Independent experimental publication. Not an official DWP service, benefits advice or an entitlement calculator.

Changelog

Source-linked reading-help release — 27 September 2026

Bind the 513-document corpus, 12/24 source controls, 28-case and nine-chapter rechecks, unchanged 40-question replay, and 401/403 exact workbench links in a checked release manifest.
Record all 16 successful public browser checks at Explorer b633038c, preserve earlier failures and the Chapter 60 v1 fallback, and publish the beginner demonstration script.
Add CI checks that reject mismatched source, verifier, consumer, data and Pages identities. Candidate meanings, extraction gaps and legal answerability remain explicitly unreviewed.

Notable changes for readers, in reverse date order. We use the Keep a Changelog distinction between unreleased work and dated changes; this repository currently uses dated research deliveries rather than claiming a Semantic Versioning release series. Implementation ownership and handovers belong in the multi-agent work log, with stable work items in the backlog.

Unreleased

Gate the versioned reading-help release record in CI: confined hash bindings, fresh 12/24 controls, all 40 retained packages, and exact verifier, consumer and data identities. Preserve failed public attempts.
Add a bounded public reading-help verifier: immutable catalogue and 40 sidecars, exact source text, continuations, cache measurements, keyboard focus, dated links, automated accessibility snapshots and the preserved Chapter 60 route. Existing receipts cannot be overwritten.
Clarify the optional cross-document target contract and distinguish its destination identity from the printed reference row. Frozen corpus rows remain unresolved where no target is declared.
Reconcile reading-help backlog acceptance with the authorised literal-highlight scope; proposed meanings and legal applicability still require qualified review.
Link the beginner glossary and learning-path documentation to the reading-help walkthrough. Keep the combined Reader and its frozen downstream context inputs unchanged; preserve the withdrawn learning-navigation observation separately.
Link all 40 retained question cases to the bounded reading-help catalogue: 401 of 403 distinct evidence records match exact unit and source identities. Keep the two unmatched capital composites explicit and every package unchanged.
Generate both backlog views from the authored register and record the passed source controls separately from pending public verification.
Correct statutory-overlay capture provenance and distinguish full-passage hashes from exact source fragments; preserve the failed admission and verify all newly admitted records with Explorer.
Correct the learning-path insertion to preserve the checked source-roles diagram heading.
Validate the retained 40-question replay bindings and measurements in CI without repeating unchanged assemblies.

