Staff-question semantics and evidence requirements

This stage addresses DWP-BL-005 (neutral domain concepts) and DWP-BL-007 (task-specific evidence requirements). It implements the reviewable modelling work for the supplied questions. Broader benefit/variant coverage, task discrimination and bounded qualification retention remain unfinished implementation. Independent specialist approval and complete legal applicability are separate outstanding acceptance gates.

Start with the practical difference

Previously, the phrase Pension Credit could resolve to a prisoner credit rates concept. JSA and ESA had similar custody-specific aliases. That worked for the first imprisonment demonstration but was a poor starting point for questions about savings, partners, care homes or benefit interactions.

The new, additive projection resolves these names to neutral benefit concepts. It distinguishes variants, components, circumstances, assessment inputs, entitlement and payment. The custody concepts and their evidence remain in the graph, with their specific labels. The old published projections are unchanged.

A concept is an explicitly named meaning, such as Pension Credit or capital. An alias is wording that the resolver may recognise for that meaning. An assertion is a directed relationship with its own status and provenance. A profile declares the evidence and unresolved obligations for a task. These are inspectable project proposals; they are not official DWP interpretations.

Required support for a summary

A reference helps a reader find related material. A required dependency says that an interpretation needs particular material to be inspected with it. The initial household checkpoint uses the existing dcterms:requires relationship for that purpose. Seven captured pages support the complete household summary; the care-home overview also requires that summary. The component follow-up adds five housing-cost and two temporary-residence page dependencies: 15 support relationships in total. Staff 012 and 013 declare the household and housing-cost paths explicitly. Temporary-residence support stays attached to its own summary, without becoming mandatory for every permanent-care-home question. These declarations remain project-authored and unreviewed.

The disability-addition increment adds 14 dependencies: four pages for the limited overview and ten for the no-partner explanation. These groups preserve receipt exceptions, partner-only patient scope and dated memo qualifications. Staff 012 and 013 investigate the branch conditionally; its inclusion does not establish the claimant's household status. That increment has 29 support dependencies, with two captured pages newly selected into the semantic index.

The partner follow-up adds ten more dependencies and conditional paths for Staff 012, 013, 014 and 017. That increment had 39 dependencies and four qualification profiles. A source-backed navigation route invites investigation; it does not infer continuing couple status or satisfaction of any disability or caring condition.

The joint comparison verifies that the generic allocator retains the declared paths within its budget. It runs DWP 7f9feb9634e3d94004853b838462aca132c505a5 with Explorer c4f2de0a99b7bc2f8b8c8a06a3c715fb56b66d8e, separately from the earlier public service. The final pair retains all 433 activated path occurrences at 256 KiB and 512 KiB, including all seven household pages for Staff 012 and 013. This is bounded evidence retention, not complete domain modelling. All 203 obligations remain open. The qualification source has since passed protected publication and a separate public Chrome check. Later source increments need their own publication observations.

What is implemented

The current source candidate is dwp-staff-semantics-613fa1c1c1cc682e9853, with index SHA-256 92a8871b8f1f2e51f1feace0b1f57c67dfd0ddcb0434cf4574795e5942e95fd6. Its current 40-case evaluation retains 177 of 177 known candidate occurrences at 512 KiB. All contexts remain insufficient. The abroad increment reuses four international concept identifiers and retains all seven supporting Pension Credit pages for four specific questions at 512 KiB. Its separate comparison also preserves the loss of source evidence for the general question at 32 KiB. The preceding ignored-person comparison retains its own measured path losses and exact source identity.

The partner source remains the separately frozen service/trial input; consult recorded service status for deployment observations. The earlier four-cell comparison above binds its separate 901-record source and does not attest these later increments. The initial household-dependency checkpoint had 1,435 assertions. The earlier public service source 3ef0e786… had 1,427 assertions and a 4,581,721-byte index; its immutable linked evaluation remains an observation of those earlier bytes.

The concepts are authored in domain-profile/staff-semantic. They are kept outside the legacy knowledge-folder scan because adding inputs there would change the identities of the frozen pilot and its dependent releases. This follows the separate additive navigation pattern.

The catalogue records each definition, type, source-page association, alias reassignment and projection refinement. Source text remains the exact whole-page extraction. Adjacent-page links are explicit: the page next door is available in the full corpus but is not silently treated as selected or reviewed.

