A ten-minute Pension Credit demonstration

Purpose: show what can be achieved by turning complex public guidance into an inspectable, source-linked knowledge bundle. This is an independent experimental exemplar, not an official DWP publication.

Open the prepared demonstration.

1. Start with a question — two minutes

Open the bundle in OKF Explorer, then choose Where does the guide explain capital disregards? (question/pc001). Follow its links to the capital disregard concept, chapter 84 and PDF page 35 (page/84/0035).

Explain the three layers: project-authored question, source navigation and machine-extracted official-source text. Only the original publisher controls the official guidance.

2. Inspect the evidence — two minutes

Open the original PDF at page 35. The extract includes paragraph numbers, footnotes, legislation citations and a reference to DMG Memo 11/20. Show the original file SHA-256 and page locator in the record's provenance. The memo is not silently reconstructed or assumed current.

3. Show the relationships — two minutes

Switch to Graph or Links. Follow page → chapter (dcterms:isPartOf) and question → topic/page (dcterms:references). Inspect a relationship's direction, source and target identities, derivation, evidence and rights.

These relationships provide traceable navigation. They are not executable eligibility rules. A legal dependency requires separate expert modelling.

4. Show why context matters — two minutes

Open the part-week payments question (question/pc006) and read PDF pages 8–9 together. A broad search for part week can instead rank later, scenario-specific worked examples first. Search rank does not establish which rule applies. Keep apparent differences visible for specialist review.

Then open the assessed income period question (question/pc003). Show the dated context on chapter 83 page 3 and its known extraction-quality limitation.

5. Finish with a concrete next step — two minutes

Show question/pc008: the bundle cannot determine an individual's entitlement. Ask the representative to select a small set of real professional research questions, check the evidence routes and identify important missing context.

Possible next increments are expert-reviewed paragraph segmentation, terminology and cross-references, separately governed legislation and tribunal decisions, update comparisons, and a tested adviser-facing interface. The later-source log is deliberately outside today's snapshot.

Local fallback

If the public Explorer cannot load, use the generated Markdown at bundle/index.md and the original PDFs in source/pdf/.

For a deterministic keyword retrieval demonstration:

uv run --locked python scripts/query.py '84351' --limit 5
uv run --locked python scripts/query.py '79104' --limit 5
uv run --locked python scripts/query.py 'unfindablepensionkeywordxyz' --limit 1

The last command should return no evidence. The CLI retrieves extracts with citations; it does not generate legal answers. Source-native examples remain in their original context and are not treated as claimant case data.