Reading-help legislation bridge

Add a separately selected statutory context and 16 source-checked dated citation mappings. Preserve unresolved references, all earlier requirements and the failed fixed-budget natural-language selection probe.
Retain an 80-invocation offline replay of all 40 staff occurrences: cold and warm packages match the frozen baseline exactly, with latency and cache measurements. This is non-regression, not an answer-quality improvement.
Add a separate v1-compatible Chapter 60 manifest with 20 exact body/reference-row occurrence pairs. Keep the earlier manifest and unresolved citation meanings unchanged.
Add four source-span-bound work-title mappings and a separately validated statutory context overlay. Reuse the existing 20 statutory units and 43 navigation relationships.
Retain the five-request Chapter 60 acquisition and its initial section 70 version failure. An offline successor preserves both geographical variants, complete selected text and source-native dated links, without further requests.
Record exact rollout baselines and the failed live health observation separately from historical deployment evidence. Consumer admission and legal applicability remain explicit outstanding gates.
Add a deterministic DMG/ADM reading-help corpus over the adopted 53,727 structured units in 513 documents. Bounded, source-bound leaves retain 893 extraction-blocked pages, printed abbreviation-row candidates, unresolved local references and explicit review limits. The source-led initial gate passed 12/12 after a retained failed run; the first fresh held-out gate remains failed overall, with 23 passes and one failure on an expected label that omitted a printed range qualifier. A separately frozen and source-reviewed fresh 24-case gate passed 24/24, including abbreviation collisions, continuing tables, exceptions and unresolved references. The Chat pack now reports truncated selection and bounds input/output paths. The corpus leaf writer uses an OS-neutral, mtime-zero gzip header; warm-cache reads are confined to hash-bound leaves in their own document. This remains an unreviewed candidate; no model or legal review was run.
Add a Chapter 60 reading-help rollout proposal for deterministic-first expansion and later bounded model review; no wider trial or specialist acceptance is claimed.
Curate a beginner research guide and file-level index for the eight-file, 90-entry public CPAG contents capture, the 31-file unvalidated UC evaluation starter, concise client-access reports and one historical Pension Credit calculation design prompt. Preserve package checksums and distinguish observed tool fields, model-reported claims and unrun comparative proposals. No source corpus, runtime bundle or legal acceptance status changes.
Add a bounded Chapter 60 reading aid for paragraphs 60025 and 60033. Its exact occurrence spans, source-backed expansions, local references and unresolved printed defects are checked deterministically; original model-authored proposals remain unchanged. Broader highlighting and specialist review remain open.
Complete a bounded LiteParse 2.14.6 experiment: freeze 12 initial and 24 conditional source windows before outcomes, retain the first 12 paired extractions and their failed source-quality gate, and verify results offline in CI. Numeric table rows and coordinates are useful, but complete table associations and exact-anchor preservation do not pass. Stop expansion; preserve all corpus defaults and distinguish punctuation differences from missing guidance. No answer-model comparisons were run.
Retain a user-supplied Claude repository-access transcript with separated prompt, commands, outputs and final answer. Record its missing model/revision metadata and prior exposure to the development case; preserve unreviewed answer claims without treating the example as Ask OKF access, independent evaluation or specialist acceptance.
Publish service 0.7.0 with the Evidence Connect corpus and source-compatible assembler, preserving older replay pairs. The exact merged public SDK check passes twelve cases in 135 requests. Record the initial stale connector schema, unchanged-permission refresh and hash-checked native diagnostics read; retain sidebar, transport and specialist-acceptance boundaries separately.
Schedule passage-boundary inspection after the service/sidebar release in linked Explorer #150 and DWP #42, under the existing BL007 structural-review package. The first increment covers 28 known cases; later gates require nine-chapter impact review, all 40 fixed-budget replays and held-out controls. Model detection remains optional and requires an agreed call budget.
Separate saved workbench inspection, fresh Explorer evidence assembly and the versioned remote MCP connection in the client guide. Add a short remote catalogue/read starter and cache-refresh check. Retain the accepted 24 September sidebar observation, the initially unaccepted fresh-session attempt and the later scoped 25 September staff-016 native page-tool success. Neither sidebar run establishes dedicated WebMCP discovery, Voice, room audio or answer quality.
Separate the current status and ordered remaining work from historical checkpoints. Record the completed bounded model-inspection implementation, retain its separate client and specialist gates, and close superseded PR 23 without removing its history or the unfinished amendment work.
Record the branch/worktree merge audit, preserve the unfinished amendment parser and superseded handover separately, and document the observed Edge sidebar demonstration. Correct the workbench launch link and stale pre-publication wording without changing evidence or browser permissions.
Record a new, anonymised UC disabled-child question outside the frozen 40. The SHA-bound offline replay selects ADM F1123 but misses A4361's effective-date branch and remains insufficient. Add conditional source reading, unknown assessment-period dates, statutory and specialist gates, and separately labelled Demo 2 and hospital leads. Track the open BL006 legal check and BL007 source-classification repair in the machine backlog; do not change source, runtime ranking, public packages or answer status.
Add a beginner source-and-roles map to the learning path, with separate law, tribunal decisions, DWP staff guidance, independent CPAG handbook and claimant/adviser/decision-maker routes. The diagram and text keep source authority, editions and appeal routes distinct; no handbook content is copied.
Add an opt-in, source-bound workbench inspection model for Pension Credit dependencies and directional Carer’s Allowance interactions. The producer verifies exact retained package, record, page, hash and text bindings; execution stays blocked because rates, applicability, exceptions and independent review remain unresolved. Existing 40 packages and the public learning-site allowlist are unchanged.
Repair two Chapter 84 source boundaries, add ten proposed evidence groups and a source-linked typed table, and retain the final 52-assembly comparison. Correct summary provenance and reserved labels after agent review, preserving earlier trials. At 512 KiB it retains 11/11 declared groups against the newly authored baseline’s 9/11, a six-whitespace-byte difference; at 32 KiB neither arm delivers source evidence. All packages remain insufficient. Keep dependency closure, compact delivery, specialist review and production promotion open; no model calls or new source acquisition.
Repair the Demo 1 CI setup after the legacy job discovered the new tests in its own shallow checkout: fetch the same frozen source commit before test discovery. Preserve the five-error failure and a clean shallow-clone reproduction showing all five tests pass after the fetch; do not alter trial inputs or weaken the gate.
Retain the four authorised Demo 1 answers without retries: 12 exact source quotations, three format failures and explicit interpretation concerns. No token saving or specialist acceptance is established. Add the forty-question ledger and observe matching human-visible/WebMCP source text from the pinned public corpus.
Freeze a two-question, four-call subscription comparison protocol. Retain complete selected records in a labelled reading view, preserve the full audit packages and test the projection offline before any model invocation.
Fix the Demo 1 scope against the supplied GitHub brief: all 40 questions match exactly. Prepare a 15:00 BST freeze with the tested engine/source pair, explicit coverage gaps and a hard cap of four new answer calls. Park separate custody and historical-amendment repairs; do not infer answer improvements or token savings from retrieval results.
Expand source-read selection to 37/40 unchanged staff question occurrences, with complete retention of 47 inherited, 181 earlier and 208 new declared source-selection paths at 512 KiB. Preserve the three ambiguous questions, all 203 original obligations, 32 KiB refusal and reduced location-route retention. Restore original custody requirements in a separate evaluation plane. This is evidence preparation and delivery, not an AI-answer or specialist-acceptance result.
Repair the Reader's endpoint capacity failure without deleting aliases or raising its limit: index discovery aliases through their dedicated metadata channel. The full 53,727-unit corpus now loads with 19,193 relationships; exact source text, pages and historical classifications remain preserved.
Retain the successful fixed-source allocation trial: all 40 large staff packages contain source, with 47/47 inherited and 181/181 new source-read paths. Independent checking preserves 82 exact legacy results. Keep the location-navigation trade-offs, 32 KiB refusal and all 203 original obligations explicit; no model-answer improvement is inferred. Track the separately discovered original-imprisonment migration and Reader label-capacity gaps before source publication.
Repair the observed discovery delivery regression with a same-source comparison: all 47 inherited paths now remain at 512 KiB. Preserve the 32 KiB limitation, earlier failure and all insufficient results. Add generic conjunctive route guards, seven source-read selection profiles and separately labelled legacy location navigation; keep all 203 original obligations open. Final integrated evaluation and publication are still separate gates.
Retain the failed forty-question discovery-corpus comparison and its exact engine/source identities. Repair source-role boundaries after a full-corpus census found reserved ranges and short notices absorbing appendices; rerun the fixed four-case gate and doubled eight-case gate successfully. Evaluate runtime delivery and new semantic profiles separately before adoption.
Build an additive source-led manual guide, PDF-structure observations, complete source-byte unit catalogue and separate discovery-card corpus. Preserve authored unit identities and obligations; use explicit source references for navigation only. Retain the four-to-eight structural experiment, failed attempts and independent boundary/integrity controls. Full question comparison and public adoption remain separate gates.
Start the source-led manual-guide build: record the reading-convention discovery process, a four-case structural gate followed by eight cases after a pass, and the separate full-corpus/staff-question evaluation. Track manual guides, structural repair and discovery cards as active work; no new semantic or answer-quality result is claimed yet.
Fix two learning-site defects found during the preceding public check: keep the focused skip link outside page layout so pointer navigation stays stable, and render table alignment through external stylesheet classes without weakening the content security policy. Retain the failed observations and add renderer and desktop/mobile browser regressions; publication acceptance remains a separate exact-commit check.
Retain two independently checked offline ranking probes on the same 40 questions. Token-level alias expansion loses tracked candidate locations; length-aware BM25 gains four without losses but still selects incomplete or wrong-subject fragments. Preserve fixed protocols, exact inputs and all results. The live ranker, source units and answerability statuses are unchanged; summary-card implementation remains open.
Record the source-bound discovery diagnosis and a separate summary-card experiment: literal ranking does not use declared aliases, while a bounded sample exposes incorrect inherited headings. Broader retrieval improvement remains unimplemented and explicitly tracked.
Add 29 source-bound logical excerpts, scoped dependency routes and two PIP profiles through an explicit additive authoring registry; preserve old inputs and all unresolved legal obligations. Keep the F1093 reserved-range and P4019/P4051 scope mismatches visible.
Preserve the earlier logical evaluation before replaying 184 assemblies with 11 scope controls and the reviewed compact engine. Add a reproducible 40-occurrence coverage ledger: eight activate a unit profile, 32 do not, and all remain insufficient. Explain assembly and delivery budgets separately in the beginner documentation.
Allow 45 minutes for the complete source and consumer replay after the merged learning release exceeded the previous 30-minute CI limit. Keep every validation step and publication gate; the cancelled run remains recorded.
Rebind the logical-context projection after the combined descriptor changes, exclude combined-only learning routes from that Reader, and replay all 184 contexts. Preserve the prior evaluation under evaluation/logical-units/history/pre-learning-2026-09-22/; semantic-review limitations remain unchanged.
Add 12 demonstration learning paths and 112 activities to the combined Reader, covering all 40 supplied question occurrences. Bind authored objectives, source locators, prerequisites and facilitator public keys into the generated snapshot. Preserve the separate logical-unit corpus and all frozen pilot/source projections.

