OKF-DWP

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

View this version’s Markdown source

On this page

Repository status

Current overview: 26 September 2026

Start with remaining work and next priorities, the Evidence workbench and the unmerged-work review. The full declared DMG/ADM source capture, source-led discovery, bounded delivery and 40-question workbench are implemented. DWP PR 39 also delivered the optional interaction and blocked calculation-inspection models; the sidebar demonstration records its tested developer-connection route.

The passage-boundary review increment adds source-bound before/after evidence for the 28 known historical-amendment cases and an isolated successor parser. It preserves the original corpus and records structural outcomes separately from retrieval and legal answerability. Its implementation is merged in DWP PR 44 and Explorer PR 155. The live workbench passed all 28 case previews, PDF rendering and review export; its 34 application files match the reviewed build. The release record keeps CI, publication and source acceptance separate. Fifteen target boundaries are corrected, twelve are partial and one remains unresolved; the 40-question replay establishes no improved expected-page coverage.

The LiteParse source pilot is complete as a bounded experiment: 12 windows were compared with Poppler, the source-quality gate failed, and the 24-window expansion was not admitted. Coordinates and numeric table rows are promising, but full header associations and exact text checks need further work. This result changes neither corpus defaults nor answer scores.

All 40 retained packages remain insufficient. The known capital transfer-wording and UC selection failures, unreviewed unit boundaries, legal dependencies, specialist acceptance and room rehearsal remain open. Broader semantic modelling is incomplete implementation as well as incomplete review. A source count, successful page or delivered inspection model does not establish answer quality.

The sections below are historical checkpoints with their own versions and denominators. In particular, the eight-profile result from 22 September must not be read as the coverage of every later source-selection increment.

Earlier semantic and delivery increment: 22 September 2026

The source-review and compact-delivery report records 75 authored unit boundaries and 49,551 uncertain candidates. Seven profiles activate for eight of 40 supplied question occurrences, including two broader DLA/PIP matches with explicit task ambiguity; 32 do not activate a unit profile. The 184-context replay remains insufficient throughout. All 203 earlier obligations, frozen sources and combined learning routes are preserved.

Read the coverage ledger for exact question, source, profile and lesson mappings. Local compact checks reconstruct the five larger packages through bounded responses; all five small inline packages still contain no source evidence. The work log keeps implementation, CI, merge and public observation separate. No new remote service default, specialist acceptance or improved AI-answer score is claimed.

Earlier additive source review: 21 September 2026

The abroad audit distinguishes a historical lexical failure from current semantic and capacity gaps. The additive source contains 57 authored concepts, 112 selected pages and 74 support dependencies. A retained comparison preserves 24 packages and 10 negative controls: four Pension Credit questions retain all seven qualification pages at 512 KiB, while 32 KiB results still fail to retain useful evidence. All 203 obligations remain open. The source increment does not change the separately versioned public-service default; use the generated service publication status for that baseline.

Earlier Monday delivery checkpoint: 21 September 2026

The current handover separates the public 0.6.0 service, exact source versions, retained evidence reader and paired model trial. DWP PR 20 and its exact merged validation and Pages checks passed. The ignored-person source contains 53 concepts, 106 selected pages and 61 support dependencies; the service remains pinned to the preceding partner source because the richer source exposes a measured context-size limit.

The v4 results and agent critique retain mechanically accepted responses with specific qualification issues. All 203 task obligations remain open. PR 21 and PR 22 carry the later publication and trial work through normal protected checks. This checkpoint does not claim their eventual merged CI or a later public observation.

Earlier additive staff increment: 20 September 2026

DWP PR 12 and Explorer PR 126 are merged. The additive content is pinned to 9de52acf1db84b27f8933d80480eaa850e74fa33. Canonical DWP validation passed. Public combined Reader checks passed in Chrome, Firefox and WebKit; their exact scope includes five facets, source links, graph paths and targeted accessibility checks. Explorer PR 127 is also merged and its Pages build passed. Service 0.4.0 is deployed; its SDK observation verifies five cases and compact reconstruction across all three approved source versions. The public staff reader passes functional checks in all three engines; Firefox hosting cookie warnings keep the overall strict hosting gate open.

Use the team handover and current work log for the latest delivery status. The sections below preserve dated earlier release observations; their original source counts, consumer versions and limitations are not current whole-project claims. Both DMG and ADM have since been captured: 513 PDFs and 19,090 pages.

The new staff semantic increment separates neutral benefit meanings from task-specific rules, records missing evidence obligations, and supplies an additive combined Reader. The legal references identify observed provisions; they do not establish statutory body evidence or current applicability. The source refresh compares fresh official listings, while explicitly leaving fresh PDF and extraction hashes unknown.

