Full DMG semantic coverage workplan

Prepared on 15 September 2026. Authored worklist, not execution results. Content freeze: 24 September; seminar: 30 September.

The goal is a source-complete research bundle and a precise report of attempted semantic work. It is not a promise of a complete ontology, legally current rule system or specialist-approved benefit decisions. The machine worklist binds the frozen census and current authored inputs by SHA-256.

Fixed denominators

Chapter 7 is seven separate source units. Its parts must not overwrite one another under a single chapter key. The 78 substantive-file classification is based on publisher metadata; it is a worklist input, not a conclusion that every passage is current or usable. Existing concept descriptions remain source-discovery material unless separately developed and reviewed.

Source-family work queues

Every one of the 331 PDFs has an explicit source-unit ID, official URL, acquisition task, research task and profile in the JSON. No source class silently disappears from coverage.

| Profile | PDFs | Required research outcome | |---|---:| | Listed substantive chapters | 78 | Paragraph/evidence coverage, term candidates, relationship proposals or reasoned abstention, dependencies and two chapter cases | | Annexes | 4 | Establish substantive role, parent scope, tables/rates/date risks and source locators | | Transitional chapters | 3 | Explicit transition cohorts, dates and unresolved applicability | | Memos | 66 | Exact referenced chapters/provisions and applicability/incorporation state, or unresolved state | | Amendment records | 151 | Change events and affected locators; never flatten into current chapter text | | Change summaries | 15 | Administrative change evidence, distinguished from legal effective dates | | Abbreviations and legislation reference lists | 3 | Candidate expansions and resolvable work/provision citations; no implied legal acquisition | | Spare chapters | 11 | Inspect/account for source and any discrepancy; no invented concepts |

Order and depth

All PDFs enter acquisition and source accounting. P0 prioritises common decision-making, evidence, appeals, international questions and the Pension Credit/pension boundary. P1 adds cross-benefit and operational comparisons. P2 completes the remaining specialist-benefit research queue. Priority changes order, not the denominator.

For each substantive source unit, complete RQ01–RQ08 and RQ09 where the JSON requires it:

• RQ01: Acquire/resolve exact PDF bytes and SHA-256; measure pages; retain original URL, source family, publisher dates and acquisition time. Record failed sources without fabricated text.
• RQ02: Segment contents, body paragraphs, examples, tables and footnotes with page/paragraph locators. Preserve raw extraction and mark uncertain boundaries; headings alone remain navigation.
• RQ03: Review candidate terms against body evidence. For each retain preferred label, aliases if evidenced, benefit/date/jurisdiction scope, definition or definition gap, evidence and unreviewed status. Resolve reuse of existing terms explicitly.
• RQ04: Review possible relationships using exact source passages. Retain direction, predicate meaning, rationale, conditions/exceptions, source role and evidence. A supported proposal or reasoned abstention closes the research task; no relationship quota forces an assertion.
• RQ05: Record source publication, revision, acquisition and stated effective dates separately, including precision. Inspect transitional text, old examples, memo changes and incorporation claims; unknown applicability remains unknown.
• RQ06: Extract explicit legislation, case-law, memo and cross-guide citations with exact citation strings and locators. Resolve work/provision/version separately; mark unacquired, ambiguous, broken or conflict states. Do not claim legal interpretation from a URL.
• RQ07: Author and execute one chapter source-grounding case and one boundary case, linked to existing persona/story/question routes. Separate deterministic checks, model outputs and specialist judgement.
• RQ08: Publish a source-unit coverage row with task state, counts, attempted areas, unresolved issues, evidence links and the reviewer type. Reconcile counts against the pinned 331-unit denominator.
• RQ09: Compare shared terminology only through scoped mapping candidates. Record DMG/ADM, legacy/New Style, Pension Credit/UC and territorial distinctions; do not infer equivalence or transfer a rule.

A research task can conclude that no supported relation was found. There is no numerical quota for new concepts or edges. A heading becomes a navigation entry; a concept definition needs body evidence, benefit/date/jurisdiction scope and an explicit review state. Use exact source identifiers and retain uncertain segmentation.

The existing four predicate definitions remain the current contract. General thematic association may use the documented SKOS relation. Do not apply the three capital-specific experimental predicates to other benefits merely because a heading looks similar. A new predicate requires an explicit definition, source evidence, inverse/direction decision and separate contract review.

Complete substantive chapter worklist

The focus column contains research prompts inferred from titles, not validated terminology or approved definitions. Each row has two finite evaluation tasks and exact source-unit references in the JSON.

Bounded ADM context

Start with the following seven ADM files, totalling 370 publisher-declared pages, for the specific cross-guide questions. They remain a separate acquisition and research family; no present acquisition or substantive comparison is claimed by this plan.

