26 September 2026

- Add a generic `/reading-help/` view for a separately governed manifest. It
  verifies frozen page JSON and exact source, occurrence and support spans,
  keeps cross-page text in labelled sections, and opens project-authored help
  only at declared occurrences.
- Distinguish abbreviations from adjoining source markers, retain unresolved
  references and support local passage navigation. Keep byte identities and
  limitations available in folded details. Add links from the Evidence
  workbench and passage-review pages.


## Passage-boundary Evidence workbench — 25 September 2026

- Add the generic `/evidence/passages/` review route and versioned, closed
  manifest, case and correction profiles. Reviewers can inspect hash-verified
  PDF pages beside exact extracted text, navigate the supplied 28-case queue,
  compare parked and successor units, and preview source-bound split, join or
  role corrections without changing a corpus.
- Show parser rules and hash-checked settings in folded provenance, retain
  failed previews, and export local review records with exact input and result
  identities. Keep preview, independent review and producer adoption separate.
- Verify the 28-case source fixture and PDF rendering in a browser. Refresh the
  build-bound Heritage browser receipt, preserving the preceding framework
  receipt and results under `validation/passage-boundaries/pre-workbench/`.
- Keep full document and corpus after-counts unknown because other units and
  IDs are not validated in an isolated preview. Distinguish producer-supplied
  downstream evidence from effects of an edited browser correction. Preserve
  the preceding receipt before refreshing this build's browser evidence.

## Explorer framework and bundler update — 25 September 2026

- Update SvelteKit to 2.70.3, the Svelte Vite plugin to 7.3.0, Svelte to 5.57.0,
  svelte-check to 4.7.6 and Vite to 8.3.0 as a reviewed framework set.
- Pin transitive Rolldown and its native bindings to the reviewed 1.2.9
  release. Keep the earlier security overrides and regenerate the dependency
  inventory from the new lockfile.
- Record the build and browser checks required for framework changes in the
  dependency maintenance guide. Make the Foundry Site browser harness own its
  server as well, so it cannot reuse an older worktree's Site.
- Refresh the Heritage browser evidence for the updated application and bind
  its receipt to the final assembled Site. Preserve the preceding results.

## Dependency security and maintenance repairs — 25 September 2026

- Update Vitest and its matching coverage package to 4.1.11, YAML to 2.9.1,
  Playwright to 1.63.0 and axe to 4.13.0 while preserving the cookie security
  override. Regenerate the dependency inventory from the reviewed lockfile.
- Patch audited transitive devalue and nanoid copies to 5.9.2 and 3.3.18 with
  conditional overrides and checks against vulnerable duplicate copies.
- Add a deterministic dependency-policy guard and regression cases for lost
  overrides, vulnerable copies and mismatched coverage dependencies.
- Update the pinned pnpm setup and Pages deployment actions without changing
  workflow permissions. Document the review, inventory and validation steps.
- Make local browser tests own their server, with an explicit port for parallel
  worktrees, so an older checkout cannot silently satisfy a new candidate's tests.
- Preserve the preceding Heritage receipt and refresh its 100-question and
  three-journey browser results against the exact repaired application build.

## Remote service 0.7.0 public observation — 25 September 2026

- Bind the Explorer service guides to retained DWP hosting and public SDK
  receipts for 0.7.0. Preserve the earlier 0.6.0 observation and separate
  delivery verification from connected-client, Voice and legal acceptance.

## Passage-boundary review design — 25 September 2026

- Record a versioned proposal for generic Evidence workbench passage inspection,
  isolated correction previews and source-bound impact review. DWP's 28 known
  historical-amendment candidates remain acceptance fixtures; no corpus, parser
  or service behaviour changes in this design.

## Remote evidence-connect admission and connection guidance — 25 September 2026

- Prepare service 0.7.0 to admit the newer, immutable DWP evidence-connect
  corpus with a compatible pinned engine and retain historical replay pairs.
- Distinguish remote assembly, saved workbench inspection and browser-host
  support in the connection guides. Record the successful developer-assisted
  Edge sidebar observation without upgrading it to automatic discovery.
- Keep deployment and public verification tied to the shared DWP receipt
  record, separate from implementation and local test results.

## Cross-repository rich-runtime reconciliation — 24 September 2026

- Require the reviewed `okf-uk-government-apis` producer to declare its
  bounded rich relationship runtime, route locator, executable schemas,
  validation receipt and governed `uv` commands. Preserve the Wayfinder source
  as fictional synthetic comparison material with rights not specified and no
  default real-world runtime admission.
- Resolve descriptor, manifest, runtime, locator and chunk references from the
  Explorer publication root. Keep separate compatibility relationship chunks
  outside the exact rich-runtime shard set.
- Add a nested `bundle/` regression fixture and retain complete digest, route,
  schema and Reader-ceiling checks in strict reconciliation.

## Workbench page tools and model inspection — 24 September 2026

- Record the public PR146 workbench observation separately from local checks:
  exact application and DWP manifest hashes, qualified interaction rows,
  blocked calculations and the unavailable native-tool gate in Chrome.

- Add seven optional native WebMCP page tools over shared, bounded workbench
  services, with snapshot-bound references, exact passage continuations,
  cancellation and revision-checked display changes.
- Add renderer-neutral directed graph, interaction, requirements, rates and
  calculation-readiness views. Human controls and tool presentation share the
  same values, source links, row-coverage notices and accessible tables.
- Admit optional blocked model and interaction proposals without changing the
  original retained packages. No rule execution, rate inference or award
  calculation is provided. Document the Java/MongoDB destination separately.
- Keep native registration, real AI-host invocation, panel rendering and
  deployment as distinct verification claims. See the
  [page tools guide](docs/workbench-page-tools.md).

## Evidence workbench and bounded delivery — 24 September 2026

- Add a reusable Evidence workbench for inspecting retained question packages,
  source PDF locations, whole passages, concepts, relationships, dependencies and
  selection traces. Export review proposals locally without changing sources.
- Deliver and reconstruct canonical context packages through responses of at
  most 32 KiB, preserving provenance, gaps and context identity independently of
  the larger assembly budget. Verify hashes, bounds and exact continuations.
- Add opt-in weighted discovery and bounded source-bound navigation routes;
  prioritise the leading discovered unit's dependencies before broad expansion.
  Preserve the existing ranking contract and frozen replay behaviour.
- Retain the original staff question review brief separately from current
  engine findings. Attempt a sandboxed cited-PDF preview with a direct-link
  fallback, and expose the complete machine-readable package.
- Keep candidate selection, useful answers, specialist acceptance and public
  client observations as separate claims. See the [workbench guide](docs/evidence-workbench.md).

## Required-path admission — 23 September 2026

- Refresh the exact candidate browser receipts: 100 Heritage questions and three
  journeys pass. Retain the preceding allocation regression and independent
  review. The independently checked fixed-source DWP trial retains all 47
  inherited and 181 new declared source paths, with insufficiency preserved;
  the next source release and deployed-client checks remain separate.

- Give applicable declared paths priority only as their exact next directed hop
  is encountered from a resolved concept. Missing edges are never supplied by
  requirements; prefix work is explicitly bounded and truncated when necessary.
- Retain integrity-checked units with unresolved source boundaries for inspection
  during v3 path allocation. Preserve their warnings and insufficient status.
  Keep earlier failed comparisons and all legacy replay bytes unchanged.

## Discovery and evidence allocation — 23 September 2026

- Give bounded lexical discovery one fully checked source unit before graph
  expansion, then prioritise resolved-concept paths ahead of the remaining
  lexical candidates. Keep ranking parameters and all resource limits fixed.
- Under v3 byte pressure, explicitly omit unused full relationship rows before
  whole evidence. Preserve selected and declared required paths, every omitted
  assertion and endpoint ID, exact lazy incident metadata, and missing-dependency
  checks. Retain the failed full-corpus observation and unmodified v1/v2 replay.

## Scoped context routes — 23 September 2026

- Add optional conjunctive routing guards to bundle assertions. Evaluate them
  from directly resolved public concepts before loading or following a route;
  retain matched and unmatched reasons without treating a navigation guard as
  legal applicability or turning evidence requirements into retrieval seeds.
- Preserve unguarded replay and exact source metadata. Reject fabricated guard
  diagnostics, report blocked required paths, and keep metadata refusals explicit.
- Retain the preceding current-app receipts before refreshing the guarded
  candidate. Correct their observation time using the actual UTC clock; source,
  application and browser-result hashes are unchanged by that timestamp fix.
- Record the guarded candidate's fresh local acceptance: all 100 Heritage
  questions and all three journeys pass against the exact deterministic build.
  Independent review and earlier failed attempts remain inspectable.

## Discovery budget allocation — 23 September 2026

- Prioritise real resolved-concept paths before lexical candidates spend the
  shared corpus-reading allowance; preserve the existing limits and v1/v2 replay.
- Replace repeated full card and incident diagnostics with exact metadata
  references and bounded lazy readers. Keep whole evidence, provenance, matching
  channels and unresolved obligations distinct and inspectable.
- Retain the failed first full-corpus acceptance and stale-receipt CI run.
  Small inline contexts can still refuse insufficient space; larger assembled
  contexts use the existing exact-read transport without hiding that limit.

## Source-bound discovery cards — 22 September 2026

- Add an opt-in corpus v3 contract for separately sharded discovery cards,
  whole-record evidence bindings and complete directed relationship inventories.
  Cards help find passages without becoming source evidence or domain concepts.
- Add fixed, inspectable BM25 source and discovery channels, explicit resource
  bounds and fail-closed card, source and incident-relationship checks. Reuse
  Reader loading and bounded exact evidence delivery; retain v1/v2 replay.
- Document producer-wide validation, future presentation boundaries and the
  separate source, semantic, client and answer-quality acceptance gates. No
  remote service defaults, approved sources or historical engines are changed.

## Compact logical-context delivery — 22 September 2026

- Add bounded browser WebMCP catalogue and exact-evidence reads using the existing lossless delivery helpers. Preserve the full assembly budget, package identity, source spans and explicit gaps across paginated responses.
- Group repeated item-local diagnostics in logical-unit v2 corpora without dropping affected IDs or combining dependency pairs. Preserve the v1 page-corpus representation and frozen receipts.
- Add strict input, cancellation, reconstruction and browser/UI parity controls. Remote-service admission, default versions and specialist answerability remain separate.
- Preserve earlier study-club and Heritage receipts, then refresh the current
  local candidate against the exact app: 100 Heritage questions and three
  journeys pass. Publication still requires the merged CI and live identity gates.

- Retain a separate compact-delivery observation for the later DWP semantic candidate: exact 32 KiB responses preserve the selected 512 KiB packages, while all five 32 KiB inline cases still lack source evidence and all contexts remain insufficient.

## Learning-path review and large-bundle teaching — 22 September 2026

- Add DWP learning paths to the home catalogue and Reader bundle selector, with explicit evidence and training limitations.

- Add opt-in assessed large-bundle programmes with source-record links, prerequisite checks, local journals and signed facilitator decisions. Keep practice ticks separate from assessed passes and bind decisions to the learner, programme and bundle snapshot.

- Integrate the draft small-bundle learning presentation with current main.
- Gate conceptual facets on the declared learning presentation.
- Add optional ordered learning paths to large descriptors, with outcomes,
  practice prompts, time estimates and reversible progress for the loaded session.
- Keep record and evidence inspection on existing Reader routes and loading paths.
  Preserve legacy bundles and document producer bounds and acceptance limits.