Relationships carry qualifications

Examples include:

• State Pension is distinguished from its new and earlier Retirement Pension regimes, with DMG 74001 as the source route.
• Guarantee Credit and Savings Credit retain distinct assessment dependencies.
• Housing Benefit's treatment as income in Pension Credit is separate from Housing Benefit eligibility and housing-cost additional amounts.
• Carer’s Allowance and the Pension Credit carer additional amount are linked through their stated conditions; payment and underlying entitlement are separate questions.
• PIP daily living is linked to the named qualifying-benefit conditions, without asserting automatic entitlement to another benefit.
• JSA and ESA variants are explicitly represented. The broad benefit label does not choose a variant.
• PIP upper-age and DLA-transition research has direct source routes into ADM P1, P4 and P5. This corrects the earlier reliance on a DMG AA/DLA page as a candidate for those questions.

These associations use established SKOS relationships and Dublin Core references. SKOS is a vocabulary for describing concepts; Dublin Core provides general metadata relationships. A broader-concept relationship is a navigation hierarchy, not an executable legal rule or an OWL class axiom.

flowchart TD
    Staff["39 distinct staff questions<br/>40 recorded occurrences"]
    Concepts["51 model-authored concepts<br/>benefit, variant, component, circumstance"]
    Profiles["40 proposed task profiles"]
    Relations["232 source-grounded associations<br/>including 15 explicit support dependencies"]
    DMG["Frozen DMG source pages<br/>exact text, hashes and PDF locators"]
    ADM["Frozen ADM source pages<br/>including PIP age and transition guidance"]
    Legal["44 verified provision references<br/>metadata only"]
    Bodies["20 selected dated statutory units<br/>derived and unreviewed"]
    Open["203 named open obligations<br/>scope, closure, applicability, law version, review"]
    Package["Bounded inspectable context<br/>remains insufficient"]
    Staff --> Concepts
    Staff --> Profiles
    Concepts --> Relations
    Relations --> DMG
    Relations --> ADM
    DMG -->|"62 citation links"| Legal
    Legal -->|"explicit source-backed references"| Bodies
    Bodies --> Package
    Profiles --> Open
    Relations --> Package
    Profiles --> Package
    Open --> Package

Why the result still says insufficient

The profile register separates:

Each obligation has a stable identifier, a readable label and a resolution action. The identifier is intentionally absent from source evidence. Retrieving a page, a legal hyperlink or a second model opinion cannot fabricate its completion. The requirements checker consequently reports the missing obligation rather than declaring a complete answer.

SDA remains deliberately ambiguous. It may mean Severe Disablement Allowance or informal shorthand for a Pension Credit severe-disability additional amount. The resolver shows both candidates instead of choosing silently. “War Pensions Constant Allowance” also retains its unresolved scheme-name boundary.

The earlier legal reference nodes remain scope records labelled reference-only-unreviewed. The new statutory-body increment adds separate derived evidence records for 20 selected units, connected through 43 source-backed references. A metadata record does not become source text, and a source passage does not establish applicability. All 203 obligations remain open. The separate legal reconciliation retains the earlier identity and footnote work.

Current source-only comparison

The current evaluation compares the original discovery base with the final semantic source using the same Explorer c4f2de0a… engine. This differs from the four-cell joint comparison, which also changes the engine independently. Both preserve their input hashes.

Both runs use the same shared Explorer engine, budgets and frozen lexical shards. Only the explicitly authored semantic base differs.

This is a development-case comparison: the supplied questions and candidate pages were known during modelling. It is not a held-out accuracy result. A page hit does not prove the page supports a claim or supplies every exception. The earlier released staff model retained 169 of 177 candidates. The separate ambiguity/loading experiment retained 171, and the immutable household-source evaluation retained 176. Those historical counts do not refer to the newly regenerated local evaluation. The current candidate retains all 177 occurrences, including the formerly omitted Staff 005 page, while all 43 evidence-closure obligations remain open. These are distinct development comparisons, not accuracy scores.

Thirty-eight of the 40 current after-runs contain ADM pages; all report truncation. The earlier projection had ADM pages in all 40. This change reflects bounded selection, not removal of ADM from the corpus. ADM presence is not a relevance score. The lexical stage finds broad matches in both manuals, and traversal may supply additional source paths. The receipt separates all ADM pages from those reached by a semantic path and reports actual truncation codes. Candidate, node, relationship and byte limits remain visible.

