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

Demonstrate source-led evidence discovery

This is an independent research candidate. Use it to inspect how guidance is prepared and selected, not to decide a person's benefits. Its source-led projection is separate from the earlier public service. Publication and client acceptance must name their exact source and engine versions.

1. Understand the collection

Open the beginner manual guide, then its linked machine-readable conventions and document catalogue. DMG means Decision makers' guide; ADM means Advice for decision making. These are different DWP staff manuals with different structures and benefit regimes.

The captured collection contains 513 PDFs and 19,090 pages. A PDF page is a checking location. A source unit is a proposed passage, which can cross pages. The candidate has 53,727 retrievable units, including 75 earlier authored units. An exact hash proves which bytes were used; it does not certify the interpretation.

2. Inspect a unit before asking a question

Load structured-context/okf-explorer.json in a compatible Explorer. Select a passage, inspect its heading and source role, and follow the original PDF link. For a passage crossing pages, inspect every span. Distinguish the original publication date from the date the project captured or processed it.

A discovery card is a short source preview linked to the complete retained passage. It helps discovery. It is not a legal summary or evidence of a complete rule. Uncertain boundaries stay labelled. A source reference means that one passage mentions another; a legal dependency needs separate review.

3. Compare Search and Ask OKF

Search for a subject such as abroad. Search finds candidate records.

Then open Ask OKF and paste a complete question from the staff-question matrix. Keep its wording, including missing facts. The current Explorer starts with Package bytes 524288 (512 KiB). Check this under Evidence limits if you changed a limit earlier, then use Build evidence package. This is 512 times 1,024 bytes. The recorded 32 KiB assemblies are smaller-budget controls and retain no source evidence; selecting that budget is a useful refusal control, not a shorter benefits answer.

Inspect these separately:

Use Read whole source passage, the record link and View cited source to move from selection to checking. An uncertain boundary can supply review material while the result remains insufficient. Its presence is not a pass for answerability.

4. Open the same package an AI would receive

Use Inspect package JSON or Copy evidence package. JSON is a structured text format. The package records the question, selected evidence, reasons, paths, source identity, budgets and unresolved requirements. No model has to answer the question for this package to be useful.

When a compatible remote release is admitted, its manifest and exact reads can deliver the package in small parts. Delivery size is the size of one reply; assembly size is the total evidence selected. Small replies cannot recover evidence excluded during assembly. Verify the source version, engine version, complete-package hash and every part before claiming a complete reconstruction. The current page tools default to 16 KiB per delivery response while keeping the separate 512 KiB assembly default. A historical remote release may have different defaults; check its recorded version instead of assuming parity. See the client-check guide. A configured connection is not proof that a particular ChatGPT, Data Agent or Voice session has its tools.

An AI may interpret the package, but its answer needs a separate claim-level check: which exact evidence supports each claim, and which qualifications were preserved? This experiment does not establish improved model-answer accuracy.

5. Read the evaluation without combining different measures

The structural experiment starts with four cases and then repeats with eight. It is development acceptance, not a blind accuracy estimate for every paragraph. A separate census accounts for every captured document and extracted byte.

The question trials retain all 40 staff occurrences plus an unknown-term control, in two projections and at two budgets. In trial 06, source-read profiles activate for 37/40 occurrences, up from 8 in the earlier logical-unit baseline. At 512 KiB, all 40 retain source passages and relationships. All 47 inherited, 181 previously added and 208 newly added source-selection path occurrences are retained. These are routes through evidence, not counts of correct answers.

Three questions still lack a suitably scoped source-read profile: unspecified benefits abroad, an ambiguous use of SDA, and an unnamed benefit in a care home. Their existing research leads stay available; the system does not guess the missing facts. Location navigation activates for all 40, but does not close their 203 original open obligations. All results remain insufficient. The 32 KiB limit still retains no source evidence. Failed allocation attempts and location-route trade-offs are kept alongside later repairs. See the separate measures and remaining work.

The original imprisonment and hospital acceptance cases are a separate check; they must not disappear merely because the staff-question census passes. Recorded specialist acceptance, current legal applicability and accurate AI answers each need their own evidence.

Reproduce the local checks

Use the locked Python environment and the immutable Explorer engine named in evaluation/manual-structure/context-probe/engine.json:

uv sync --locked
uv run --locked python scripts/build_manual_guide.py --check
uv run --locked python scripts/build_structured_units.py --check
uv run --locked python scripts/build_structured_context.py --check
node --experimental-strip-types scripts/check_structured_replay.mjs \
  --explorer-root /absolute/path/to/the/pinned/okf-explorer

The final command requires an explicitly accepted current.json pointer. Its absence means adoption is still held; do not create a pointer merely to bypass that gate. For an individual earlier trial, use its retained runner, engine and source identities. See the work log and backlog for the exact remaining work.

Frozen Demo 1 comparison

Use the forty-question ledger and four-answer review for the 30 September presentation. The current care-home source was checked through both the public UI and native in-app WebMCP. The source text and context identity matched. A separate remote MCP, Voice or Data Agent session still needs its own observed connection test. The four authorised answer calls are complete; rehearse with the retained answers rather than spending more model allowance.