## Service 0.6.1 client question compatibility — 21 September 2026

- Make the shared non-blank question pattern compatible with clients that match
  the whole string. Preserve the accepted questions, 2,000-character limit,
  read-only tools, source versions and evidence engines. Add regressions for
  ordinary, multiline and Unicode questions across all three tools.
- Explain how to refresh an installed connection and establish tool access in
  each new AI session. Keep missing tools, connector validation, public MCP
  delivery and model answer quality as separate observed outcomes. Distinguish
  the connection smoke test from the fixed staff benchmark.
- Check historical 0.6.0 receipts against their retained build and runner;
  preserve every archived observation when the current Worker changes.

## Recorded service publication and documentation consistency — 21 September 2026

- Correct the remote guide and service README to reflect the retained 0.6.0 public deployment and SDK observation. Keep source, runtime, engine and verifier identities separate.
- Derive the three dated status blocks from immutable DWP receipts and check them in pull-request and Pages gates. The shared DWP publication page is the reference for later recorded deployments; a dated observation is not a live health check.
- Preserve frozen candidate records, failed observations and earlier browser results. Public delivery does not establish Data agent, ChatGPT Voice, legal or answer-quality acceptance.

## Service 0.6.0 preparation — 21 September 2026

- Prepare Ask OKF service 0.6.0 with the final partner/household
  qualification corpus. Retain all four historical sources and their exact
  replay paths; admit the new source only with the current assembler.
  Verify ten local cases, including 512 KiB care-home and empty-control packages,
  without promoting missing evidence, legal review or model acceptance.
  Extend the separate bounded live verifier to both exact public questions;
  preserve all earlier public observations.

- Add versioned evidence replay. Bind new review links
  to approved source and assembler identities; check older links against a
  bounded compatible set and return only the complete matching package. Keep
  unknown original assembler identity explicit, preserve full-package output,
  and expose replay provenance alongside compact delivery. Retain local tests,
  eight source/assembler integration cases and browser observations separately
  from the earlier 0.5.0 deployment.
  The later 0.6.0 public observation is recorded above.

## SDK discovery verification — 21 September 2026

- Compare complete tool definitions across MCP SDK versions while validating their distinct server/cache envelopes. Preserve exact schemas, read-only annotations and trust metadata.
- Retain bounded stage diagnostics for failed verification. Fifteen controls include both real SDKs against the local service; the public failed attempt remains recorded separately. Runtime and deployed Worker bytes are unchanged.

## Preserved evidence navigation — 21 September 2026

- Add a static history landing page after the publication audit found a directory link without an index. Preserve all frozen observation, classification and inventory bytes; this is a separately labelled navigation addition.

## Unreleased

- Open learning-hub guides with normal browser navigation so static documents do
  not first fail as application routes. The DWP Try link now reaches a beginner
  launch guide for the actual learning programme, with its evidence and assessment
  limits, rather than the separate guide for programme designers.

- Add optional logical evidence units with exact source spans and explicit
  boundary uncertainty. Corpus v2 separates physical page and retrieval-unit
  counts and loads bounded declared destinations beyond lexical candidates.
  Preserve whole-unit qualifications, source provenance, read-only delivery,
  corpus v1 package identity and frozen service engines. Add integrity, budget,
  ambiguity and compatibility controls; expose spans in Ask OKF. Retain exact
  compressed v1 replay inputs so platform-specific gzip output cannot alter
  the historical comparison.

- Share English question-scaffolding classification between corpus discovery
  and unresolved-term diagnostics. Preserve declared aliases, ambiguity,
  meaningful qualifications and identifiers; remove misleading missing-concept
  reports for pronouns and broad request wording. Add cross-domain regressions
  while preserving frozen service engines and historical context packages.
  Refresh the generated study-club execution and local Heritage browser
  evidence against the new application build, using actual evaluations rather
  than rebinding earlier passes. Document these separate evidence lifecycles.

- Add an offline exporter and small static reader for explicitly approved,
  fixed evidence examples. Preserve canonical packages and source/engine
  declarations; expose bounded, hash-checked records, provenance, gaps and
  directed relationships without storing anonymous questions or changing the
  live service. Add eleven offline controls to the context-assembly CI gate
  and publication contract. Retain a local Chrome journey with seven focused
  checks and its earlier failures; public publication remains a separate gate.

- Give valid, bundle-declared evidence paths priority when node, relationship or
  byte limits would otherwise omit them. Use at most one additional allocation
  pass from the original resolved seeds, retain whole passages and explain the
  priority without changing source authority. Keep missing paths and truncation
  explicit when the required evidence still cannot fit.
- Report a retained source's declared dependencies even after supporting edges
  are trimmed. This corrects a diagnostic gap and changes affected context
  identities; historical packages and model trials remain unchanged.
- Show each large-corpus record's declared type in the detail badge before and
  after loading its full record. Use collection vocabulary only as a fallback,
  so mixed corpora do not label every item as the same kind of document.
- Retain a separate local service integration observation for the changed
  engine: four approved source versions and four compact cases pass. Record
  changed corpus context identifiers explicitly and preserve the earlier 0.5.0
  receipts. A fixed source version alone does not freeze a future engine's
  assembled package.
- Re-execute all 100 Heritage questions and three browser journeys against the
  revised application, retaining exact build identities and fresh results.

- Document the published Ask OKF service 0.5.0, its household/statutory default
  and all four explicitly selectable source revisions in the top-level guide.
  Record seven actual full-context and four compact SDK passes, three functional
  browser passes and the retained Firefox hosting-cookie warnings. Keep evidence
  delivery, legal completeness and AI answer quality separate.
- Restore publication documentation lockstep after the service-only PR updated
  its nested guide and changelog without the contract-required top-level entries.
  Preserve the failed Pages run; this documentation correction changes no engine
  or service runtime bytes. Check lockstep before impact-based job selection so
  service-only pull requests cannot skip this publication requirement.

- Keep ambiguous Ask OKF meanings inspectable as separately labelled evidence
  branches without resolving them or activating their evidence requirements.
  Preserve bounded paths through shared junctions and expose path-limit omissions.
- Load up to four hash-bound corpus files concurrently with deterministic budget
  reservations and unchanged aggregate limits. Retain exact source and ranking
  checks; increase only the semantic/base-index ceiling to an explicit 8 MiB.
  Package, manifest, posting and record-shard limits remain unchanged.
- Replay synthetic context, three-version remote integration and exact-build
  Heritage acceptance after the engine change: all 100 questions and three
  fixture journeys pass. Document the older corpus context-identity change;
  mismatched historical replay recipes continue to fail closed.

- Serve the additive DWP staff-semantic corpus by default in Ask OKF service
  0.4.0, retaining both earlier approved corpus versions and exact replay links.
  Verify the real three-version loader and compact delivery against hash-bound
  source files. Keep current staff profiles insufficient with named obligations.
- Request `no-transform` for service HTML while preserving Content Security
  Policy and `no-store`; verify actual host behaviour separately. Align the
  service architecture, README, changelog and reproduction commands.

- Count missing-evidence explanations while trimming Ask OKF packages to their
  byte budget. Retain whole evidence records when they fit, rather than falling
  back to an empty package after adding final diagnostics. Add generic boundary
  regressions; keep genuine metadata-only budget failures explicit.
  Replay the non-DWP acceptance receipt against the updated engine hash;
  its evidence package and context identifier remain unchanged.

- Include the bounded, transitive Markdown reading closure in the documentation
  cache identity, so changes to linked service guides and changelogs rebuild
  their rendered pages and exact-source alternates. Keep unrelated service
  sources outside that cache dependency.

- Correct a stale replay link in Ask OKF service 0.3.1: changing the question or
  source version, or starting another replay, removes the old link. Browser
  regression checks that a newly requested link matches the new context.
  Preserve 0.3.0 deployment evidence and verify the corrected runtime separately.

- Extend conceptual-navigation acceptance with an optional published-bundle
  journey. Check real network responses against pinned local files and verify
  the deployed application manifest, while retaining the offline fixture mode.

- Keep browser observations of an external Ask OKF service in a separately
  selectable output directory. Receipts distinguish the tested target from the
  local build reference and leave deployment identity to its own receipt.

- Refresh exact-build acceptance for the navigation candidate: 100 Heritage
  questions and three local journeys passed. Add separate service-browser CI
  coverage and full-corpus conceptual-filter browser checks. An actual DWP
  check exposed a producer topic-field mismatch, corrected in the additive
  producer projection; the regression now compares hydrated record counts.


- Preserve exact Markdown alternates when newly linked reading pages come from
  outside the ordinary copied documentation directories, including the changelog
  and remote-service guides. Validate those links in the assembled site.


- Add bounded, lossless evidence catalogues and read-only evidence slices to the
  remote Ask OKF service, preserving full-package output and context identity.
  Add a stateless browser evidence reader, explicit delivery limits and replay
  tests. These changes are a version 0.3.0 candidate; hosted and AI-client
  observations are separate from implementation checks.


- Show producer facet descriptions and optional conceptual-classification
  methods, review status and whole-snapshot coverage. Preserve ordinary facet
  filtering across Graph and Timeline without turning text mentions into
  applicability relationships. Add explicit source, inferred and audit date
  roles, source-only Timeline selection and partial-date groups that never
  invent a month. Keep a visible Explorer changes link in the header.

- Verify the deployed full-source HTTPS service with the official MCP SDK:
  complete imprisonment and hospital packages match the shared core, and the
  explicit historical imprisonment package remains unchanged. A separate raw
  HTTP client verifies exact package equality for 40 staff questions and three
  controls, all correctly insufficient. Keep subsequent ChatGPT observations
  and Explorer Pages publication as separate acceptance records. Fix the confirmed
  Worker redirect incompatibility using manual handling that rejects every 3xx
  response without following it.

- Remove personal pronouns and broad forms of “go” and “get” from full-source
  question ranking after a real ChatGPT call exposed conversational-word noise.
  Retain substantive evidence terms and verify the correction with non-DWP
  fixtures and a separate full staff-question run; no answer rules are added.

- Record fixed failure stages for remote evidence assembly without logging
  questions, source URLs or exception text. Refresh exact-app Heritage and
  non-DWP regression receipts after the corpus extension; preserve their source
  snapshots and existing context results.

- Extend the remote MCP service to an approved immutable full DMG and ADM
  corpus, retain the original profile by version and cache only verified public
  source files. Add an accessible landing page with beginner guidance and clear
  source coverage, protocol and answerability boundaries.

- Add hash-bound full-source lexical discovery to Ask OKF, with explicit page,
  query, transfer and package limits. Preserve source governance, scoped graph
  traversal and the historical narrow-index adapter; never treat lexical
  matches as a complete answer. Share the corpus reader across UI and tools.
- Fix silently omitted large-graph relationships, inactive stack expansion and
  Links overview assumptions about record identifiers. Add bounded paging,
  counts and cross-browser regressions.

- Add an independent read-only remote MCP adapter for the existing Ask OKF
  engine, with approved immutable source bindings, bounded requests and the
  unchanged governed evidence package. Document ChatGPT setup, host boundaries,
  privacy, threats and separate deployment/client acceptance.
- Keep the remote adapter's locked tests and build in a required CI job. Route
  service-only changes independently from unchanged Heritage browser artefacts,
  while preserving unknown-path failure controls and frozen profile contracts.
- Deploy the public HTTPS test service and verify complete package equality with
  two protocol clients. Record actual ChatGPT calls and distinguish its runner's
  output limit from the assembler's explicit evidence budget. Retain smaller
  package demonstrations and limitations without changing the shared engine.