Full comparison and controls · Exact before/after context archives

Ten shared-engine controls cover neutral resolution, explicit SDA ambiguity, no results, missing source evidence, reversed relationships, three paraphrases, the original imprisonment routing question and a counterfactual entitlement premise. Thirty-one producer controls check source hashes, exact spans, all question occurrences, qualification dependencies, legal-reference boundaries, preserved scopes and rejected governance/authoring mutations.

Context sizes and practical limits

A byte budget limits the complete returned package: passages, relationships, question interpretation, missing-evidence diagnostics and provenance all count. The shared engine now includes requirement diagnostics before trimming records. Previously, it could trim until the records fitted, append the diagnostics and then discard everything as an oversized result. The generic fix retains whole passages that fit and still refuses when metadata alone exceeds the budget. It does not shorten evidence text or suppress missing obligations.

The reproducible budget_observations in the evaluation receipt cover four staff questions also used in earlier model trials. These are new engine/source observations, not replacements for their frozen trial inputs:

All eight packages remain insufficient and disclose truncation. At 64 KiB, Staff 012 returns a 1,926-byte metadata_budget refusal: its interpretation and evidence obligations cannot fit, so no records, relationships or requirements are returned. The other seven observations retain non-empty evidence. A zero-record refusal does not mean that the source has no requirements; the independent joint harness counts activated authored requirements before output trimming.

At 256 KiB, Staff 012 retains its seven household support pages. The separate 512 KiB joint observation retains 62 records and 124 relationships, but also reports a missing temporary-care-home → page/78/0024 dependency. That optional summary is not an unconditional requirement of the permanent-care-home profile. The no-partner and severe-disability overview dependency expansions remain open. These sizes are observations for these questions and this snapshot, not a universal minimum. One question can activate several task profiles when their concept conditions overlap: these four activate 4, 3, 2 and 1 respectively. Profile activation is a conservative evidence obligation, not a relevance score.

Reproduce

Use the locked environment and an Explorer checkout containing the corpus context engine:

uv sync --locked
# Use a separate Explorer checkout at the recorded consumer version.
git clone https://github.com/chris-page-gov/okf-explorer.git ../okf-explorer-staff
export EXPLORER_STAFF_ROOT="$(cd ../okf-explorer-staff && pwd)"
git -C "$EXPLORER_STAFF_ROOT" checkout c4f2de0a99b7bc2f8b8c8a06a3c715fb56b66d8e
uv run --locked python scripts/build_staff_semantic.py --check
uv run --locked python scripts/test_staff_semantic.py
node --experimental-strip-types scripts/evaluate_staff_semantic.mjs \
  --explorer-root "$EXPLORER_STAFF_ROOT" --check

Without the check flag, the two build/evaluation commands write their additive outputs. The evaluation makes no network or model calls. Its local fake-origin fetcher serves only hash-bound repository files; it is not a public deployment observation. The receipt records the exact engine file hashes.

The current evaluation is rebuilt when source-metadata or authored-input bytes change, even if their meanings and selected evidence stay the same. The paired model trials retain their original archived assembly index and exact context packages. Their answers were not rerun after the legal metadata publication projection and whitespace-only YAML clean-up. Replaying those trials checks the archived inputs; it does not claim they used the newer current snapshot.

The index is compact JSON. The first staff projection shortened one repeated generic discovery scope string in the new projection only. The household projection preserves this declared normalisation and whole source pages; it requires the explicitly tested 8 MiB semantic-index bound. Ordinary search and record shards remain limited to 4 MiB each; returned packages keep their existing limits. The catalogue retains the exact before/after mapping. Specialised custody scopes, source text, source hashes, provenance and the original frozen index remain unchanged.

Remaining modelling work

The bounded staff-question model is now implemented and executable. A complete departmental knowledge model is not claimed. Remaining work includes:

1. Specialist review of meanings, alias choices and task requirements.
2. Closing each profile's source-continuation, exception and scope gaps.
3. Applicable legislation, effects, commencement, territorial and case-law review.
4. Additional task families and genuinely held-out question evaluation.
5. More precise relation types or rule models where reviewed evidence justifies them; current SKOS associations do not implement a benefits engine.
6. Further retrieval selection and evidence-size work without concealing truncation or weakening the completeness boundary.

These acceptance gates should stay separately visible in the backlog rather than being collapsed into either “done” or “broader modelling incomplete”.