Full-DMG research candidate

On 15 September 2026 the owner instructed completion of the bundle unattended. This authorises the public-source processing, semantic drafting, verification and normal reviewed PR publication in the completion plan. The checkpoint records actual task state. The final research candidate contains 331 PDFs and 14,743 measured pages. Eight semantic batches add 245 concepts and 328 proposals across 78 of 78 substantive PDF units. Including the pilot, it has 271 concepts, 343 proposals, 15,390 entities, 16,210 assertions and 15,363 resource records. These are selected-passage research outputs, not a complete entitlement model.

Research attempts for the other 253 source-family units have terminal outcomes: 242 completed with documented gaps and 11 spare units marked not applicable with evidence. The source-family review preserves those limits; it does not claim exhaustive body review. Across the 331 PDFs, 802 pages have no extracted text. No exhaustive visual review or OCR has established their contents.

All 160 context-aware, source-guided answer trials have observed responses and separate model assessments: 90 supported, 56 partial and 14 rubric-underspecified. These are assessor categories rather than a calibrated accuracy measure. The summary binds the unchanged answers, source reads and per-case findings. This is not a blind or end-to-end retrieval benchmark. Specialist acceptance remains zero.

The full-DMG candidate is publicly available at browser-checked content commit 80b6f08426aea39dd2934fb8795b61215e2cc0ad, snapshot dwp-full-dmg-2026-09-15-2dd78242297c. Open the JSON candidate or follow the walkthrough. Publication and canonical CI history are tracked in PR 5; the immutable browser receipt applies to this content commit. Earlier browser checks retain their named immutable scope. The Pension Credit publication remains available with its original 36-PDF inventory, 744 default page records and frozen domain profile.

The full-DMG browser receipt records public Edge observations against that content: JSON and YAML-LD loading; an ESA Graph with 11 nodes and 10 edges and no missing labels; the exact DMG 42320 relationship evidence and hashes; six labelled PDF resource results for 84351; the CPAG Timeline's April 2026 publication month; and bounded source, search and lazy-record checks. YAML-LD Narrative retained paragraphs 84351–84353 and the direct chapter 84 PDF-page-35 link. The 073499 search returned six results and loaded the split chapter's six-digit passage; a term in the final record chunk also loaded. No warning or error logs were captured in the checked browser session. The local repository suite passed 43 tests. These results do not establish every route, comprehensive accessibility or policy correctness.

One search limitation is explicit: unavailableclaimantdetails expanded to the indexed term unavailable and returned seven results. The nonsense control zzzxqvnomatch produced an unmatched term and no results. Browser query interpretation and the exact Python retrieval controls are different evidence paths.

The generic Concept facet also passed its public Graph check, with an explicit bound of 200 loaded records out of 271 exact index matches. Targeted resource loading can initially show zero in Evidence Sources while Narrative retains the direct original PDF link. The exact content commit's CI run 35028883819 and CI run 35028878408 passed. These candidate results remain distinct from subsequent canonical branch checks.

CPAG substantive reuse and specialist acceptance are separate external decisions. Neither is required to complete an explicitly unreviewed research bundle from public GOV.UK sources. No handbook body acquisition or specialist acceptance is authorised or asserted by the unattended instruction.

Lifecycle: experimental public preview. The owner explicitly authorised a public chris-page-gov/okf-dwp repository, an OKF+ YAML-LD exemplar and a demonstration for a pensions representative on 15 September 2026. The owner later supplied three links with an instruction to log them without changing the 4pm workstream.

The initial repository scaffold was committed separately before feature work. Source acquisition, discovery, build and independent review ran in parallel; publication proceeds through a feature pull request. Source and generated artefacts remain distinct.

Evidence

Full-DMG candidate:

Verify the retained trial evidence with uv run --locked python scripts/verify_full_dmg_trials.py --check. This replays integrity checks and reproduces recorded assessment categories; it does not run or grade new answers.

Original Pension Credit pilot and its later small-bundle additions:

Foundry boundary

The pre-existing Explorer Foundry warm-up and build prompts informed this work; their exact pins are in the domain profile. This is a bounded preview, not a claim that the complete Foundry production process has run. In particular, expert legal/terminology review, exhaustive external legislation and memo acquisition, full accessibility and cross-browser assurance, full positive/negative locked-consumer fixtures before acquisition, backwards compatibility assurance and production release-gate closure are deferred. The time-bounded user request authorised progressing a reviewable exemplar; no deferred gate is marked passed.

The draft domain profile's blocking_for_build entries describe blockers for a later governed production build. They do not revoke the owner's explicit authority for this documented preview. The limited build and browser receipts establish only their stated checks.

Editorial check

Original project material uses British English and sentence case. Official titles, exact source extracts, schema/API fields, code identifiers, URLs and unmodified vendored profile bytes retain upstream wording intentionally. Source text is not silently edited for spelling, contemporary rates or apparent inconsistencies.