- Reject invalid optional review status, original assertion identities and extra alias fields before context assembly or rendering.

- Make context-control replay portable across zlib builds while retaining exact
  archive digests, bounded decompression and byte-for-byte fresh content checks.
  Reject archive tampering, changed content and oversized evidence.

- Add a transport-independent governed context assembler and additive producer,
  package and evaluation-case contracts. Keep Search and the earlier lexical
  MCP context pack separate. Check declared phrases, directed dependencies,
  whole evidence, provenance and explicit gaps within bounded packages; do not
  generate model answers or infer domain rules from hyperlinks.
- Add independent A–H package evaluation, exact identity and path controls,
  portable execution receipts and synthetic negative cases. Preserve earlier
  research trials and distinguish optional tool registration from observed AI
  host invocation.
- Execute the generic engine's synthetic controls, independent assessor
  controls and non-DWP retained case in a required context CI job. Check exact
  source URL and capture-time expectations alongside hashes and directed paths.
- Keep the standalone study-club context case under `tests/fixtures`, preserving
  its exact inputs and execution evidence without claiming a complete
  Evaluation Foundry fixture family.

- Restored indexed Timeline when opened directly on a selected record or after
  targeted Graph loading, retaining the browser memory limit and bounded graph
  requests. Recognised explicit operational source release dates separately
  from catalogue timestamp fallbacks, preserving month and year precision.
  Refreshed the exact-build Heritage evidence with 100 passing questions and
  three passing local journeys, retaining the original source snapshots.

- Labelled small-bundle Timeline dates by their provenance. A described source's
  declared publication or issue date takes precedence, with month and year
  precision preserved. Record publication, capture, observation, generation and
  legacy timestamps remain separately labelled instead of implying publication.
  Refreshed the exact-build Heritage runtime evidence: all 100 questions met the
  required score and all three local journeys passed against the new app, with
  the original source snapshots preserved. Documented the evidence refresh.

- Reorganised onboarding into linked task guides for trying AI, choosing examples,
  audience journeys and a complete fictional first bundle. Shortened the README,
  separated development/reference material and reconciled current interaction
  instructions. Added one generated learning catalogue with a frozen external
  government WebMCP entry, explicit AI-access limits and deterministic starter
  files, question checks and download. Production gates and historical anchors
  remain preserved. Linked Chris Page’s recorded AI/WebMCP demonstration from
  the README and access guide, with a watch-and-try exercise that distinguishes
  the recording from host compatibility and the separate human browser check.

- Preserved panel focus across Chrome's native blur during responsive resizing,
  while leaving an explicitly focused visible control undisturbed.

- Preserved the byte-frozen Heritage evaluation profiles while classifying the
  four current repository journey manifests through their existing control rule.
  Corrected impact planning to validate both artefact and semantic identities in
  version 2 plane receipts, retaining fail-closed handling of malformed roots.

- Added a paired Search and Details layout for medium screens, compact inline
  footer controls and a Results tab for small-bundle navigation. Collapsed paired
  selection actions into an explicit disclosure, aligned tab/control heights and
  reduced oversized detail headings. Kept facet selection summaries within a
  stable header so double-click targets do not move. Colour segments
  now toggle values while keeping facets folded; a second ordinary click removes
  a selection. Double-click keeps the original multi-value selection.
- Restored the required geospatial index when opening a Reader URL with a map
  restriction, including browser history restoration, and retained loading state
  until that scope is available.

- Opened each new bundle with all facets folded and Title as the browsing default,
  preserving explicit URL sorting. Replaced facet bars with alternating luminance
  colours and a separate black-on-white highlight track, labelled counts and
  selection ticks, with measured adjacent-colour contrast tests.

- Preserved the full-index fallback for older search bundles when no highlight
  or keep selection is active. Section facets now use the same declared section
  as the record inspector. Added regression coverage for both behaviours and
  clarified the information included in shared URLs and exported pins.

- Harmonised Explorer selection around SeeLinks: highlight-first previews,
  OR within a facet, AND across facets, reversible keep/remove operations and
  presentation folds with live counts. Added shared cards and lists, a mobile
  three-panel workspace with footer and swipe navigation, keyboard splitters,
  bundle-qualified bookmarks and tabbed small-bundle provenance. Extracted the
  interaction algebra, panel shell, facets, results, bookmark shelf and small
  record inspector into independently reviewable modules. Local rehearsal
  evidence and remaining browser gates are recorded in
  [the implementation note](docs/explorer-interaction-implementation.md).

- Started the estate-wide OKF build and publication standardisation programme.
  Added a separate publication-method v1 profile, an executable repository
  contract, contract-driven documentation and changelog lockstep, a fail-closed
  dependency planner, a 17-entry operational estate registry with an accessible
  generated view, safe workbook-folder controls and a live adoption tracker.
  The semantic bundle registry now validates every bundle against its frozen
  schema, while ordinary Explorer and future heritage CI use runner Chrome and
  reserve bounded Firefox/WebKit installation for cross-engine assurance.
  Pages now binds the assembled candidate to its exact commit and governance
  bytes, then verifies the deployed estate-registry journey in installed Chrome
  without rebuilding the candidate.
- Recorded the estate adoption wave in the lockstep tracker and hash-bound
  every publication contract to its reviewed commit. All 11 applicable
  producers are adopted: Explorer, AI Infrastructure, UK Government APIs,
  GOV.UK Content, ONS, UK Living, ELS, Planning, WCC, Land Registry and
  legislation. Land's focused routing
  [PR #8](https://github.com/chris-page-gov/okf-LandRegistry/pull/8) passed its
  exact-main immutable-maintenance path; legislation passed hosted validation,
  exact-main Pages deployment and six byte-identical HTTP route checks. Six
  reviewed fixtures, immutable publication units, compatibility surfaces and
  upstream references have justified non-applicable decisions. The WCC
  workbook case proves that an unadmitted
  `.xlsx` change can select four content-free inventory and documentation gates
  without a semantic or release rebuild. It also replaced a brittle Playwright
  cache-path link with the supported installed-Chrome channel.
- Corrected the generated estate-registry page after its first exact-commit
  Chrome run proved the identity and route bytes but exposed an origin-root
  favicon 404 through the clean-console gate. The page now links its
  project-relative SVG icon and verifier diagnostics include the resource URL
  and location. The final registry binds audited correction `6060d4a2`, with
  11 applicable repositories adopted and six reviewed units non-applicable.
  The correction needs one affected Site build, not a semantic bundle, corpus
  or release rebuild.
- Rebound the heritage local-candidate receipt to the exact deterministic
  Explorer build changed by the learner registry correction. A genuine Chrome
  run passed all 100 questions and all three local journeys; the corpus,
  semantic bundle and released external heritage publication remain unchanged.
- Added a no-code walkthrough of the immutable AI Infrastructure OKF v0.6.0
  knowledge graph. It starts with browser-only GitHub and Explorer routes,
  explains the optional Obsidian download, follows one Markdown relationship
  into its direct and reified semantic representations, and distinguishes the
  roles and limits of YAML-LD, JSON-LD, validation, release receipts and MCP
  retrieval.
- Avoided the GitHub-hosted runner's unreliable Ubuntu package mirror for
  Chromium contract tests by using its installed Chrome channel. The less
  frequent Firefox and WebKit dependency installation now has a bounded timeout.
- Replaced the learner-hub header's unscoped Explorer launch with an exact,
  source-backed three-record Heritage worked example opening the Coventry
  Cathedral record. On narrow screens, large-corpus navigation now starts
  collapsed so the selected evidence appears before optional facet controls.
- Published the evidence-backed OKF evolution review, including its chronology,
  bundle inventory, standards decision ledger, conversation evidence register,
  current best practice, grounding analysis and bounded MCP retrieval prototype.
  The generated OKF bundle now contains 165 nodes and 588 semantic assertions.
- Replaced the Pages root with a static learner-first project studio for an
  18-year-old non-specialist, while preserving the complete interactive
  Explorer at `/explore/` and retaining legacy deep-link behaviour. Added an
  eight-stage independent-learning pathway, governed bundle catalogue,
  cross-bundle personas and user stories, project templates, accessibility
  checks and three-engine browser journeys.
- Expanded the beginner learning path with a practical data-to-ontology
  on-ramp, route-based study plans, foundational definitions, an annotated
  authoritative-source register and reconciled geospatial and glossary
  material. Standardised the affected documentation on British English with
  context-aware guidance and checks that preserve exact code, schema, URL and
  official-title spellings.
- Added stable visible permalinks to beginner headings while preserving
  established fragment identifiers. Direct fragment loads, refresh and browser
  history now focus and visibly highlight the selected section without
  replacing the existing accessible collapsible and pinned learning-path
  sidebar.
- Added a deterministic SharePoint Word-record experiment for all 293 governed
  `okf-uk-living` service families, with a frozen synthetic evaluation schedule,
  exact provenance and citation gates, append-only continuation controls and an
  explicit ignored boundary for tenant-specific transcripts and attempt
  evidence. The completed Microsoft 365 Copilot development run kept all 293
  family responses inside the safety boundary and selected the exact expected
  family in 292 cases. The public report preserves the one Universal Credit
  near-neighbour exception, distinguishes service protection from semantic
  failure, estimates the counterfactual pay-as-you-go charge and states that an
  independent holdout and the OneNote condition remain outstanding.
- Reconciled the external-source inspection research with the implemented
  typed `source_access` contract. The guidance now separates today's bounded
  `link`, `json`, `xml` and `text` display modes from future multiple-
  representation selection, format-specific renderers, clearer transport and
  parse diagnostics, and their required regressions.
- Added a fail-closed `okf-uk-living` reconciliation gate for the complete rich
  relationship runtime, route locator and four executable Draft 2020-12
  schemas. The audit validates every projected assertion and route commitment,
  applies the Reader's aggregate loading ceilings, and requires an explicit
  reviewed preset when a producer is audited from a renamed worktree.