Passage review workbench — 25 September 2026

Add a source-bound passage review queue for the 28 known historical-amendment outliers, exact PDF and extraction identities, before/after spans and a beginner review guide. Preserve the parked proposal and its uncertainties.
Separate a narrowed amendment parser into an explicit successor profile. Reject the broad bare-Appendix rule after PDF review found it split a sentence. Preserve the frozen baseline and record corpus, identifier and qualification checks separately from retrieval and specialist acceptance.
Retain all 40 candidate evidence packages and an offline equal-budget replay check. Selection changes in 21 cases; expected page coverage and requirements do not improve. All results remain insufficient, with no new model calls.
Publish and verify the generic workbench at Explorer 4d76cddb: all 28 public cases, exact hashes for 34 app files, PDF rendering/page synchronisation and local review export pass. Record DWP merge c971e4cd and protected CI separately from specialist acceptance.
Mark the bounded workbench implementation delivered in the authored backlog and regenerate its report. Keep 12 partial boundaries, one unresolved source extraction and separate parser adoption open; full-corpus preview impacts remain unknown until a wider producer replay. See the release record.

Evidence workbench — 24 September 2026

Publish the additive full-DMG/ADM capital overlay, ten repaired source passages, source-backed links for four captured dependency groups, and exact bounded packages for all 40 public staff-question occurrences. Preserve frozen source and earlier trials; all 40 contexts remain insufficient and 11 dependency groups remain open overall.
Merge Explorer PR145 and DWP PR36 after protected checks, then verify the version 3 learning site against merged DWP commit 7eeded763 in 1,141 exact public requests. The deployed workbench opens all 40 cases and supports source inspection, deep links and local review export. The live capital Reader-to-Ask journey selects U07, DMG 84861 and their direct link, but remains truncated and insufficient. Specialist and legal acceptance remain open.

Logical evidence units — 22 September 2026

Add a source-preserving logical-unit projection across 513 frozen DMG/ADM documents: 49,680 units, exact UTF-8 source spans and complete byte accounting. Retain 49,634 machine boundaries as uncertain; 46 excerpts have explicit agent-reviewed boundaries.
Add five scoped profiles and 28 source-backed relationship proposals, including complete cross-page examples, definitions and conditional household/absence support. Preserve eight unresolved references and all earlier review obligations.
Add a separate indexed Reader and corpus v2 for reusable Explorer context assembly. Share PDF resources per document, preserve source dates and roles, expose boundary/concept facets, and validate actual posting bounds without dropping evidence.
Retain the 184-assembly fixed-source comparison and seven negative controls. All results remain insufficient; the five 32 KiB focused packages retain no source evidence. At 512 KiB the five new profiles retain 24 of 24 declared paths. Do not compare those paths directly with the old profile denominator or claim improved model answers.
Keep frozen page projections, remote-service defaults and replay observations unchanged. Add source, consumer, browser and publication checks, a beginner guide and separate outstanding semantic-review packages.

Abroad semantics and question diagnostics — 21 September 2026

Trace the reported five weak lexical matches to the historical 19 September client observation and distinguish it from a fresh offline replay of the published source and engine.
Reuse international concept identifiers, add overseas wording and declare whole-page Pension Credit qualification support. Preserve broad-benefit ambiguity, historical limits and open obligations.
Add focused source, paraphrase and budget checks; track cross-benefit coverage, task discrimination and historical source competition as separate unfinished work.
Point the learning path to generated current service status, date its 0.6.0 baseline explicitly and replace the obsolete unrun Data review wording.
Audit all 40 retained 512 KiB staff packages: 27 of 38 ADM-containing packages have no ADM relationship path, broad or duplicated triggers remain, and required-path losses differ from support-dependency gaps. Record exact denominators and case lists without upgrading any insufficient result or open obligation.
Add beginner explanations of Voice/client boundaries and record the returned frozen-file Data Analytics proposal review. Full artefact import and independent verification remain pending; native MCP access, Voice acceptance and improved legal-answer accuracy are not established.

Service 0.6.1 and native connector checks — 21 September 2026

Publish the question-schema compatibility patch as Sites version 11, preserving the five source versions, two engines and all earlier observations. The new SDK run reconstructed 11 cases in 121 requests with no retries or model calls.
Retain the native before-and-after controls: the same previously rejected multi-character question now succeeds. The exact Staff 012 small-budget check returns zero records and a byte-budget refusal; it is not substantive evidence or an AI answer.
Record the refreshed three-tool ChatGPT metadata with permissions unchanged. Keep client integration in progress: the existing Codex task still exposes only the full tool, while intended compact-client and nested Data Agent access remain unproved.
Verify that the shared status checker rejects the old 0.6.0 selection after new publication receipts arrive. Updating its exact immutable receipt selection produces the recorded 0.6.1 status; earlier failures and receipts remain unchanged.

ChatGPT connection and acceptance guide — 21 September 2026

Explain the difference between a recorded public service, an installed connection and tools available in a particular conversation.
Record the bounded metadata refresh observation and the remaining question-schema correction; do not claim a new ChatGPT, Data Agent or Voice acceptance.
Add fixed, manifest-first control and care-home prompts, exact-read identity checks and a client-observation work package. Keep failed calls, partial delivery and unverified hashes visible.

Shared recorded service status — 21 September 2026

Generate one service-status page from an explicit immutable deployment/SDK receipt selection, replacing cross-repository guesses about whether a release was deployed.
Reject stale selections when a newer successful compact-service publication or matching SDK observation is recorded. A deployment without matching verification remains pending; no browser, Voice, legal-accuracy or real-time availability claim is inferred.
Keep historical receipts unchanged and add bounded offline integrity, freshness and failure controls to CI.

Paired fixed-evidence responses — 21 September 2026

Retain four direct-v4 subscription attempts: two empty controls and both Staff 012 care-home answers. All pass mechanical checks with zero observed tools; the substantive answers contain six claims and seven exact citations.
Preserve identical complete evidence, authored prompt and schema for both clients. Both answers retain insufficient status and refuse to decide whether the whole Pension Credit award stops or continues.
Retain an independent agent critique: the main scoped claims are traceable, but one answer omits selected exceptions and overstates a gap. Preserve every original output and older failure; one paired question does not establish specialist acceptance, comparative accuracy, affordability or complete legal answerability.