The initial meeting browser-verified snapshot is dwp-pension-credit-2026-09-15-bdee692f1126: 791 nodes and 850 evidence-bearing assertions. Direct JSON and YAML-LD imports both passed in the live Explorer. All 744 source pages also passed literal-text rendering checks after a numbering defect was found and corrected.

The earlier pilot publication consists of the public GitHub repository, immutable commit URLs and its verified Explorer demonstration. No separate GitHub prerelease or uploaded release assets have been created.

CPAG external reference addition

The owner subsequently requested a CPAG handbook assessment and bundle addition. The updated snapshot dwp-pension-credit-2026-09-15-e49131bd7ef0 contains 792 records and 852 navigation/containment assertions, including one metadata-only external reference. The 744 DWP source-page records are data-equivalent to the initial meeting snapshot. The original domain-profile handoff remains frozen; the CPAG access review is separate under docs/cpag-handbook.md.

Bundle observation timestamps follow the latest recorded source or authoring observation; they do not establish a referenced source's publication date. YAML whitespace is normalised only when decoding proves that the data is unchanged. Existing source-rendering and initial browser receipts remain scoped to their named snapshot; the CPAG browser check is recorded separately.

The CPAG reference passed live Explorer checks in Edge using both YAML-LD and JSON. Search, access wording, publisher links, Graph and the exact snapshot notice were checked; neither tab captured warning or error logs. See validation/cpag-browser.json for the immutable content commit and hashes.

Semantic stage two

Started following the owner's request to move beyond the meeting preview and prepare a wider exemplar. The initial stage-two candidate added individual concepts, source-backed semantic proposals, a non-executable rule review packet, legislation catalogue references and expanded evaluation journeys. The new full-DMG/ADM census initially contained publication metadata only. The later full-DMG candidate described above acquires all 331 DMG PDFs; the separate 182-PDF ADM census remains metadata-only.

The original domain profile, initial behavioural designs and earlier browser receipts remain frozen. The full-DMG stage adds the separate 160 observed source-guided trials and model assessments described above; the 14 baseline journey checks establish route preservation, not observed answers to every original question. Specialist approval remains zero. The CPAG public contents harvest remains a separate local research artefact; this change does not publish its 90-entry outline or incorporate handbook text. The semantic pilot does not depend on those local files.

Content freeze is 24 September 2026; the seminar is 30 September. WebMCP and MacBook/PA integration are at feasibility stage, with ChatGPT live voice identified and its actual Mac/task/browser/audio integration awaiting rehearsal. No new website, audio configuration or voice-tool integration is delivered by this candidate.

The initial stage-two candidate snapshot was dwp-pension-credit-2026-09-15-42c0a30e120c: 821 records and 964 assertions (744 page-containment, 205 navigation and 15 model-derived semantic proposals). All 744 source-page runtime records are unchanged from the previous public baseline. Local deterministic generation, all-assertion/RDF validation, 12 semantic negative controls, eight retrieval controls and the frozen wider metadata census pass. Independent agent review covered semantic wording and validation failure modes; this is not human domain review. See validation/semantics.json and validation/semantic-review.json.

The exact stage-two content commit 48abaff6c640e4a3e206af1adc71e21efd12738f passed live Explorer import checks in installed Edge for YAML-LD and JSON, semantic relationship inspection, graph labels, the non-executable candidate and its original PDF-page link. No warning or error logs were captured in either test tab. Browser receipt.

CPAG source date correction

The updated candidate dwp-pension-credit-2026-09-15-1b84574268ad preserves the same 821 records, 964 assertions and source pages. CPAG's public metadata establishes April 2026 as the handbook release month. The referenced book now records that month separately from the original 15 September metadata capture, record generation and later publication-metadata review. The source statement does not establish an exact release day or a legal effective date. Earlier browser receipts retain their original snapshot scope.

In this 821-record snapshot, CPAG is the only record with an explicit structured publication date. The Timeline correction uses declared date roles; it does not infer source dates for the other records. Their observation or generation dates remain labelled as such. Wider DMG document publication and revision dates still need evidence-backed mapping, separately from legal applicability.

The corrected content commit 7d3c69df0571b8d5206c8edce92963d99979cb5c passed live Timeline checks in both YAML-LD and JSON: Source published — April 2026, machine-readable month precision and three separately labelled September audit dates. The current capital-disregard graph, source link and exact snapshot also passed. No warnings or errors were captured in these tabs. Current browser receipt.

The earlier pilot's companion Explorer correction was merged in PR 120. Its 51601b5d94ac33ce5be654e81943ca2e049743d8 deployment passed all required application, cross-browser and exact deployed-identity checks. The preserved deployment receipt records the Explorer host check; it does not establish DWP policy correctness or bind the later full-DMG candidate. The latter's reviewed consumer source pin is recorded in profiles/explorer-runtime/vendor-lock.json; its deployed-host identity and checked public interactions are recorded separately in the full-DMG browser receipt.