## v0.7.0 - 2026-08-12 - Explore OKF authoring and review

- Reviewed the reusable OKF authoring methodology against the completed Land
  Registry retrospective and public graph evidence without reopening that
  producer. Added versioned Authoring Profile v1 controls for readable labels,
  compact graph endpoint indexes, denominator-based useful semantic linking
  and an explicitly non-release Explore OKF feedback stage.
- Made semantic-link coverage identity-bound and fail-closed: declared
  candidate inventories reconcile exactly across outcomes, assertions and
  dereference receipts; mapping strength agrees with its predicate; targets
  remain inside governed URI namespaces without encoded-delimiter bypasses;
  duplicate assertions cannot inflate coverage; and approved observations
  require one approval-grade receipt carrying the canonical complete-result
  digest and exact observation time. The guidance explicitly separates this
  mechanical guarantee from
  the domain review needed to judge whether an eligibility rule omitted an
  entity.
- Added the strict, versioned Explore OKF producer contract, JSON Schemas,
  examples and Python build/validation tooling for snapshot-bound endpoint
  labels and exploratory-publication metadata.
- Added Explorer consumption of integrity-bound endpoint labels across
  route-based views. Opaque generated identifiers now appear as a visible
  **Missing label** quality defect while the route, IRI, type and label
  authority remain inspectable. Source-native record routes and selected
  relationships now survive copied links, reload and browser history.
- Added a persistent exploratory-publication banner, route- and filter-
  preserving feedback links, explicit invalid-contract warnings and fail-safe
  `noindex` behaviour across all Explorer views.
- Fixed the beginner learning-path panel so its final chapter remains reachable
  by scrolling the panel independently of the article. After chapter
  selection, the panel can collapse to a narrow current-chapter rail, expands
  on hover or keyboard focus, and provides an accessible persistent pin. Deep
  section fragments now restore the exact article position after reload in
  Chrome, Firefox and WebKit.
- Updated both copy-ready Foundry prompts, the bundle-authoring guide and the
  beginner explanation, and researched a bounded Coventry everyday-services
  open-data pilot to test the method before reviewing `okf-uk-living`.
- Replaced host-dependent Python and repeated unpinned `pip` installation with
  a committed `uv.lock`, CPython 3.12.11 pin and `uv` 0.12.2 requirement.
  Local guidance and all root workflows now use one locked invocation path,
  with a reusable commit-pinned CI setup action and regression enforcement.
- Kept `requirements-okf.txt` only as the byte-preserved legacy compatibility
  requirements manifest required by the current digest-bound Heritage
  publication-unit contract; its separately versioned host-Python workflow
  does not lock transitive dependencies, while Explorer development and CI no
  longer install from it. Preserved the exact released Heritage
  evaluation-profile bytes while preventing current root workflows and the
  Foundry browser server from using a host interpreter.
- Preserved hidden publication files in the unchanged GitHub Pages artefact,
  including `.nojekyll`, and added a workflow regression assertion. This keeps
  the downloaded release artefact's complete Site tree identical to its
  candidate receipt rather than silently omitting one governed file.
- Separated the editable bundle-URL draft from the committed loaded-source URL
  so an initial asynchronous load cannot overwrite a person's newly typed URL
  or leak unsubmitted text into copied routes and feedback links. Delayed
  registry start-up and slow local-file reads also cannot supersede a later
  explicit load or clear its loading state. Accepted remote and local
  descriptor identities are bound before a later validation or subordinate-
  hydration failure can expose feedback, so that feedback names the descriptor
  actually under review. Added deterministic delayed-registry, delayed-bundle
  and delayed-file browser regressions with exact failure-identity assertions.
  Refreshed the exact-app Heritage consumer evidence and corrected its receipt
  generator to use British-English publication wording.

## v0.6.3 - 2026-08-12 - Accurate large-corpus overview counts

- Corrected the unloaded large-corpus overview metric to prefer the declared
  record count over the dataset-grouping count. This prevents 14 Land Registry
  dataset groupings from being labelled as 14 discovery records when the
  bundle declares 2,203 records.
- Preserved compatibility with older bundles that publish only a dataset count
  and added cross-browser regression coverage for bundles whose record and
  dataset-grouping counts differ.

## v0.6.2 - 2026-08-11 - Bounded producer search policy

- Added an exact, opt-in `okf-static-search.v2` query policy so a producer can
  declare its component tokeniser, bounded stopwords and integer-rational
  minimum-should-match rule. Manifests without the extension keep strict AND;
  declared policies count unresolved tokens as unmatched, require the stated
  number of distinct token groups and expose the deterministic decision in the
  worker response.
- Added `wait_for_ranked_result` and `ranked_result` to external bundle
  journeys. Both bind a result by its credential-free canonical URL, wait for
  the exact URL query to reach Explorer's settled marker and fail immediately
  when the settled result is absent, while genuine loading retains the bounded
  90-second journey ceiling.
- Covered policy compatibility across strict-AND producers, entity aliases,
  embedded entity prose, typo tolerance, capped query bounds and malformed
  contracts. Added real-browser acceptance tests for delayed settlement and
  prompt settled-empty failure.

## v0.6.1 - 2026-08-11 - Governed runtime evidence

- Added bounded, declarative capture of ordered DOM attributes to external
  bundle journeys. Producers can now place the Explorer's actual ranked
  canonical result URLs in the digest-bound browser receipt, allowing release
  evaluation to calculate retrieval and forbidden-result metrics from the
  rendered product rather than from a separate lexical approximation.
- Extended governed entity recognition to unambiguous organisation names
  embedded in natural-language questions, while keeping residual subject terms
  as required lexical constraints. Corrected count-only and entity-filtered
  result totals so capped postings are reported as lower bounds or unknown,
  rather than as exact counts.
- Made Explorer acceptance commands checkout-scoped single writers. A second
  invocation fails before deterministic build output, Chromium or receipt
  creation; inner runners reject direct invocation, bind the receipt to the
  exact wrapper and lock-module bytes, and verify their live parent-owned lock.
  POSIX process groups are terminated and verified before release; an
  unverifiable descendant deliberately leaves the lock in place.
- Made external runtime input immutable for the browser run by copying the
  bundle and Explorer build into independent private snapshots, serving only
  those copies and rechecking the source trees afterwards. Added aggregate
  manifest, telemetry and receipt limits, strict readiness selectors, atomic
  ranked-row capture and fail-closed credential-free canonical URL checks.
- Preserved an optional non-empty `review_status` on every governed rich
  relationship assertion while continuing to require it for model-derived
  assertions. Normalised producer review state is therefore available to the
  Reader instead of being silently discarded.
- Added Predicate Registry v2 as an additive, integrity-locked profile. Its
  bounded schema derives exact `active-emitted` and
  `authorised-zero-evidence` states from the complete relationship set, binds
  the whole registry rather than only its predicate rows, and leaves the
  frozen Bundle Wiki v1 profile byte-identical.
- Closed the Heritage Foundry producer-provenance boundary by including the
  shared semantic module in both impact profiles and binding every declared
  producer input into a sorted, exact-byte `producer_materials` root in the
  candidate receipt. Python generation and JavaScript evaluation enforce the
  same limits and digest rules.
- Strengthened the legislation runtime gate with exact candidate and Explorer
  Git identities, immutable bundle and application snapshots, write-once
  evidence publication, aggregate resource ceilings and a fail-closed
  diagnostic mode. Release evidence now binds the completed deterministic build
  and all executable acceptance material before and after the browser journeys.
- Archived the byte-preserved 26 July v1 runtime receipt as historical evidence
  instead of presenting its legacy `passed` field as current authority. The
  canonical path is populated only by the clean-commit, annotated-tag and
  exact-candidate v2 release gate.
- Governed selective CPSV-AP 3.2.0 adoption across the OKF producer portfolio,
  preserving DCAT/DCAT-AP, ELI and statistical-model boundaries while
  requiring pinned standards assets, evidence-bearing mappings and honest
  subset-validation claims.
- Pinned the canonical Bundle Wiki v1 profile to the complete, byte-exact
  Explorer v0.6.0 16-file tree and an adjacent machine-readable vendor lock.
  Reconciliation now fails closed on a missing, extra or drifted canonical
  mirror, checks every declared relationship schema, and offers an explicit
  symlink-safe profile sync with opt-in replacement of divergent files.

## v0.6.0 - 2026-08-10 - OKF 0.2 semantic contracts and governed YAML-LD

- Completed the local nine-repository OKF 0.2/YAML-LD migration: every
  reviewed `okf-*` repository now declares one machine-readable semantic
  contract, absolute semantic identities plus safe local routes, synchronised
  direct and reified assertions, deterministic runtime projections and exact
  setup/build/check guidance. Added positive, negative and compatibility
  fixtures; shared-schema sampling to the cross-repository auditor; compact
  digest-bound semantic shards for large graphs; and a bounded rich-runtime
  Reader path with lifecycle-separated planes, SHA-256 route locators and
  per-plane incident assertion commitments. This release commits the Explorer
  contract, plugin and Explorer-owned migration as v0.6.0. The remaining
  producer migrations are independent working-tree changes and are not claimed
  here as committed, released, deployed or publicly verified.
- Closed the independent cross-layer review findings before declaring local
  completion. The shared assertion schema now requires preferred labels and
  canonical credential-free HTTP(S) authority/evidence/rights sources; every
  pinning producer validates its complete population against the exact shared
  bytes. Explicit YAML-LD graphs now fail closed on missing, duplicate, unsafe
  or inconsistent routes and incomplete assertions instead of deriving routes
  from IRIs. The portable contract auditor validates the complete contract
  shape. The older heritage-corpus producer now propagates the preferred label
  into every reified assertion too, so tightening the shared schema does not
  leave a second Explorer-owned generator behind. Rich route hydration now
  takes priority independently of record
  locators, decodes shards sequentially, drops undeclared row properties and
  enforces aggregate compressed-byte, retained-text, evidence and supporting-
  assertion ceilings in addition to its existing row/chunk bounds.
- Hardened the portable contract and plugin review boundary: declared paths and
  globs must stay repository-relative and contained—including authoritative
  inputs and symlinked prefixes—and glob expansion has an explicit match-count
  ceiling. Tooling strings remain untrusted declarations until reviewed against
  trusted guidance, and sampled plain/gzip artefacts are read under explicit
  on-disk and decoded-byte limits. Malformed relationship, semantic or runtime
  JSON now fails the audit instead of disappearing as an empty sample or
  aborting reconciliation, and the shared assertion schema uses the same
  whitespace-free RFC-style IRI rule as the reconciler and Reader.
- Completed the Explorer repository's own OKF 0.2 semantic migration. The
  Markdown graph now deterministically emits `okf-bundle.yamlld`,
  `okf-bundle.jsonld` and a compatibility runtime bundle from one normalised
  source. Its 579 directed Markdown links are represented conservatively as
  evidence-bearing `dcterms:references` assertions with absolute identities,
  suffix-free semantic routes, direct triples, reified assertion nodes,
  provenance and rights; section placement is not promoted into an invented
  domain predicate. Site assembly verifies and publishes both semantic
  representations alongside the runtime projection.
- Added an additive large-record contract for repository-authored Narrative
  Markdown, enclosing-process context, previous/next steps, variants and
  related routes. A selected record now uses this content instead of the
  generic reduced-corpus narrative when the producer supplies it.
- Added typed resource source access with `link`, `json`, `xml` and `text`
  display modes. Explorer fetches only explicitly browser-readable JSON, XML
  and text, renders XML/text as inert escaped content, retains
  `source_api_url` as JSON-only compatibility behaviour and always preserves a
  direct official-source link. No response is persisted or redistributed.
- Standardised authored release pages, semantic profiles, CI labels, plugin
  guidance and visible Explorer copy on British English and GOV.UK plain
  English conventions. Repository and generated producer-agent guidance now
  preserves exact schema values, identifiers, official titles and upstream
  product terminology where localisation would be incorrect or incompatible.
- Made focused-graph navigation explicit and reversible: Graph actions now
  recentre the inspected record, preserve the prior focus in the URL, reset
  the viewport and let browser Back restore the previous focus and inspector.
  Bounded focus graphs enter the established ordered relationship regions at
  four relationships rather than falling back to an unstructured radial view.
- Clarified the large-corpus facet controls as `Clear filters` and `Reset facet
  layout`, grouped them together with Guidance, and kept the compact bundle
  loader beside the Explorer title at tablet widths.
- Tightened graph pointer targets so an icon no longer owns the empty span to
  its label or overlaps another node. List labels now sit inline on their
  outside edge, staircase labels prefer the available left/right width, and
  each visible label retains its own exact clickable target.
- Labelled single record-type stacks as `All matching …`, distinguishing a
  bounded loaded subset from the full match count. Opening a large stack now
  always produces semantic subgroups, with deterministic title bands/ranges as
  the fallback, instead of fanning out as many as 72 individual records. Each
  subgroup can be opened again, and every open/close step is URL-backed so
  browser Back restores the preceding stack depth.
- Split a single dense relationship family into paired left/right columns,
  placing two icons per row and keeping both label columns outside the focus.
  Smaller lists retain the established single-column relationship region.
- Added an explicit open-stack hierarchy above the graph canvas. Every opened
  level keeps its sibling choices together across the row, marks the active
  branch as `Open below`, names the level shown in the graph, and moves inactive
  siblings out of the child canvas. Switching branches or using browser Back
  therefore produces a visible structural change as well as different counts;
  relationship rows use those groups' human labels rather than encoded routes.
- Compressed the graph chrome without flattening its visual hierarchy: node or
  relationship keys now share one horizontally scrollable rail with authority,
  hierarchy levels use slimmer breadcrumb-like rows, and facet cards use tighter
  headers, value rows and three-value summary samples so more of the corpus and
  graph remain visible at laptop and tablet heights.
- Added an evidence-backed Heritage Evaluation Foundry engineering postmortem,
  complete public-safe prompt-response trace, machine-readable process
  registers and a modular content-addressed report builder. The analysis
  reconstructs all three pull requests and six green workflows, identifies
  late-discovery amplification, and specifies an Evaluation Profile v2 impact
  planner and selective-rerun architecture. Rendered reading pages now declare
  the project-relative SVG favicon so direct documentation journeys do not
  create a spurious `/favicon.ico` browser error.
- Implemented every postmortem recommendation: Evaluation Profile v2 and its
  consumer lock; a fail-closed, explainable impact planner; promotion evidence
  outside candidate roots; `normalized-core` and per-plane emitters with
  manifest-owned changed-only writes; 13 adversarial late-finding
  microfixtures; conditional parallel CI plus nightly three-engine shadow
  assurance; content-addressed Site components; hash-sharded semantic and link
  intents; an independently rooted heritage publication unit; and an annotated,
  attested, immutable-release policy. The faithful, tiny and synthetic
  candidates now share one explicit external URL family while remaining
  identity- and count-isolated.
- Made the public verification harness resolve GitHub Pages project-root
  bundle paths, wait for client-rendered identity, select a single terminal
  journey and return a failing exit status when any browser journey fails.
- Made every rendered reading-page return link target the project directory
  root rather than the non-route `/index.html`, and replaced the generated SPA
  fallback with a standalone, accessible 404 whose assets and return link stay
  inside the `/okf-explorer/` GitHub Pages project root.
- Replaced the retired HAR `/list-entry/{NHLE}` inference with the live
  Historic England register search bound by an exact `q=ListEntry` parameter;
  direct NHLE rich pages remain unchanged. YAML-LD record links now prefer the
  Pages project root so SvelteKit hydrates the governed deep route rather than
  treating `/index.html` as an application route.
- Made the external tiny-bundle acceptance fixture wait for the expected
  selected-record heading, so browser success proves hydration completed
  instead of merely observing the pre-hydration panel shell.
- Added a production-scale Evaluation Foundry exemplar for every supported
  National Heritage List for England feature intersecting the exact Coventry
  and Warwickshire district boundaries, plus sanctioned annual Heritage at
  Risk workbook evidence from 2013 to 2025.
- Added a separate source-backed tiny assurance fixture and a clearly labelled,
  default-off synthetic supplement so sparse semantic capabilities can be
  demonstrated without contaminating faithful records, counts or search.
- Extended OKF Markdown additively with governed YAML-LD identity, predicates,
  qualified assertions and deterministic JSON-LD projections, while retaining
  backward-compatible rendering for consumers that understand only ordinary
  YAML front matter and Markdown.
- Added Explorer presentations for heritage detail, source geometry and
  qualified graph evidence, together with bounded alternative-name and
  one-edit typo-tolerant search whose corrections remain visible to users.
- Added a beginner-facing process, 100-question evaluation suite, executable
  consumer journeys, source acquisition and reconciliation receipts,
  identifier-bound link validation, release-root integrity checks and rendered
  HTML publication for the faithful, tiny and synthetic products.
- Promoted the Coventry and Warwickshire profile from provisional to evaluated
  after its exact GitHub Pages descriptor and release root passed the terminal
  browser journey; published copy-ready faithful, tiny and synthetic entry
  points and registered the exemplar in the YAML-LD bundle registry.
- Added a final release-page journey action and a digest-bound, policy-frozen
  tagged-release evidence handoff, while excluding accidental Foundry
  evaluator output from both Git and Site assembly so post-deployment receipts
  cannot alter the verified publication closure. GitHub currently reports the
  release itself as mutable, which the postmortem records as remaining work.