Direct trial v4 preparation — 21 September 2026

Prepare a separately versioned client-format successor using documented installed CLI fields and retained value-free observations. Keep the original authored prompt, answer schema and complete public packages identical.
Add bounded structural diagnostics and strict scalar/structured usage validation. Thirty-nine controls and independent review pass; unknown events, tools and formatter calls still fail closed.
Preserve every v3 input and failed attempt. A new immutable freeze and successful controls are required before any substantive v4 calls; no answer result is implied by this preparation.

Recorded Monday delivery checkpoint — 21 September 2026

Make the eight retained direct-v3 observation tests portable across macOS and Linux by using the system temporary directory and resolving its path. Preserve all frozen executables, inputs and outcomes.
Export three actual public-service examples into 225 bounded static files, preserving complete package hashes, source and engine identity, insufficient status and historical limits.
Retain both rejected direct-v3 empty controls and hold substantive calls. Record client-format failures without claiming AI correctness or retrying the frozen attempts.
Add thirteen offline controls for the fixed-origin v2 website verifier, including late-response rejection, and update the beginner learning route and Monday handover.

Retained evidence publication — 21 September 2026

Add a narrowly approved static publication route for up to three fixed evidence examples, preserving exact packages, provenance, missing evidence and original authority labels.
Check committed inputs, complete reconstruction and bounded resource files; preserve the existing Markdown-only website when no registry is declared. Twenty-three offline controls and independent review pass.
Explain the browser reader, explicit publication approval and integrity limits in the beginner publication guide. A local build is separate from public-site verification and legal acceptance.

Public versioned evidence and direct trial freeze — 21 September 2026

Preserve the deployed 0.6.0 service record and separate public SDK observations: the first failed envelope comparison and the corrected 121-request, 11-case pass. Runtime and verifier identities remain distinct.
Freeze the actual received current care-home and empty-control packages for a separately governed direct-JSON paired trial. Thirty-two offline admission controls pass; no model outcome is implied by freezing its inputs.

Direct JSON model-trial harness — 21 September 2026

Add a separately reviewed fixed-evidence trial harness for the original Staff 012 care-home question and an independently assembled unknown-term control. Both subscription clients receive identical, complete governed evidence; every tool event is rejected.
Add 30 offline controls and strict source, engine, service, compact-reconstruction and hosting bindings. Limit execution to one attempt per provider and case, with successful empty controls required before substantive attempts.
Keep the protocol pending until the exact public service observations and final inputs are frozen. This entry records the harness, not new provider calls, answer success or specialist approval. Earlier trials remain unchanged.

Ignored-person and normal-residence qualifications — 21 September 2026

Add separate model-authored concepts for normally residing with someone and disregarding a person's presence for the Pension Credit severe-disability addition. Retain 17 whole captured pages, including eight newly selected pages; preserve dated conditions and distinct statutory definitions.
Declare conditional evidence dependencies without changing question triggers, existing evidence or any of the 203 open obligations. The compiled index has 913 records, 53 concepts, 1,526 assertions and 61 support dependencies within existing limits.
Record independent source review and 61 focused controls. Context retention, public delivery, model quality and specialist acceptance remain separate gates; no individual entitlement is established.

Later public observations — 21 September 2026

Preserve a separate learning-site observation: exact source 91b99078…, 123 successful responses and 2,382 checked internal links.
Preserve six actual public Chrome journeys for disability candidate df352daa…, with complete care-home and unresolved-SDA packages. Add offline source/hash/path admission and six controls; retain every original observation unchanged.
Keep public browser observations, offline comparison, protected publication, remote delivery and model trials separately labelled in the Monday handover.

Partner qualification support — 21 September 2026

Preserve the separate partner lower- and higher-rate branches, actual caring payment, complete treated-receipt provisions and dated memo changes. Correct the interpretation of a flattened superscript without changing source text.
Add ten supporting dependencies and conditional routes for Staff 012, 013, 014 and 017. Keep Staff 018's unbounded question unchanged. The semantic index now has 903 records, 1,482 assertions and 39 support dependencies; all 203 obligation identities and statuses remain open.
Rebuild the combined Reader and all 40 development cases. Independent source review, 52 semantic and 20 Reader controls pass. A separate 320-assembly comparison and exact replay retain 585/585 declared path occurrences at 512 KiB and 449/585 at 256 KiB. Preserve missing evidence, all earlier receipts and the distinction between source consistency and specialist acceptance.

Disability-addition qualification support — 21 September 2026

Distinguish the limited severe-disability overview from the detailed no-partner branch. Require their four and ten supporting pages, including two already captured Chapter 78 pages newly selected into the semantic index.
Preserve the difference between actual carer-benefit payment and the specified disability-benefit receipt qualifications. Keep partner-only patient scope and separate memo dates explicit; retain both unresolved meanings of SDA.
Add conditional support paths for Staff 012 and 013 without presuming that the claimant has no partner. Preserve all 203 obligation identifiers and statuses.
Independently review the source changes; pass 35 staff, 11 household and 20 combined Reader controls. Rebuild the 903-record, 1,464-assertion index and 21,224-relationship combined Reader without changing captured source bytes.
Retain a separate 320-assembly comparison and exact replay: all 497 declared path occurrences survive 512 KiB, while only 407 survive 256 KiB even though original candidate overlap remains 177/177. Preserve the visible missing qualifications, 64 KiB metadata refusal and every earlier observation. Publication and model-answer acceptance remain separate checks.
Preserve the separate public Chrome observation of qualification source 7f9feb96… and the published allocator. Verify its 273 immutable inputs, complete source passages and directed paths offline with 14 failure controls; distinguish selected evidence from references to explicitly omitted targets. This receipt does not attest the later disability source or remote service.
Fetch the receipt's immutable historical source before the complete unit-test suite in shallow CI checkouts. Preserve the initial failed run rather than treating a local checkout with complete history as proof of CI readiness.
Use the resolved platform temporary directory for the new receipt controls, preserving symlink checks while allowing the same tests on macOS and Linux.

Learning publication verified — 21 September 2026

Verify the published learning site against commit 7815b17bb3db3903738c745d3fb9508919ebab15: 122 actual HTTPS responses match the public manifest and all 121 generated outputs. Check 2,320 internal links in the received HTML, with no missing targets, missing fragments or duplicate IDs.
Retain the dated publication receipt, exact generated manifest, executed verifier and successful publication gates. This is a point-in-time publication-byte/link observation; it does not attest the later qualification source/engine pair, browser accessibility or AI answers.

