Logical evidence units

Learning path · Backlog · Earlier abroad review

Implemented and verified in the public Reader, 22 September 2026. This is an independent research publication, not an official DWP service or an entitlement decision.

Why a page is not enough

A PDF page tells us where text appears. It can stop in the middle of an example, condition or citation. It can also contain several unrelated rules. The source pages remain the immutable record of what was captured; they are not sufficient boundaries for retrieving meaning.

The intended logical evidence unit is a complete source passage with its governing heading, list, notes, citations and attached examples. It can cross several pages. Automatically detected candidates remain explicitly uncertain until their boundaries have been reviewed. A separate dependency names other evidence needed to interpret it, such as a definition, exception or update memo. An example needs its parent rule; finding its concluding sentence does not establish a general rule.

flowchart TD
    Source[Frozen PDFs and extracted pages] --> Blocks[Headings and logical content blocks]
    Blocks --> Units[Scoped evidence units with exact source spans]
    Concepts[Concepts and task profiles] --> Units
    Units --> Dependencies[Required definitions and qualifications]
    Units --> Context[Bounded context assembly]
    Dependencies --> Context
    Context --> Review[Human inspection and separately evaluated AI answers]
    Units -. page locations and hashes .-> Source

Current increment

The semantic closure and compact-delivery report records the next bounded source review: 75 authored boundaries, seven profiles and a reproducible ledger covering all 40 staff question occurrences. The implementation and comparison below describe the earlier 46-unit baseline; the current report gives the new results and separates them from public checks.

Source integrity and uncertainty

Each unit records ordered source spans, explicit UTF-8 byte offsets, source and fragment hashes, and any inserted joining characters. A hash identifies exact bytes. The producer checks the original extraction; the consumer checks the supplied unit and its internal span consistency. Consumer checks do not re-acquire the PDF or prove that a machine extraction reproduced its layout.

Detected boundaries remain labelled as machine candidates. Explicit authored boundaries have source-backed reasons and an agent-review status; neither is specialist acceptance. Contents, reserved paragraph ranges, empty extraction, ambiguous headings and unassigned material remain accounted for. Repeated paragraph numbers need document, version and occurrence context.

Segmentation and semantic modelling are separate. A smaller complete passage does not automatically acquire a benefit concept, a task route or current legal applicability. Old page-based requirements and frozen evaluations retain their original identities. New unit-based requirements need their own reviewed scope.

Earlier implemented baseline

The new additive projection covers both frozen manuals: DMG, the Decision makers’ guide, and ADM, Advice for decision making. These manuals concern different benefits and variants; being in the same corpus does not make their rules interchangeable.

A task profile declares which concepts activate a question scope and which source paths it needs. It is authored knowledge, not an instruction that the retriever may use as a hidden answer key. The new profiles cover Pension Credit absence, Universal Credit temporary absence, household absence, temporary care-home residence and temporary/permanent care-home alternatives. Earlier page-based profiles remain in their original projection; they have not all been migrated.

The Reader contains 50,524 records: the units, 331 concepts and 513 source documents. It shares one PDF resource per document, retains exact page links on each unit, and offers boundary, unit-kind, manual and authored-concept facets. The 523 graph relationships include existing concept relationships and the new source routes. A facet saying No authored reference exposes a semantic gap; it is not evidence that the source contains no relevant meaning.

The page-break example

The authored unit for DMG 077001 joins the end of PDF page 8 to its continuation on page 9. It retains both complete Jason and Nicola examples, the continuing entitlement conditions, duration branches, citations and the reference to DMG 04642. Other units retain the death-related and medical branches and the qualified-practitioner definition, including its citation on the following page. The unresolved supersession reference stays visible. Household and qualifying-young-person passages have separate, conditional scope.

ADM source fixtures retain the Universal Credit absence passages, their cross-page qualifications and an unresolved memo reference. Boundary review by an agent is not specialist legal acceptance, and the memo’s applicability remains unresolved.

Earlier recorded comparison

Preserved baseline results compare 40 staff questions, five focused cases and one unknown-term control in both modes at 32 KiB and 512 KiB: 184 assemblies, plus seven scope controls. KiB means 1,024 bytes. The run used the same frozen sources and questions, no network and no model calls. Its exact evaluator and engine hashes are retained alongside 20 focused context archives.

At 512 KiB, the new focused profiles retained these declared paths:

These 24 paths are not the old page-profile denominator, and do not close 203 previously recorded review obligations. All 184 results remain insufficient, with ai_answer: null. At 32 KiB all five focused unit packages retain zero source evidence: metadata and required context cannot fit. The small-budget refusal is a limitation, not a successful benefits answer.