## v0.5.7 - 2026-07-29 - Fail-closed descriptor identity

- Fail closed when a document declares a large-corpus schema without the
  matching `okf-large-corpus` kind, or declares the large-corpus kind without
  the supported schema, instead of silently normalising it as an empty small
  bundle.
- Added unit and real-browser malformed/degraded descriptor coverage for
  producer compatibility gates.

## v0.5.6 - 2026-07-29 - Consumer-proof Foundry releases

- Added a dry-run-first repository bootstrap scaffolder/checker for empty,
  imported and explicitly adopted targets. It creates only reviewed
  foundations and disabled CI, preserves existing files and never initializes
  Git, creates a remote, pushes or publishes.
- Added a bundle-agnostic, package-level real-browser acceptance command that
  binds the exact Explorer source/lock/build, bundle tree and descriptor
  identity to declarative journeys, requests, console output, restored state
  and terminal outcome.
- Extended the authoring profile with repository lifecycle, executable
  consumer identity, explicit compatibility windows, concrete validator paths
  and a 60-second tool-first public-URL share gate.
- Made the Foundry authoring contract machine-check a pinned consumer
  inventory and lock, a connected producer-to-public-route dependency graph,
  two-stage real-consumer fixtures, digest planes, bidirectional compatibility
  and post-deployment deep links.
- Updated both Foundry prompts, the prompt kit, beginner guidance, schema,
  template and semantic validator so a producer-only pass cannot authorise a
  release.
- Made a selected large-corpus record hydrate its resource index before the
  Resources view renders, with a three-browser regression proving the linked
  resource is visible while targeted Graph hydration remains bounded.

## v0.5.5 - 2026-07-29 - Bounded relationship navigation and OKF Foundry

- Replaced the misleading whole-relationship loader on large-corpus facet
  cards with bounded related-record previews, exact match totals and
  index-derived Graph/Links projections. Worker-backed facets no longer
  attempt full record or relationship hydration when their bounded postings
  already provide the required context.
- Made search-result selection load the chosen record's route-scoped adjacency
  immediately, while preserving the whole-corpus memory guard.
- Documented the distinction between the small semantic representations,
  chunked JSON assertion plane and facet membership; added a copy-ready UK
  Legislation AI traversal and declared the current absence of a
  source-and-target route index for the official-effects datapack.
- Published the complete documentation dependency graph as responsive HTML,
  including links embedded in tables and links from documentation into the
  sample knowledge corpus. Internal Markdown navigation now rewrites to
  deterministic HTML routes with stable heading anchors; a post-build crawl
  fails publication for raw same-site Markdown links, missing targets or
  missing fragments, duplicate identifiers, and missing scripts, stylesheets
  or images. Raw source remains machine-discoverable as an exact-build
  alternate, while the visible source link uses GitHub's rendered HTML view.
- Published the Foundry prompt kit, both copy-ready prompts, worked examples
  and authoring profile as safe, responsive HTML. The prompt pages provide
  accessible one-click exact-copy controls, plain-text downloads and an
  AI-neutral procedure and completion checklist while retaining canonical
  Markdown and schema URLs.
- Made the rendered beginner guide's desktop chapter panel independently
  scrollable without moving the article, while retaining normal mobile flow.
- Added deterministic, browser-rendered HTML publication for the complete
  beginner learning path, with accessible chapter navigation, responsive
  tables, safe Markdown handling and links back to the source files.
- Updated Playwright to 1.62.0 and `svelte-check` to 4.7.4 in one
  release-assured maintenance change, including the synchronised lockfile and
  CycloneDX dependency inventory.
- Updated Playwright to 1.62.1, synchronised the CycloneDX inventory, and
  verified the patch with the affected Chrome journeys before the full
  Chrome, Firefox and WebKit assurance matrix.
- Updated SvelteKit to 2.70.1, Svelte to 5.56.8 and Vite to 8.1.5 while
  retaining the compatible TypeScript 6 toolchain, and limited the Dependabot
  Explorer stack group to minor and patch releases so future major migrations
  receive isolated pull requests and validation.
- Updated SvelteKit to 2.70.2 and Vite to 8.2.0, refreshed and checked the
  CycloneDX dependency inventory, and verified the resolved dependency set
  with a clean moderate-or-higher registry audit. The update resolves the
  repository's transitive PostCSS advisory with PostCSS 8.5.25.
- Added a 20-chapter, zero-background learning path covering the complete OKF
  Explorer foundation: browser and static-app concepts, Markdown and bundle
  publication, large-corpus loading, deterministic retrieval, graph and
  semantic-web modelling, validation and provenance, coordinated views,
  geospatial and legislation data, AI infrastructure and federated AI,
  OKF v0.2 trust and lifecycle, bundle federation, Foundry authoring, governed
  enrichment and release assurance, security and accessibility, publication
  checks and contributor workflows.
  The documentation landing page, repository guide, README and advanced
  ontology architecture now route unfamiliar readers into the relevant
  beginner chapters.
- Added a general two-stage OKF authoring protocol: a read-only domain and
  standards warm-up followed by a deterministic build, assurance and
  publication controller.
- Added the versioned `okf-domain-profile.v1` JSON Schema and a complete YAML
  template so research decisions pass between tasks as a bounded,
  hash-lockable contract instead of an unbounded transcript.
- Added a profile validator for JSON Schema, evidence/task/rights/decision
  cross-references and byte-independent JSON/YAML data equivalence.
- Made the handoff carry an explicit collection profile, material claim
  register, calibrated-confidence rules, source format/identity/version
  behaviour, identifier schemes, task risks and standards conflicts so the
  machine contract can represent every fact required by the warm-up prompt.
- Encoded source-native identity, orthogonal authority/evidence axes,
  standards applicability, tiny-fixture preflight, immutable acquisition,
  content-addressed reuse, optional governed model enrichment, frozen-candidate
  security and byte-identical RC promotion.
- Grounded the protocol with cross-domain checks for UK legislation, ONS data
  discovery and GOV.UK content while retaining the strict boundary between OKF
  0.2 core conformance and the additive Foundry production profile.

## v0.5.4 - 2026-07-27 - Pages artefact closure

- Preserved the deterministic Svelte fallback during final Pages assembly
  instead of replacing its manifest-bound `404.html` with a second file.
- Added a fail-closed post-assembly check that rehashes every application
  material declared by the canonical build manifest before Pages upload.
- Retained the independently observed v0.5.3 mismatch as evidence for the
  corrective release rather than weakening the immutable-artefact gate.

## v0.5.3 - 2026-07-27 - Reproducible production builds

- Bound SvelteKit's application-version seed to the synchronised package
  version, removing its timestamp default from application-version metadata
  and client chunk generation.
- Added a fail-closed two-build regression that cleans generated state before
  each production build and compares every generated path and file digest,
  including the exact `index.html` bytes.
- Made both pull-request CI and the Pages publication workflow run the
  deterministic two-build proof before site assembly.
- Published the canonical per-file build manifest with the Pages app and made
  the release-bound runtime stage its exact bytes plus every described build
  file, closing the prior self-attested whole-tree digest boundary.
- Added fail-closed validation for missing, extra, duplicate, unsafe, linked
  and tampered build materials while retaining in-memory manifest derivation
  for mutable local v1 acceptance runs.