Qualification retention verified locally — 21 September 2026

Compare two immutable source versions and two archived engines across all 40 cases at 256 KiB and 512 KiB: 320 assemblies and deterministic replay pass. The final source/engine pair retains 177 of 177 known candidate occurrences and 433 of 433 activated declared path occurrences. These are evidence retention measures, not answer-accuracy results.
Retain all seven household support pages for Staff 012 and 013. Record the smaller 27-record/63-relationship packages at 256 KiB, the 62/124 packages at 512 KiB and their remaining optional temporary-care-home dependency gap.
Rebuild current source-only evaluation: 12 to 177 candidate occurrences, ten generalisation controls and eight budget observations. Staff 012 at 64 KiB correctly refuses with metadata_budget, zero records and 1,926 bytes; the other seven observations retain evidence. All contexts stay insufficient.
Bind the joint receipt to DWP 7f9feb9634e3d94004853b838462aca132c505a5 and Explorer c4f2de0a99b7bc2f8b8c8a06a3c715fb56b66d8e. Keep all 203 obligations, earlier models and public observations unchanged. Protected publication, fresh public acceptance and larger component support sets remain separate open work.

Care-home component qualification dependencies — 21 September 2026

Narrow the housing-cost summary to the captured former-home treatment and the Housing Benefit “may be payable” condition. Keep the no-partner opening of the temporary-care-home rule explicit.
Add five required housing-cost pages and two required temporary-residence pages. Together with the initial household checkpoint, this produces 15 model-derived support relationships. Staff 012 and 013 require the household and housing-cost paths; temporary-residence support remains a record-level dependency rather than an unconditional permanent-care-home task requirement.
Regenerate the semantic index and combined Reader, retaining 901 semantic records and 1,442 assertions. Preserve all source bytes, historical observations and 203 open obligations. The no-partner and severe-disability overview dependency sets and publication remain separate work. The later joint comparison above records bounded retention without closing those larger sets.

Initial household qualification checkpoint — 21 September 2026

Declare seven captured guidance pages as required support for the household summary, and make the care-home overview require that summary. Compile eight source-backed, model-derived dcterms:requires relationships and explicit qualification paths for Staff 012 and 013.
Preserve source bytes, original candidate identifiers, all 203 open obligations and earlier model and browser observations. A support declaration does not establish complete legal applicability or specialist acceptance.
Carry the new edges through the combined Reader with the existing forward and inverse requirement labels, source provenance and review boundaries.

Household delivery follow-up — 20 September 2026

Independently verify the successor model critique against its frozen inputs, answer hashes and selected evidence. Recompute literal citation diagnostics, retain the defective locator and keep semantic opinions subject to human review.
Verify the full household Reader on the public website in Chrome: conceptual facets, statutory text and graph links, source/audit date separation, care-home headings and unresolved SDA branches. Retain the failed first harness assumption and all 16 execution artefacts, with bounded offline integrity controls.
Retain a separate nine-attempt model experiment: seven parser-accepted responses, six mechanical passes including both empty-evidence controls, one bad locator, a rejected formatter sequence and a timeout. Record 14 claims, 21 citations and model-authored scope concerns; no substantive paired success or specialist approval is claimed. Preserve all earlier experiments.
Prepare a current Voice rehearsal sheet with separate speech, tool-access and room-audio checks, plus a verified-evidence fallback. Actual Voice and room acceptance remain untested.
Publish service 0.5.0 from the merged runtime with four approved source versions. Seven actual full-context SDK cases and four compact reconstructions pass; all three public browser evidence journeys pass. Retain two Firefox hosting-cookie warnings and keep its strict console check failed.
Add an explicit observation layout so a new release can declare historical browser journeys not run. Verify the bound artefacts without inventing a pass or changing earlier receipts.
Extend the staff concepts to 51 and selected whole DMG/ADM pages to 96, keeping household headings, conditions, dated transitions and neutral Income Support.
Add 20 selected dated statutory units, 43 references and one verified metadata bridge; retain both acquisition attempts, source hashes and extraction limits.
Keep statutory evidence separate in the Reader, including source-family filters, official links and requested-version dates that do not become publication dates.
Retain 176 of 177 known candidate-page occurrences across 40 development tasks; all remain insufficient, with 203 explicit obligations and no specialist approval.
Preserve previous browser observations against archived exact source files; prepare separate current-candidate checks and paired model trials.
Retain a portable engine-only ambiguity/performance experiment and strengthen source-plane identity to cover the statutory acquisitions as well as PDFs.
Verify all 20 statutory extracts and links, directed graph routes, source/audit date separation, care-home qualifications and unresolved SDA branches in three local browsers. Preserve two failed harness attempts and 41 hashed artefacts.
Freeze six new model inputs and retain five actual subscription attempts with no accepted answers. Hold further calls after event-format mismatches and two timeouts; investigate separately without changing the frozen experiment.
Bound retained-observation file reads and reject symlinked manifests, artefacts or parent directories before opening them; preserve the original receipts.
Prepare a separately reviewed model experiment with strict observed CLI-event recognition and reversible context dictionaries. Six exact round trips pass; substantive inputs are 12–15% smaller, without an answer-quality or speed claim.

Learning website — 20 September 2026

Add a script-free web edition of the learning path, glossary, guides and public evaluation notes. Keep Markdown as the source and bind every page to its commit.
Publish only tracked allowlisted documents after protected-main validation; private correspondence and untracked research remain outside the site.
Add safe rendering, stable links, keyboard navigation, source fingerprints and deterministic build controls. Live publication checks are recorded separately.

Publication follow-up — 20 September 2026

Publish service 0.4.0 with the new staff corpus and both preserved earlier source versions; pass five actual SDK cases and three compact reconstructions.
Verify the combined Reader publicly in Chrome, Firefox and WebKit, with 261 distinct observed corpus files and five facet-parity checks. Preserve the first loading timeout and bounded rerun timings. Add a ten-minute Monday script.
Record the merged staff-semantic release and separate remaining domain work from human acceptance. Name the neutral Income Support modelling gap explicitly.
Add a portable public evidence-reader verifier with exact SDK/hosting bindings, catalogue and slice checks, fifteen offline controls, recorded request pacing and separate functional and console outcomes. Actual deployment observations are recorded separately; the verifier alone is not a public acceptance claim.
Retain three-engine staff functional passes and all twelve earlier functional regression journeys. Chrome and WebKit pass strict console checks; Firefox hosting-cookie warnings keep BL023 open. Preserve the failed Chrome driver measurement and reviewed browser-native observation amendment.
Refuse to overwrite retained public Reader observations and check the saved service artefact census, digests and outcome bindings offline in CI.