The JSON names six conditional files: A4, A5, C1, M5, R2 and U1. Activate one only when an existing journey has a recorded unresolved dependency that needs it. Initial plus conditional selection is bounded at 13 of 182 ADM PDFs; the other 169 are not queued. Expanding further needs a separate scope decision. The claimant-facing UC eligibility source supports service discovery and scope questions; it does not replace staff guidance or legislation.

Evaluation and staff needs

Reuse all eight personas, ten stories and fourteen questions listed in the JSON; do not replace the original Pension Credit questions when widening coverage. The 156 chapter cases provide broad coverage. Fourteen explicitly enumerated baseline cases preserve the original questions, and 22 family cases cover source navigation and a boundary check for each publication group. The six existing stage-two journeys remain the staff-facing end-to-end tests.

For every substantive unit, the positive case asks for a source-backed statement, exact locator and scope; its expected evidence must be authored after acquiring and checking the source. The negative case asks whether a heading or isolated passage authorises a rule for another benefit/date. Executed answer text, input/output digests, method, errors and reviewer state must be retained. A prompt and expected-behaviour description are not a passing result.

Thirteen reusable negative controls are defined in the JSON: missing support; heading-only promotion; benefit/regime confusion; wrong locator; unsupported supersession; date-role/precision substitution; reversed direction; example-as-rule; unavailable CPAG text; unacquired legal content; unsupported specialist status; individual entitlement demand; and source-byte replacement. Apply relevant controls by profile and the directed-relation control to the capital pilot.

Temporal and legal dependency work

Every explicit dependency receives a ledger entry with the literal citation, source page/paragraph, possible target identity, acquisition state and resolution confidence. Distinguish work, provision, version and digital file. Retain unresolved abbreviations, case identifiers, memo references and conflicting interpretations. The sibling legislation catalogue helps identity resolution; its metadata is not captured provision text or a Pension Credit effects graph.

Do not infer that an amendment was incorporated, a memo expired or a rule commenced merely from a document date. Publication, revision, observation, acquisition and stated legal effect are different fields, with precision preserved. A legal dependency that cannot be acquired can close an accounting task with a gap, but it cannot support a legal proposition or specialist acceptance.

Unattended execution and honest completion

Resume work using (source unit, source hash, requirement, tooling revision). Keep completed attempts immutable and link retries to prior attempts. Queue independent source units in bounded batches; report acquisition, extraction and semantic errors separately. Every task produces a receipt with method/model identity, exact inputs, outcome, evidence paths and unresolved dependencies. Never manufacture a model identity or usage value when unavailable.

Pass/fail/not-run/blocked counts must remain separate. A documented blocker satisfies accounting, not a passing evaluation. Report how many source units have a definition, a supported proposal, a rejected proposal and an unresolved legal/temporal dependency; none of those counts estimates all possible concepts in welfare law.

The report columns are fixed in the JSON, including source identity, priority, acquisition, page/text quality, segmentation, candidate/proposal counts, unresolved dependencies, evaluation outcomes, research state, specialist state and receipts. Render those columns as the human coverage report. This gives a finite stopping rule while leaving expert-review gaps visible.

Practical release checks

1. Reproduce the frozen census with python3 source/discovery-2026-09-15/acquire_metadata.py --check; reconcile acquisition output to the worklist by exact official URL and source hash.
2. Run the repository’s governed build, scripts/validate_bundle.py, scripts/evaluate_queries.py and scripts/evaluate_semantics.py through uv run --locked python after the full-corpus pipeline is integrated. These current scripts alone do not yet prove every new coverage requirement.
3. Add report checks that join every source-unit ID and requirement to a receipt, reject unknown IDs and duplicate terminal outcomes, and reconcile all source/class/family/chapter/persona totals. Keep the plan unchanged; execution results are separate artefacts.
4. Validate every exact quotation and original-page locator, proposed endpoint/predicate, authority/review state and hash; include the applicable negative mutations.
5. Measure loading, search, graph navigation and memory on the actual expanded corpus. Use bounded projections or source-family bundles if the existing consumer limits are exceeded; do not silently omit records.
6. Check a representative route from every source family and semantic priority, plus all seven Pension Credit substantive chapters, in the installed consumer against the frozen bytes.
7. Freeze content on 24 September with all completion labels computed independently. Website/WebMCP/voice remain separate implementation and rehearsal gates for 30 September.

Decisions still requiring people

The original stakeholder plan, named specialist reviewers, accepted synthetic outcomes and any wider ADM scope remain open. Those decisions do not prevent source acquisition, provenance validation, conservative research proposals or a coverage report. Unattended work must finish at a candid research result when specialist decisions remain unavailable, not invent acceptance or a complete ontology.