## v0.5.2 - 2026-07-26 - Frozen runtime evidence closure

- Made the release-bound legislation runtime receipt self-contained beneath its
  external evidence directory, with safe relative paths and byte/SHA-256
  identities for the runner, both bundle descriptors, the Pages build index
  and every Chrome screenshot.
- Added exact build-index/tree and screenshot-set integrity checks, the
  canonical external receipt basename, and an explicit WCAG 2.2 AA
  accessibility gate for the Whole-Law release controller.
- Kept the Explorer UI and browser journey unchanged while making a passing
  `okf-explorer-runtime-acceptance.v2` receipt independently rehashable.

## v0.5.1 - 2026-07-26 - Governed enrichment runtime hardening

- Added byte- and SHA-verified cross-binding for the accepted assertion
  manifest, independent audit and reviewer receipt before model-assisted
  relationships can become available.
- Made model-enrichment shard loading fail closed and retryable, with global
  assertion-ID uniqueness, exact route/predicate/target validation and
  incomplete-route recovery that cannot be hidden by cached healthy shards.
- Preserved and displayed assertion rights provenance, restored reactive
  availability state, and announced model-enrichment status and failures to
  assistive technology.
- Updated the legislation runtime and evaluation journeys for compact facets,
  bounded record hydration and the current 906,754-relationship publication.

## v0.4.2 - 2026-07-11 - Explorer identity alignment

- Aligned the private Explorer package and citation metadata with the current
  independently published Explorer product and patch release.
- Replaced the inherited AI Infrastructure bundle citation identity with the
  generic OKF Explorer software identity, canonical repository and Pages URL.
- Updated the live GitHub repository description and homepage to describe the
  federated Explorer, registry, profile and conformance-tooling role.

## 2026-07-11 — Federated bundle publication and YAML-LD foundation

- Adopted independently published OKF bundle wikis as the production
  architecture, with the Explorer acting as a generic consumer and curated
  registry rather than the canonical corpus host.
- Added the experimental OKF Bundle Wiki Profile v1, versioned JSON-LD context,
  JSON Schemas, SHACL shapes and a YAML 1.2/YAML-LD parser with pinned context
  loading.
- Replaced the line-based Markdown frontmatter parser with structured parsing
  while retaining the legacy Explorer projection.
- Added a single YAML-LD registry source that generates Explorer JSON and
  JSON-LD projections.
- Added a machine-readable source constraint ledger and human escalation guide
  so fair-use, access-control and licensing concerns remain visible without
  silently removing prototype features.
- Added semantic identity metadata to large-corpus descriptors and deterministic
  route-scoped relationship adjacency shards, allowing the Explorer to show a
  selected concept's links without hydrating the full relationship table.
- Added governed model-assisted legislation enrichment with a reproducible
  Responses API runner, checked-in reviewed rules, explicit cost/usage logging,
  universal provenance-bearing subject/type/entity edges and compressed
  route-scoped adjacency. The initial API project quota failure is retained in
  the constraint ledger rather than reducing the prototype's feature scope.
- Published the complete legislation bundle independently at
  `chris-page-gov/okf-uk-legislation` with protected `main`, green CI, GitHub
  Pages and release `v0.2.0`; the Explorer registry now treats that publication
  as canonical.
- Published the 41,520-record UK Government APIs bundle independently at
  `chris-page-gov/okf-uk-government-apis`, with protected `main`, CI, Pages,
  276,996 relationships, route adjacency and release `v0.4.0`.
- Published the 155-concept/579-relationship AI Infrastructure Markdown wiki
  independently at `chris-page-gov/okf-ai-infrastructure` with YAML-LD,
  JSON-LD, protected `main`, CI, Pages and release `v0.4.0`.
- Renamed the product repository to `chris-page-gov/okf-explorer`, moved the
  Svelte Explorer to the Pages root, retained `/next/` as a query/hash-preserving
  compatibility redirect, and made the independent AI bundle the default.
- Released Explorer v0.4.0 support for machine-readable `okf-moved.v1`
  descriptors so legacy bundle URLs can forward to canonical independent
  publications without copying their corpora.
- Checked in and applied the standard protected-`main` policy: strict required
  CI, one approval, stale-review dismissal, conversation resolution,
  administrator enforcement, linear history, and no force pushes or deletions.
- Overrode the SvelteKit transitive `cookie` dependency to the patched 0.7
  series, closing the low-severity `GHSA-pxg6-pf52-xh8x` alert reported after
  the repository rename.

All notable changes to this repository are recorded here. Entries are grouped by
date and describe the user-visible publication effect, validation run, and any
source-of-truth changes.

## v0.5.0 - 2026-07-26 - Whole-Law federation and bounded runtime

- Added governed metadata-term registries and closed-world validation reports
  to the bundle-wiki profile. Explorer now provides searchable bundle-wide and
  record-scoped term definitions, provenance and validation evidence; bundle
  UI terms can also supply existing `(i)` help text. Semantic-resource labels
  are derived from their actual YAML-LD or JSON-LD serialisation.
- Added a reproducible production-build acceptance runner and evidence receipt
  for the final local Whole-Law/Legislation publication. Chrome, Firefox and
  WebKit now prove federation-child completion, sub-1 MiB compressed startup,
  bounded cold/warm search, facet layout, coloured graph relationships,
  keyboard operation and release-blocking axe results.
- Bounded histogram segment counts by the provider presentation contract so
  24-pixel keyboard targets no longer overflow the navigation panel, and
  normalised source-native `derived-non-official` authority to the distinct
  derived graph style without losing its provenance label.
- Added the additive `okf-explorer-federation.v1` control plane and
  `okf-relationship-assertion.v2` contract. Federations load only their
  overview first; child descriptors and data shards remain untouched until a
  user selects a child.
- Added fail-closed child coverage, authority, freshness, discovery-route and
  relationship-summary validation. Official, deterministic-derived,
  model-assisted and unclassified relationships now have distinct labels,
  detail evidence and graph styling.
- Added declared descriptor fallback routes so Pages, raw-content and mirrored
  descriptors can recover in a fixed order without probing guessed repository
  paths. Repository, documentation and archive links remain visible recovery
  destinations but are never parsed as descriptors.
- Added bounded YAML-LD loading for `.yamlld`, `.yaml` and `.yml` documents,
  including safe JSON content sniffing when static hosts use
  `application/octet-stream`. YAML uses the 1.2 core schema, unique keys, no
  merge keys, no aliases and no custom/executable tags.
- Registered the future UK Whole-Law OKF descriptor and its declared raw
  fallback, repository, documentation, release archive and repository
  subpath.
- Made single-child federation landings useful above the fold and added a
  separate compact inventory for researched source families, so planned legal
  source coverage is visible without misrepresenting it as implemented child
  bundles.
- Prevented complete facet indexes from falling through to whole-corpus record
  hydration while a search worker is starting or lacks filter postings.
  Low-cardinality searchable facets retain coloured distributions, facet
  inventory counts are explicit, stale federation routes are cleared when a
  child opens, and a 50,000-record safety gate rejects unsafe full-index loads
  with a recoverable explanation.
- Added `okf-record-locator-sharded.v1` support so a selected result hydrates
  one integrity-aware record shard by route or search ordinal. Corpus-wide
  relationship hydration is likewise bounded at 100,000 advertised rows;
  larger bundles continue through summaries and route-scoped adjacency.
- Replaced the intermediate-width stacked workspace with a two-column
  navigation/stage layout and a collapsible detail row, preserving the
  overview-first Explorer experience in the in-app browser.