Additive staff semantics and both-manual evidence — 20 September 2026

Retain five paired Claude/Codex fixed-evidence cases, 12 attempts, 49 claims and 79 citations, including failed quotations, a timeout and a formatter-policy rejection. Keep model critique distinct from specialist acceptance and cost.
Map all 40 supplied question occurrences to projected personas and journeys, retaining exact wording, repeated questions, ambiguity and evidence needs.
Add neutral benefit and variant concepts, source-backed relationships and 40 executable task profiles. Preserve 203 explicit missing obligations and label the supplied questions as development cases. No complete answer or specialist acceptance is inferred from improved candidate retrieval.
Add official provision-identity metadata, exact citation mappings and six bounded tribunal searches. Keep statutory bodies, legal effects completeness, current applicability and judgment reuse outside the achieved evidence scope.
Add a combined DMG/ADM Reader with source-manual and conceptual facets, full captured-page search, directed relationships and separate source/audit dates. Preserve earlier release bytes and identifiers.
Compare a fresh official source-listing observation with all 513 captured attachment identities. Report fresh PDF/extraction hashes as unknown because this observation did not acquire the PDF bodies.
Retain independent-review corrections for citation inheritance and direct context delivery. Reproduce producer outputs and boundary controls in CI.
Explain the new outputs through the beginner learning path, semantic and legal guides, ontology map, methodology, retrospective and delivery ledger. Prepare a CPAG permission-scope draft without sending it or changing reuse rights.

Team handover and explicit delivery tracking — 20 September 2026

Add a shareable team handover with the merged baseline, verified demonstration links, supplied-question evidence and a dependency-ordered continuation plan.
Name broader semantic modelling explicitly under BL005 (neutral concepts and relationships) and BL007 (task evidence profiles). Their implementation is in progress; existing mention classifications do not complete these tasks.
Split each backlog item into delivery and acceptance work packages. Record the executor, next action and evidence separately; generate the detailed Markdown ledger from the register and check it in CI.
Add controls rejecting completed work without evidence, aggregate completion with open packages, and model review presented as independent human acceptance. The focused backlog suite passes 15 tests. Preserve all earlier source, context, model and browser receipts; this checkpoint makes no new semantic-release claim.

This dated entry records the evidence-review and navigation work. DWP PR 10 records its integration and exact check/merge status. Explorer PR 125 is merged and its published DMG navigation has a passing scoped Chrome observation. Public service 0.3.1 has separate hosting, SDK and browser observations below, including its unresolved host-console failures. Earlier dated observations keep their original scope.

Added

A discovery-first departmental method, HMRC discovery brief, actual namespace map and factual retrospective, with the difference between adopted vocabulary, proposed standards and domain concepts made explicit.
Stable backlog IDs, dependencies and acceptance checks, plus a separate multi-agent work log with ownership and handover boundaries.
Separate the completed classification/DMG Reader milestone from unfinished ADM Reader and cross-manual navigation (DWP-BL-024). Ask OKF already includes both manuals; this remaining gap concerns the human Reader projection.
An additive review descriptor with 44 navigation labels and an explicit classified/unclassified accounting of all 19,090 DMG/ADM pages. Literal facet assignment does not assert legal applicability; the Reader keeps its DMG scope.
Forty human-readable staff review packs, with 42 shared source-evidence resources and machine-readable packs of 3,795–12,062 bytes. Source candidates remain distinct from retrieved pages; every case awaits specialist review.
Offline CI checks for backlog identities, dependencies and prose consistency, and retained answer-trial input/output integrity. No model calls run in CI.
Three actual, fixed-public-evidence Claude subscription trials, with tools disabled and no model override. Preserve the abroad trial's strict verbatim quotation failures, the custody citation checks and the no-evidence abstention. Human claim review remains pending; these are not engineering or legal passes.
A separate actual Claude local MCP observation: seven compact-tool calls, exact replay of diagnostics and two source records, and offline corruption controls. Preserve the preceding zero-call attempt and fabricated model catalogue as a failure; do not conflate transport with answer quality.

Changed

Deploy service 0.3.1 after independent final review found a stale replay link: clear it when the question or source changes and on resubmission. Twelve local browser journeys passed across three engines, and the live SDK retained exact package parity. Preserve all 0.3.0/v6 receipts and hosting failures. The corrected public reader's twelve historical-profile journeys passed their functional assertions, including the changed replay identity, then failed strict console checks on host errors. A separate v7 full-corpus Chrome journey verified exact displayed evidence and also retained the host-console failure.
Correct the reusable Explorer documentation cache to include transitive linked Markdown and its exact-source alternates. Forty focused tests and the assembled site check passed; no evidence or application-runtime change was needed.
Make this changelog directly visible from the main and beginner guides.
Public service 0.3.0 exposes the unchanged ask_okf plus an evidence catalogue, exact bounded reads and a human replay route. The official SDK verifies three unchanged full packages, all five read sections, exact reconstruction of the 31,312-byte abroad package and three fail-closed controls. Hosting and SDK receipts are separate from browser, AI-answer and public Explorer acceptance.
Preserve the public service reader's failed strict console gate: all nine historical-profile browser cases reached their final console check with functional assertions satisfied, then failed on host-injected Cloudflare code blocked by CSP (plus Firefox cookie-domain errors). Record DWP-BL-023 for hosting integration; keep CSP and fail-closed controls unchanged.
Record a separate public full-corpus Chrome journey on hosting version 6: four real tool calls, six insufficient-context records, and rendered source, provenance and diagnostic hashes matching the SDK. Functional checks passed; the strict host-console gate still failed. Preserve this observation before the 0.3.1 replay-link correction.
Verify published Explorer navigation after PR 125 and Pages deployment: all 21 app files and 188 observed corpus files matched the expected bytes, and benefit/circumstance/topic/authored-concept filters matched across Reader, Graph and Timeline (74/343/268/1 records). The scoped Chrome check found no console errors or targeted accessibility violations. The Reader remains DMG-only; all full-corpus questions remain insufficient pending specialist evidence profiles. Preserve the earlier local observation separately.