Local median assembly time was 176.64 ms for units and 126.17 ms for pages at 512 KiB across all 46 cases. This includes confined local reads and cache reuse, not remote delivery or client latency. Unit packages retained more source text in these focused cases, including uncertain lexical candidates. This is not a performance improvement, relevance score or model-answer accuracy result.

Demonstration and reproduction

The additive Reader descriptor is logical-context/okf-explorer.json; its Ask index is logical-context/manifest.json. Open the pinned logical-unit Reader with the compatible Explorer build from PR 140. The source pin identifies unchanged logical data from this implementation; the public browser and WebMCP observation verifies this pinned source with Explorer commit 69d38b1c. UI, JSON and actual WebMCP build/explain calls agree on the context ID; no public console warnings or errors were recorded. Exact merged CI and documentation deployment status remain in the DWP PR handover. The existing combined Reader and remote service keep their separately recorded versions. Do not describe this build as a new default deployment of ask-okf.crpage.chatgpt.site.

1. Open the logical-unit descriptor in Explorer.
2. Search 077001. Compare the contents candidate with the authored gateway unit; open the whole passage and its exact source spans.
3. Select Ask OKF and ask: What happens to Pension Credit if I go abroad?
4. Under Evidence limits, use 524288 package bytes. Build the package.
5. Show resolved concepts, the gateway and complete examples, directed references and requires relationships, provenance and missing obligations.
6. Inspect or copy the package JSON. A compatible WebMCP host can call okf_build_context and okf_explain_context and compare the same context ID.
7. Repeat at 32768 bytes to demonstrate an explicit insufficient-context refusal. Do not fill the gap with a model’s own knowledge.

uv sync --locked
uv run --locked python scripts/build_logical_units.py --check
uv run --locked python scripts/build_logical_context.py --check
uv run --locked python -m unittest discover -s scripts -p 'test_logical*.py'
node --experimental-strip-types scripts/evaluate_logical_context.mjs --check --explorer-root /path/to/approved/okf-explorer

The approved immutable engine and four module hashes are in engine.json. Check out that commit before the final command. The check verifies those Git blobs before importing code, recomputes bounded packages and compares the retained results. It does not call an AI provider or overwrite earlier observations.

What remains open

• Review the remaining 49,551 machine boundaries and difficult layout, table and heading cases. Full byte accounting proves preservation, not correct meaning.
• Migrate the remaining staff profiles with explicit concept and qualification routes. The current increment supplies seven scoped profiles, not full semantic coverage of the 40 staff questions or every DMG/ADM benefit.
• Resolve the twelve recorded references, current legal versions, territorial and claimant applicability, and obtain specialist review.
• Reduce unrelated lexical candidates and metadata overhead; prove useful compact delivery before adopting a new remote-service source version.
• Run new paired fixed-evidence model trials only after evidence scope review. No claim of improved answers or affordability follows from this build.

These are tracked separately from the completed implementation packages in the backlog. Earlier failed integrations and preliminary comparisons remain in the work log.

Acceptance gates

1. Reconstruct exact unit fragments from the frozen source; retain every original span or explicitly classify why it is not a logical evidence candidate.
2. Preserve complete conditions, examples, footnotes, tables and heading scope.
3. Keep source references distinct from required support and applicability.
4. Load required unit destinations within explicit resource bounds and name any missing or ambiguous dependency.
5. Compare identical questions, source and budgets; report required support, irrelevant text, metadata cost, truncation and unresolved concepts separately.
6. Preserve old page routes, immutable sources, service versions and replay receipts. Admit a new service version only after its own checks pass.
7. Keep the beginner guide, machine backlog, changelog and publication status in step with the actual implementation and observed results.

Learning-programme integration

The combined descriptor is a template input to the logical-context Reader. Adding learning metadata therefore changes the dependent build identity even when all source-unit text, boundaries and semantic relationships are unchanged. The logical Reader explicitly excludes the combined corpus's teaching overlay: its lesson routes do not exist in the separate unit corpus.

The integration rebuild replays the same 184 contexts and seven negative controls, with no network or model calls. Every result still reports insufficient evidence. The earlier evaluation is retained unchanged under evaluation history. Historical public observations continue to identify their original builds; they do not certify this new snapshot. The learning programme does not resolve corpus-wide review, compact evidence delivery or remote-service adoption.