- Made OKF Explorer a dual-version OKF v0.2 exemplar. Structured
  `generated`, `sources`, `verified`, lifecycle, freshness and Attested
  Computation fields now have first-class passive presentation; v0.1
  `timestamp` and body `# Citations` remain labelled fallbacks, with v0.2
  fields taking precedence. Loading a bundle never executes its computation,
  executor or attester.
- Migrated the canonical AI infrastructure Markdown to v0.2, including
  spec-shaped reserved indexes/log, actor-bearing generation and verification,
  lifecycle status and structured provenance. Generators for the small bundle,
  UK Government APIs and UK Legislation now emit v0.2 Markdown while retaining
  the existing YAML-LD, federation, static-search, facet, datapack, integrity
  and presentation extensions. The migration uses the corpus's Git authoring
  commit for generation time instead of treating source publication dates as
  provenance events. Actor-free legacy `verified: yes` flags remain
  unverified rather than being upgraded into invented human review events.
- Split documentation and checks between permissive OKF v0.2 core conformance
  and the stricter Explorer authoring profile. Structured and unknown
  frontmatter fields/types survive generated bundle projection.
- Accepted the dual reader against a locally rebuilt pre-migration `okf-ons`
  publication and its v0.2 successor: both load all 5,097 records, 12 facet
  bars and the exact 382-record region reduction, while only the v0.2
  descriptor advertises its core Markdown layer.
- Changed large-corpus Timeline to distinguish the period represented by a
  release from the date its catalogue metadata was harvested. Repeating
  datasets are grouped into series and compact year/month release links;
  catalogue timestamps are used only as an explicitly labelled fallback.
- Added structured record comparison. Explorer separates other releases in
  the same series from genuine alternative datasets, links each item to its
  own durable route, displays declared field differences, and suppresses
  legacy `Compare before selecting:` prose when the structured alternative
  records are present. Title-derived series are explicitly labelled as a
  presentation aid rather than asserted semantic identity.
- Changed Map to use a bounded viewport with an internally scrolling result
  list and compact horizontally scrollable reduction controls. A labelled
  OpenStreetMap reference layer supplies geographic context while the
  metadata, coverage outline and result list remain usable if map tiles fail.
- Fixed single-predicate facet graphs to use the same stable relationship
  region layout as other metadata fan-outs, retain an accessible Layout
  control, and avoid the inconsistent diagonal stack previously shown for
  values such as `every 12 weeks`.
- Added hierarchy-aware, diverse high-cardinality facet previews. Declared
  bundle hierarchies take precedence; legacy facets receive bounded Year,
  Format, Region and Other summaries instead of a count-only sample. Documented
  SKOS hierarchy authoring in YAML-LD and added SKOS/RDFS/OWL prefixes to the
  pinned bundle-wiki context.
- Changed committed view, graph-key, label, node-type, relationship-highlight
  and graph-layout choices to create addressable browser history entries.
  Re-selecting or Ctrl/Cmd-clicking a highlighted graph node or relationship
  now clears it. The compact graph toolbar again exposes Auto versus
  relationship-region layout.
- Locked the focused node's type in the visible node key so a stale or newly
  selected hide-type control cannot imply that the graph focus can disappear.
- Added a maintained viewer capability inventory, feature matrix, conflict
  register and regression contract across the classic OKF canvas, modern
  LLM-Wiki, CFTE, OKFR, original CKAN and canonical Svelte lineages.
- Added safe small-bundle Markdown tables and Mermaid-lite flowcharts, plus
  exchange-aware Narrative and Timeline presentations that preserve user
  prompts, final answers and chronological commentary.
- Changed small- and large-corpus graphs to use complete non-overlapping
  node-and-relationship label layers on a two-second cycle, with persistent
  selected context, Pause/Resume and reduced-motion handling. Reciprocal
  directed relationships now use distinct curves, collapse identical labels
  and place different labels nearer their sources.
- Added predicate-and-direction relationship controls for large focus graphs:
  Auto or explicit semantic regions, group/member visibility, drag and button
  ordering, reset and durable `graph.*` URL state. The node key now lists only
  types in the displayed graph, and left-list labels remain visible in an
  above-left placement that keeps them out of centre-facing relationships.
  Controlled list and staircase regions keep every node label visible without
  cycling, paint labels above node symbols, compact vertical lists, spread
  staircases across the available width and separate same-side relationship
  lanes. The logical canvas now follows the centre panel aspect ratio, and the
  lower-right staircase exit is reserved before right-list placement. Repeated
  edge predicates are consolidated to one label per group; only conflicting
  edge labels cycle.
- Made graph and relationship controls compact and sticky within the centre
  panel. Plain wheel input now scrolls the panel, while the slower graph zoom
  requires Ctrl/Command+wheel or the explicit buttons.
- Replaced the wrapping graph key and layout panels with a two-line control
  surface. The first line keeps zoom, label-set cycling, node types,
  and relationship types available; the second line switches between
  only the node and predicate types present in the current graph. Node types
  can be hidden without losing their restore control, while selecting a
  relationship type highlights its source nodes, target nodes and directed
  edges. Relationship inspection now separates Source, Relationship and Target
  evidence into three tabs, and SVG keyboard focus follows the edge instead of
  drawing a rectangular outline across unrelated nodes. The Nodes and
  Relationships mode buttons show their current visible counts. Redundant
  on/off wording and the exposed Layout panel were removed; pressed styling
  identifies the active key while the automatic relationship-region layout is
  retained. Focus graphs now repeat the focused node name as an SVG title and
  anchor publisher and licence nodes at the lower left and lower right, giving
  exported screenshots a document-like final line.
- Fixed the dual-collapsed workspace grid so both context rails remain 44px,
  and fixed wrapped relationship-drawer rows so their text cannot overlap.
- Added edge width for an explicit varying numeric relationship metric while
  keeping constant or undocumented scores neutral and arrowheads fixed-size.
  Persistent label placement now looks ahead so a focus label cannot consume
  the only viable position for a boundary-node label.
- Added a layered ontology and semantic graph architecture covering RDF/JSON-LD
  instances, RDFS/SKOS vocabularies, bounded OWL 2 RL-style inference, SHACL
  validation, DCAT/PROV evidence, predicate registries and the strict
  separation of semantics from Explorer presentation.
- Fixed graph panning that previously failed when a drag began over a node or
  relationship, suppressed native drag ghosts, retained context in folded side
  rails, and made detail panels reliably touch-scrollable on mobile.
- Fixed invisible node and label interaction rectangles inheriting white SVG
  borders that cut through graph labels.
- Added unit and Playwright regression coverage for label-layer completeness,
  reciprocal edges, the ONS 21-node/20-relationship focus graph, graph panning,
  safe rich Markdown, conversation views, collapsed context and Android-sized
  detail scrolling.
- Added optional `okf-explorer-provider-datapack.v1` loading for large corpora,
  with formal pack/manifest schemas, required governed-snapshot bindings,
  SHA-256 manifest/pack integrity, bundle-base-path validation, non-exhaustive
  reviewed comparison evidence, HTTPS-only provider actions, and visible
  bundle, record/search and inherited resource status. Matching records outside
  the bounded comparison are labelled as not reviewed, while paired examples
  without a recorded difference are labelled as aligned only in reviewed
  fields. Paired rows with changed known fields require exact, unique difference
  evidence and fail closed when it is missing; declared differences must also
  reference paired examples and match their known values. The deterministic ONS
  fixture demonstrates a governed April 2026 Average house price snapshot
  alongside a May 2026 reviewed upstream reference without claiming live
  validation.
- Added an experimental bundle-level Explorer presentation profile and a more
  compact facet workspace: provider labels/order/control defaults, bundle-local
  user pin/hide/reorder preferences, complete-data distribution strips,
  numeric/date histograms, search-first high-cardinality previews, hierarchy
  browsing, and tabbed result/detail panels. Multi-valued facet reduction is
  now bounded and assignment-aware.
- Removed complete single-valued dimensions from Suggested facets because they
  cannot reduce the current result set. They remain available under All and
  remain visible when selected, explicitly shown or pinned.
- Restored small-bundle graph and link compatibility with generator-produced
  `edges`, added Markdown-body search and safe body rendering, exposed
  credential-redacted source/resource links, and made selected Schema.org,
  provenance and complete node JSON available from the detail pane.
- Updated the SvelteKit, Svelte, Vite and `svelte-check` development stack while
  retaining TypeScript 6.0.3 until the Svelte checker supports TypeScript 7's
  changed module shape.
- Added repeatable post-deployment Playwright execution against the public
  GitHub Pages Explorer, hardened the browser evaluator against cold large-index
  startup, and corrected journey validation for canonical default-sort URLs and
  keyboard relationship selection.
- Added a deterministic geospatial Map canvas for both small and large OKF
  bundles. It classifies declared coverage, coordinates, UK places,
  ArcGIS/OGC services and spatial files; persists Map reductions in public
  URLs; keeps representative centroids distinct from boundaries; and limits
  direct GeoJSON/ArcGIS geometry loading to explicit bounded preview actions
  with source-link recovery.
- Added dedicated Map personas and user stories for area-based policy analysis,
  spatial data analysis, provenance-preserving inspection, locator precision,
  bounded previews, progressive recovery, durable URL state, accessibility and
  trustworthy pack authoring, with machine-readable links to focused browser
  assertions.
- Added an 18-scenario Playwright Map suite covering deterministic small and
  large bundle paths, every reduction and selection control, exact versus
  representative locator states, GeoJSON/OGC/ArcGIS success and cap handling,
  linked-only and failed-resource recovery, durable URL history, keyboard and
  responsive behaviour, empty/loading states and the bounded record list. CI
  now runs the suite, and documentation lockstep treats its tests, config and
  journey traceability as controlled publication surfaces. Journey validation
  also fails on missing or untraced `GEO-E2E-*` identifiers.
- Fixed the UK place vocabulary to recognise canonical display labels such as
  `London` and `North East`, preventing recognised areas from also appearing
  under Other declared coverage.
- Hardened deterministic large-corpus retrieval against stale bundle responses,
  silent or crashed workers, malformed facet values, metadata-gap filters and
  approximate capped-posting totals. Dynamic facets and result summaries now
  remain explicit about fallback and truncation semantics, and legacy small
  bundles normalise scalar alias/tag fields before client-side searching.
- Added persona, user-story and question traceability for every hosted README
  exemplar, plus opt-in browser journeys for retrieval URL state, facets,
  sorting, graph relationships, drawer resizing, folded full-record sections
  and source inspection without replacing the Explorer window.
- Kept the graph relationship drawer above the sticky pins bar so its resize
  grip receives real pointer drags as well as keyboard input.
- Changed broad large-corpus searches that span more than the safe result-chunk
  budget to return the highest-ranked bounded subset instead of failing. The
  result summary retains the full match count and explains the browser-memory
  safeguard without implying AI token usage.
- Added backward-compatible large-corpus delivery through integrity-bound,
  same-origin byte-range packs. Descriptor, search-worker, record-chunk and
  route-adjacency fetches preserve their logical paths while enforcing the v1
  pack index, 64 MiB bounds, exact HTTP 206 ranges, dual SHA-256 checks and
  gzip decoding limits; existing direct static bundles continue unchanged.
- Added optional `okf-operational-metadata.v1` sidecar loading for large
  corpora, allowing source-backed operational enrichment to be refreshed
  without rewriting dataset or search chunks.

### Added

- Added an in-Explorer Source Inspector for machine-readable record links, with
  a human summary, searchable collapsible JSON tree, raw view, copy controls,
  response provenance and a bounded 10 MB client-side fetch. Known legacy
  data.gov.uk action URLs resolve through the canonical browser-readable GOV.UK
  CKAN action endpoint.
- Added an optional, provenance-bearing `operational_metadata` contract for
  canonical source, authoritative publisher, update frequency, current release,
  full/delta distributions, API access, technical specification and licence.
- Added a complete work-level UK legislation OKF pack generated from every year facet in the official legislation.gov.uk Atom catalogue, including primary, secondary, draft, devolved, retained-EU and historical document types.
- Added ELI 1.5 / ELI-I, Schema.org `Legislation`, legislation.gov.uk FRBR/CLML and Akoma Ntoso crosswalks, normalised legal categories and title-derived topic navigation with explicit non-authoritative classification warnings.
- Added progressive official CLML discovery in the Svelte Explorer. A selected work can resolve every Part, Chapter, section, article, regulation, rule, schedule, paragraph and nested P1-P7 structure with official passage links and copyable provenance citations.
- Added combined static-title and official remote full-text search, legislation-specific facets, complete access-method documentation and disclosure of the currently authenticated research bulk/SPARQL surfaces.
- Added a 100-question barrister-oriented AI-answer evaluation suite, 100-point legal/provenance rubric, answer JSON Schema and deterministic evaluator with a hard cap for missing official or proposition-level provenance.
- Added corpus completeness validation, fixture generation in CI, Pages publication wiring and a registry entry for the shareable hosted legislation viewer.
- Added deterministic gzip support for large dataset/search chunks in the Explorer and generator, preserving the complete corpus within practical repository and Pages sizes.
- Added a maintained UK Legislation documentation spine covering getting started, personas and user journeys, an illustrated legislation-specific manual, agent research, evaluation and publication maintenance.
- Added four hosted-Explorer illustrations with a machine-readable capture manifest so corpus overview, exact-title search, work provenance and live CLML passage discovery can be refreshed reproducibly.
- Promoted UK Legislation to a first-class hosted example in the README opening and cross-linked the documentation spine from repository, publication and architecture guides.

### Changed

- Made **View source data** the primary source-record action and relabelled the
  direct endpoint as **Open raw JSON ↗**. Raw external responses always open in
  a separate tab, so they cannot replace the current Explorer state.
- Relabelled harvested CKAN dates as catalogue metadata and added a current
  source/maintenance disclosure. Resource hosts remain discovery leads until a
  bundle supplies canonical-source evidence and provenance.
- Kept Explorer search fully static and deterministic while adding durable
  query, repeated facet-filter and sort state to public URLs, including
  Back/Forward restoration and compatible small-bundle type filtering.
- Added backward-compatible `okf-static-search.v2` filter postings, missing
  metadata buckets, filter-before-limit execution, dynamic facet counts,
  structured match explanations and deterministic relevance/newest/title/
  metadata-quality sorting, with the existing v1 full-index path retained as a
  correctness fallback.
- Split the retrieval panel into Search, Filter results and Sort controls, added
  removable active-filter chips and meaningful candidate totals, and replaced
  raw scores with plain-language "Why this matched" evidence.
- Added a reproducible 30-query ranking benchmark for weighted, field-weighted
  IDF and IDF-plus-exact-boost strategies. The benchmark gate retains the
  current weighted default because neither candidate reaches the required 3%
  macro nDCG@10 improvement.
- Added an illustrated static search and filtering manual captured against the
  58,461-record GOV.UK CKAN corpus, covering deterministic match explanations,
  repeated facet filters, sorting, result totals, v1 fallback and v2 indexed
  execution.
- Fixed large-corpus browser searches by copying reactive retrieval state into
  structured-clone-safe data before posting requests to the search worker.
- Standardised absent and legacy `None`/`null` metadata as `Not specified
  (metadata gap)` in record details and illustrated the wording in the new
  manual.
- Replaced the ambiguous large-corpus `tokens` label with `distinct indexed
  terms` and an explicit explanation that the count is local search-index
  vocabulary, not AI or billable model usage.
- Added deterministic organisation-aware retrieval for exact publisher names
  and unambiguous abbreviations. Existing CKAN v1 bundles reuse their compact
  delta-encoded publisher postings, while new v2 bundles can publish explicit
  entity aliases through `data/search/entities.json`; recognition remains
  visible in suggestions and match explanations.
- Fixed the graph relationship drawer so its larger drag target changes a real
  row height, retains pointer capture and also supports keyboard resizing.
- Kept an inspected relationship visibly selected in both the drawer and graph
  instead of clearing edge-only highlight state during selection reconciliation.
- Added an early record-date summary plus a contextual Dates and related records
  block. It distinguishes source update dates from resource/coverage years,
  follows explicit series metadata, and says when the bundle contains no other
  record in that series rather than guessing from title similarity.
- Moved the lightweight search card's Load full record action beside the title
  and changed secondary detail-card sections to folded disclosures, leaving the
  first Overview section open by default.