Staff questions, ADM acquisition and repository governance — 19 September 2026

Repair inherited repository semantic and publication contract drift against unchanged canonical schemas. Preserve per-delivery scope, status, authoring notes and output descriptions in separate okf.delivery.json; declare concrete outputs using governed roles and add the root OKF version declaration. Canonical reconciliation passes without warnings. Add an offline CI gate and regression controls for unsupported fields, unknown roles, unsafe/missing output paths, altered schema bytes, broken references and dependency cycles. Preserve the original publication classifications and source denominators separately; canonical fields do not upgrade extraction or interpretation authority. Keep real-browser journeys required and add a clearly labelled offline check of retained public-browser identity and context evidence.
Add a public register of 40 question occurrences and 39 distinct wordings, preserving ambiguity and recording 42 verified source candidates without publishing private correspondence, contacts or collaboration links.
Run all 40 questions through the remote MCP service against the preserved custody profile. Retain exact HTTP response bodies, source/core hashes and offline replay. All 40 packages match the shared engine and remain insufficient; no substantive or specialist-approved answers, token savings or cost savings are claimed. Eight mutation controls reject altered provenance, transport accounting, parity claims and forged sufficiency.
Acquire all 182 PDFs and 4,347 pages in a separately frozen ADM publication census. Preserve original PDFs, page text, acquisition receipts and date distinctions without changing DMG evidence. Verify every file and page locator; retain 91 pages without extracted text and other extraction-quality flags.
Record combined acquisition coverage of 513 PDFs, 19,090 measured pages, 18,197 pages with nonempty extracted text and 893 without. A live remote full-corpus run covers 40 staff questions and three boundary controls: all 43 complete packages match the shared engine, and all 40 questions return candidate evidence, 12 retain a separately located page and 21 retain a page from a candidate document. All 43 packages remain insufficient; no answer-quality, AI-answer or specialist-acceptance claim follows. Retain published-browser acceptance as separate evidence.
Add eight ADM acquisition controls for frozen census identity, classification, bounded metadata, source paths, PDF integrity, page locators and unreviewed extraction authority. Existing DMG acquisition helpers remain unchanged.
Audit authored relationships through semantic and runtime projections, endpoint identities and bidirectional adjacency. Document sparse semantic coverage separately from the Explorer display defects; data preservation does not establish complete policy modelling or certify browser behaviour.
Enable and independently verify classic main protection: reviewed pull request workflow, strict GitHub Actions validate requirement, administrator enforcement, resolved conversations and no force pushes or deletion. Record zero required formal approvals explicitly; configuration is not perpetual assurance or evidence that substantive independent review occurred.
Add a private-input filename guard and seven temporary-repository tests. Tracked or staged .email.md files are rejected without reading their contents.
Add a task-led beginner learning path and linked glossary, with official references, benefit-variant explanations and distinct Search, Ask and AI roles. Explain provenance, date meanings, full capture versus evidence profiles, remote MCP, browser WebMCP, HTTP 405 and unverified Voice access.
Lead the meeting guide with the combined-corpus candidate, an explicit version and a bounded ChatGPT rehearsal prompt. Record the completed published-browser journeys and retain the earlier custody observations under their original version. Document separate consumer pins for replaying the two generations.
Verify deployment 5 through the official MCP SDK: current imprisonment and hospital packages and the explicitly selected historical imprisonment package exactly match the shared engine. Preserve the earlier failed hosting attempt; distinguish these transport checks and the wider remote evaluation from the separate published-browser observations.
Independently replay all 43 complete corpus packages, reverify 42 source candidates and reject all 21 corruption controls. Update only the delivery verification status; retain source snapshots and evidence authority unchanged.
Verify the published Explorer at commit a8628fdb77c1c03a5d99b6d105d9e4b8722088d7, including downloaded app-file identities, Search, Ask, provenance, directed chapter routing and machine-readable context. Native WebMCP build and explain calls in the Codex browser match the UI and remote bounded abroad package: six source pages, 31,312 bytes, context 2cdfa5fe…, insufficient and truncated. Default imprisonment exposes eight resolved concepts, 64 records, 127 relationships and chapter 24/53/54/78 routing, while remaining insufficient. ChatGPT Voice and room audio are not verified by these checks.
Record ChatGPT's cached older tool schema rejecting the new corpus version before any MCP request. Document the observed existing-connection Refresh route and keep the post-refresh invocation separate from service acceptance.
Retain the actual ChatGPT retrieval failure and its corrected rerun: the same bounded abroad question now returns six source pages and 31,312 bytes, without reported host truncation. Keep insufficiency and both retrieval/assembly limits visible. Bind the earlier and corrected 43-case runs in a comparison: exact-page overlap increases from 10 to 12 cases and document overlap from 19 to 21, with two gains and no lost overlaps. These are retrieval diagnostics, not answer quality, complete legal advice, specialist acceptance or Voice verification.

Remote Ask OKF acceptance — 19 September 2026

Add an independent HTTP MCP acceptance client for the imprisonment and hospital questions, comparing complete returned packages with the existing Explorer engine and evidence assessor.
Retain compressed raw tool results, input and output hashes, source version, protocol and observation times; replay those observations in CI without a live service dependency.
Document the hospital knowledge gap separately from transport success. Existing custody evidence does not answer hospital questions; no new benefit rule or specialist acceptance is inferred.
Add ChatGPT connection and meeting demonstration instructions, with separate gates for deployment, actual AI invocation and Voice or WebMCP host support.
Record actual ChatGPT Pro-account calls, full-response host limitations and successful smaller packages. Keep client-side delivery limits separate from the assembler's explicit budget truncation; preserve the failed attempts and model interpretation limitations.
Replay six smaller-budget HTTPS captures against the unchanged core in CI, including whole source records and explicit insufficiency.

Governed context assembly public experimental candidate — 16 September 2026

Add an imprisonment case spanning legacy JSA, Income Support, State Pension Credit and the two ESA components, with exact whole-page evidence and explicit regime boundaries.
Project governed context records and directed relationships from authored YAML-LD. Preserve the distinction between source chapter routing and model-derived paragraph selection; do not change the frozen Bundle Wiki profile.
Add a separate assessor case, frozen-source preflight and real-engine A–H acceptance. Verify exact identities, directed paths, provenance, budgets and scoped answerability without generating a model answer.
Bind source URLs and capture timestamps to frozen inventory/API/census evidence, including catalogue event-date roles. Reject corruption that preserves text and digests but substitutes source identity.
Replay context acceptance and actual-index controls in CI against an explicitly pinned Explorer commit with locked dependencies; reject provisional branch names or unresolved pins.
Execute declared synthetic missing-evidence, direction, integrity, access, ambiguity, conflict and budget controls on copies of the actual index. Bind input, implementation and observed output hashes; preserve real execution timestamps on replay.
Add a demonstration and reproduction guide. Keep these engineering checks separate from current-law assurance, specialist acceptance, browser/host verification and the unchanged 160 source-guided answer trials.
Publish immutable demonstration links after checking Explorer 905e680f6d3ad385de9b8effc351566eba0ab2b3 with DWP content efb05c66616a9cd4328a86cf412780fe7bc7cf0b in the public browser. Record the 52-record, 127-relationship package, exact engine equality, source and graph navigation, JSON inspection and bounded failure; retain screenshots and app-byte verification.
Add the Search-to-Ask demonstration sequence and an explicit evidence-only handover prompt for a separate AI answerer. The observed Chrome host has no native WebMCP tools; no AI answer or voice integration is claimed.