## v0.3.0 - 2026-07-09 - Standards-Aligned API Demonstrator

- Added `docs/okf-standards-crosswalk.md`, a field-by-field crosswalk between
  the OKF record contract and DCAT/DCAT-AP (`dcat:DataService`) and OpenAPI,
  including a record-type crosswalk, an access-model to
  `securitySchemes.type` mapping, and a worked example built from the
  Ordnance Survey `search/places/v1/postcode` record. Cross-linked from
  `docs/index.md`, `docs/repository-guide.md`, `docs/okf-bundle-authoring.md`
  and `docs/ai-okf-usage.md`.
- Added local standards concept pages for [DCAT](standards/dcat.md),
  [DCAT-AP](standards/dcat-ap.md) and [DQV](standards/dqv.md), and refreshed
  [OpenAPI](standards/openapi.md) so API-related OKF bundles can link to local
  standard metadata as well as official sources.
- Extended the UK Government APIs builder so every generated API/data record
  carries compact standards-alignment metadata: `dcat_type`, `openapi_type`,
  DCAT/OpenAPI export status, OpenAPI security-scheme mapping and missing
  standards requirements for future DCAT/OpenAPI exporters.
- Updated the Explorer detail card to render DCAT/OpenAPI terms in monospace
  with explanatory info bubbles and a Standards Alignment section.
- Expanded the crosswalk, API source specification, bundle-authoring guide, AI
  usage guide and pack-parity notes with a standards gap analysis and cautious
  export-readiness policy.
- Changed the Pages build so the root URL redirects to the canonical Svelte
  Explorer under `next/`, while the old dependency-free Explorer is published
  explicitly under `legacy/` and the single-file legacy viewer remains available
  through `viewer.html` / `view.html`.
- Added a legacy Explorer handoff for large-corpus descriptors and clarified
  facet-detail totals so matched record counts are not confused with capped
  preview rows.
- Softened selected facet and record highlighting in the Explorer so active
  records remain readable without the heavy saturated-blue card treatment.
- Fixed graph stack double-click handling so synthetic stack nodes are
  highlighted or expanded in place instead of becoming unreadable graph-centre
  routes.

## v0.2.0 - 2026-07-08

### Added

- Added `docs/index.md` as the documentation landing page for repository
  navigation, Explorer use, AI usage, bundle authoring, evaluation and dated
  review records.
- Added `docs/repository-guide.md` so contributors can find the publication
  pipeline, stable public URLs, validation commands and source-of-truth
  boundaries.
- Added `docs/ai-okf-usage.md` with prompts and access rules for pointing an AI
  at small OKF bundles or large-corpus descriptors without losing provenance.
- Added `docs/okf-bundle-authoring.md` with the minimum record contract,
  Explorer feature contract, facet/relationship guidance, metadata-repair
  rules and acceptance checklist for new OKF bundles.
- Added `docs/okf-explorer-persona-manual.md`, a screenshot-led manual that
  documents Explorer UI behaviour through personas and user stories.
- Added `docs/demo-script-2026-07-09.md`, a 20 minute demonstration script for
  the UK Government APIs OKF exemplar and Explorer workflow.
- Added current Explorer screenshots under `docs/assets/okf-explorer-manual/`
  covering overview, Provider facet, search, record card, graph, timeline,
  type, resources and Rugby search states.
- Added `scripts/check_documentation_lockstep.py` and wired it into
  `okf-explorer-ci` so human PRs that change publication-critical source,
  generated corpus, tests, workflows, or Explorer source must also update
  documentation and this changelog.
- Added a 100-question OKF Explorer evaluation suite, additive 100-point
  scoring rubric, visual-regression manifest, and browser harness at
  `scripts/evaluate_okf_explorer.mjs`.
- Added graph-overlap screenshot evidence so relationship-label layering,
  overlapping white boxes, and arrow-to-icon placement remain testable review
  concerns.
- Added a separate 100-question GOV.UK CKAN evaluation suite with the same
  additive rubric and pack-aware `target_bundle` support in the harness.
- Added `docs/okf-pack-parity.md` to define parity expectations between the UK
  Government APIs OKF exemplar and the GOV.UK CKAN large-corpus pack.
- Added OS Data Hub visual-regression evidence for search-context loss,
  unreadable dense graph clusters and misleading record-type breakdowns.

### Changed

- Marked the 2026-07-07 code review as a historical record superseded by the
  2026-07-08 follow-up and later `main` merges.
- Updated the 2026-07-08 code-review follow-up with a current-state checkpoint
  and final status table after remediation and later Explorer fixes were merged.
- The UK Government APIs OKF builder now canonicalises OGL licence variants to
  `open-government-licence-v3` and records explicit licence basis/confidence on
  each API/data record.
- ONS records with no source-declared licence now infer Open Government Licence
  v3.0 from ONS terms and mark the record as
  `provider-terms-inferred`, preserving lower confidence than directly declared
  licence metadata.
- Ordnance Survey provider-native API records now infer
  `ordnance-survey-licence-required` from OS licensing guidance instead of
  remaining `not-specified` or being incorrectly treated as OGL.
- OKF Explorer large-corpus search now prepares the static index in the
  background, debounces typing, and shows explicit search/index status.
- The right-hand record card now uses clearer metadata section titles, clickable
  topic/format/tag chips, and info bubbles for licence basis, evidence counts
  and metadata-quality percentages.
- Dense graph relationship rows are stacked into count-bearing graph nodes and
  the relationship list is shown as a drawer-style panel with its own scroll
  area.
- Dense large-corpus graphs now group API/data records by record type, expose a
  visible "Grouped by record type" caption, and expand one record-type stack at
  a time.
- Dense graph stacks now count the full matching reduction while expanded
  stacks show a bounded sample with the sample size stated in the caption.
- Opened graph stacks with many records now sub-group by an available semantic
  facet such as format, topic, licence, access model, contract status, source
  adapter or update year, so dense OS/Data Hub clusters no longer expand into a
  single unreadable fan-in.
- Large-corpus facets now hide duplicate `canonical_publisher` navigation when
  `publisher` is available, expose an in-facet search box, and reveal values in
  pages instead of capping the list at 18.
- Large-corpus facets now render only the active facet body, keep the facet open
  after a value is selected, and normalise hyphen/underscore text consistently
  between the facet search query and the searched values.
- Facet value clicks now behave as single-select by default; Ctrl-click,
  Cmd-click or Shift-click adds/removes values for multi-select filtering.
- Graph node glyphs now distinguish providers, formats, topics, licences, tags
  and hosts/resource types with different compact SVG icons while preserving
  selected and stacked states.
- Graph legends now show the same shape vocabulary used in the graph, including
  opened stack groups.
- Double-clicking graph metadata nodes such as provider, host, format, topic,
  tag or licence now applies the corresponding facet reduction when available,
  so left-panel counts and graph context stay aligned.
- The graph view now separates inspection from graph navigation: single-clicking
  a graph node updates the data card, while double-clicking a navigable node
  recentres/reduces the graph context.
- Large-corpus search now exposes an explicit clear button and clears stale
  selected-record context as soon as a materially different query is typed.
- Bundle URL suggestions now close when focus or pointer interaction moves
  outside the URL control.
- Info bubbles for created, modified and timeline dates now use distinct scoped
  help keys so only the requested explanation opens.
- Large graph arrows now use source and target node shape padding so arrowheads
  terminate at the visual edge of stack/card/circle nodes.
- The graph relationship list now has a centre drag handle for resizing its
  drawer height.
- The large-corpus timeline now defaults to newest dated records first and can
  be viewed as latest records, years, quarters or months; clicking grouped time
  buckets applies the matching date facet.
- Large-corpus in-app Back/Forward now preserves the inspected data card state
  when returning from graph metadata inspection.
- Large-corpus full-index hydration now fetches static JSON chunks in smaller
  batches and retries transient CDN/server failures, reducing the risk of a
  GitHub Pages `503` when opening high-cardinality facets such as Provider.

### Fixed

- Fixed transient GitHub Pages shard failures during Provider facet hydration by
  reducing chunk-fetch concurrency and retrying HTTP 503/5xx responses before
  surfacing an Explorer error.
- Fixed the two open PR review issues: closed facet bodies no longer trigger
  repeated full-corpus facet scans, and facet search tokens are normalised with
  the same hyphen/underscore handling as facet values.
- Reduced the misleading `not-specified` licence gap for ONS CKAN-derived data
  products, data access endpoints, generated contract records, and OS
  provider-native API records.
- Generated Markdown concept pages now expose licence, licence source, licence
  basis, and licence confidence.
- Lightweight search-result detail cards now preserve licence metadata and use
  the same metadata-quality and timestamp wording as fully hydrated record
  cards.
- Graph relationship arrows now render to the trimmed visual boundary of the
  source and target icons rather than passing through card/icon centres.

### Validation

- `python3 -m unittest discover -s tests`
- `python3 scripts/check_documentation_lockstep.py`
- `python3 scripts/build_okf_bundle.py --check`
- `python3 scripts/update_viewer.py --check`
- `python3 scripts/check_okf.py`
- `pnpm test`, `pnpm check`, and `pnpm build` in `apps/okf-explorer/`
- `python3 scripts/build_uk_government_api_okf.py --check`
- `python3 scripts/build_site.py`
- `node scripts/evaluate_okf_explorer.mjs --base-url http://127.0.0.1:8002/_site/next/ --bundle ../uk-government-apis/okf-explorer.json --limit 100`
- `node scripts/evaluate_okf_explorer.mjs --no-browser --suite evaluation/gov-ckan/questions.json`
- `node scripts/evaluate_okf_explorer.mjs --no-browser --suite evaluation/okf-explorer/questions.json`
- Live GitHub Pages header checks for
  `uk-government-apis/okf-explorer.json` and `uk-government-apis/data/apis-9.json`.
- Targeted Playwright smoke against `http://127.0.0.1:8002/_site/next/`
  covering facet search/open state, graph legend/drawer, bundle suggestion
  dismissal, and timeline latest/quarter ordering.
- `git diff --check`

## Learning presentation for small bundles

- Added an opt-in learning path with ordered goal cards, contextual record pages,
  connections and expandable source evidence.
- Small bundles can declare conceptual facets. Source and review filters remain
  available as secondary information; existing bundles retain their defaults.
- File imports now honour a valid authored starting route and clear stale
  inspections. URL-loaded bundles retain explicit deep links and filter state.
- Added a reusable content-design guide and tests for presentation validation,
  legacy behaviour and conceptual facet intersections.