Full-DMG research candidate — 15 September 2026

Record the owner's unattended completion authority, finite acceptance gates and an hourly task continuation with durable checkpoints.
Complete acquisition against all 331 DMG URLs and a separate indexed Explorer projection, preserving the original 36-PDF Pension Credit inventory, stable routes, frozen domain profile and immutable demonstration links.
Add a finite semantic/evaluation worklist and make CLI retrieval scope explicit. Keep CPAG body reuse, specialist acceptance and full ADM outside this research-bundle completion claim.
Verify source hashes and all 14,743 measured pages. Add bounded evidence discovery, an indexed full-text compiler and source/semantic coverage reporting. Retain 802 pages with no extracted text and the absence of exhaustive visual review or OCR.
Add eight wider authoring batches: 245 concepts and 328 exact-passage relationship proposals across all 78 substantive source units, with explicit exceptions and pending specialist review.
Record totals including the pilot: 271 concepts, 343 semantic proposals, 15,390 entities, 16,210 assertions and 15,363 resource records. Keep those counts distinct from complete policy modelling or specialist acceptance.
Complete bounded research accounting for the other 253 source-family units: 242 outcomes with documented gaps and 11 spare units marked not applicable with evidence; do not claim exhaustive source-body review.
Correct the indexed exploratory notice envelope, add route labels and direct original-PDF narrative links, and use platform-independent gzip headers for reproducible macOS/Linux builds.
Execute indexed locator, source identity, no-result and baseline navigation controls separately from model answer trials.
Record 160 context-aware, source-guided observed answers and separate model assessments: 90 supported, 56 partial and 14 with underspecified rubrics. Preserve original responses and omissions; these are assessor categories, not an accuracy score, blind benchmark or end-to-end retrieval evaluation. Specialist acceptance remains zero.
Add scripts/verify_full_dmg_trials.py --check to replay evidence integrity and reproduce the retained trial summary without model calls or automatic grading.
Run unmodified, pinned Explorer acceptance functions in CI; retain its exact warning and normalise label whitespace without changing source titles.
Correct 278 PDF source links from authored records that were labelled as HTML; keep authorship, source-file format and media type independently accurate. Add an isolated compiler regression and negative metadata controls.
Add publisher/resource endpoint labels and source hosts, reconcile per-publisher resource counts, and preserve CPAG's typed April 2026 publication month separately from capture and observation dates. CPAG remains metadata and links only.
Reconcile literal paragraph and memo references to acquired location candidates, preserving ambiguous matches and unresolved legal identifiers across all 331 source documents.
Supply all 420 Graph metadata/facet labels, with exact consumer route encoding and deletion controls.
Pass 43 local repository tests. Verify public JSON/YAML-LD loading, source navigation, the selected ESA Graph and DMG 42320 relationship, six labelled PDF resources for 84351, CPAG's April 2026 Timeline and bounded search controls at content commit 80b6f08426aea39dd2934fb8795b61215e2cc0ad, snapshot dwp-full-dmg-2026-09-15-2dd78242297c. Retain exact scope in validation/full-dmg-browser.json.
Add immutable full-DMG launch links and a ten-minute walkthrough. Record Explorer's observed search-term expansion separately from exact Python retrieval controls. Preserve original Pension Credit links and earlier receipt scope.
Track publication and canonical CI history in PR 5 separately from the immutable content's browser receipt. Specialist acceptance and production assurance remain incomplete.

CPAG external reference — 15 September 2026

Added a searchable CPAG Welfare Benefits Handbook reference with publisher links and explicit subscription, rights and AI-processing boundaries.
Connected the reference to the welfare rights adviser persona and Pension Credit topic without inferring substantive policy agreement.
Recorded the public access review. No CPAG handbook text was acquired; the DWP source inventory and page counts remain unchanged.
Regenerated YAML-LD, JSON-LD, RDF, Explorer and Markdown projections with mixed-rights attribution.
Verified the CPAG record, search, rights notice and Graph in the live Explorer using both import formats.

0.1.0 — 15 September 2026

Created an independent experimental Pension Credit OKF+ exemplar from all 36 PDFs linked by the declared GOV.UK publication.
Captured 1,524 pages with hashes and page-preserving extraction; included the seven substantive chapters and 744 page records in default Explorer content.
Added source-linked concepts, personas, stories, questions and evidence-bearing semantic relationships.
Added a researched, schema-validated discovery handoff, pinned consumer/profile files, deterministic build, validation and retrieval controls.
Documented unofficial status, OGL attribution, extraction limits, historical context, missing external evidence and deferred production assurance.
Logged owner-supplied tribunal, calculator and CASA links outside the demonstration snapshot.
Logged future extensive evaluation, benefits-engine and customer-journey ideas without expanding the first exemplar.
Corrected Markdown display of source numbering after live-browser inspection; extracted source text remains unchanged.
Preserved upstream Explorer licence notices alongside the unchanged profile mirror.
Recorded the delivered publication scope as the public repository and verified Explorer demonstration.

15 September 2026 — semantic stage two

Preserve the original 22 concept identities while moving authoring to individual YAML-LD files; add four source-backed capital concepts.
Add 15 explicitly model-derived semantic proposals, exact passage and paragraph/page evidence, a generated semantic map and a non-executable capital-disregard review candidate.
Add five projected personas, six user stories and six questions; de-identify staff needs and distinguish them from owner requests and model suggestions.
Assess pinned UK legislation catalogue reuse and add three metadata-only references without importing statutory text or assuming amendment coverage.
Capture a reproducible official metadata census of 331 DMG and 182 separate ADM PDFs; plan a 24 September content freeze and 30 September seminar.
Record WebMCP/website/Mac audio feasibility and a model-comparison design; neither integration nor comparative model performance is claimed as tested.
Extend deterministic generation, source-passage/locator validation, RDF triple checks, identity-preservation and negative controls. Keep previous source snapshots, domain profile and browser receipts intact.
Clarify the seminar host as ChatGPT live voice; record verified desktop Voice documentation and usage boundaries while retaining the untested integration status.
Correct CPAG temporal metadata: publisher-described April 2026 edition release and 6 April online announcement remain separate from September project capture; preserve month precision.
Verify the corrected Timeline in both public bundle formats against the deployed Explorer, and retain the exact deployment receipts. Refresh current launch links, distinguish source-date coverage from labelled record dates, and clarify extraction gaps and the remaining legislation/website work.