Examine the 40 questions in the Evidence workbench
This workbench makes the evidence behind a question visible. It is an independent research tool, not an official Department for Work and Pensions (DWP) service or an entitlement decision.
The Decision makers' guide (DMG) and Advice for decision making (ADM) are staff guidance manuals. Their numbered paragraphs identify guidance; PDF page numbers identify where text was printed. Neither manual is itself legislation.
Start with one question
To investigate how a passage was divided, use the separate 28-case passage-boundary review. It compares the source PDF, exact extraction and proposed boundaries, and exports a local suggestion for separate review. It does not replace the 40-question evidence packages described below.
This page is a guide. Open the interactive Evidence workbench in Explorer. This link uses the reviewed, immutable catalogue with the optional inspection models; each downloaded package is checked against its recorded fingerprint. The original catalogue remains available. For an AI-led walkthrough, use the tested sidebar demonstration.
- Choose one of the 40 public staff-question occurrences. One question is repeated in the register, so there are 39 distinct wordings.
- Check the evidence status. Insufficient means the package cannot establish that all evidence needed for the task is present. It does not mean every selected paragraph is irrelevant.
- Open a selected passage and its cited PDF page. GOV.UK blocks embedded PDF display, so use the page-linked original-document button to open a separate browser tab. Compare the extracted text with the original, including qualifications and examples.
- Inspect the interpretation tabs: passage boundaries, concepts and relationships, dependencies, and the explanation of why the text was selected.
- Read the missing requirements and budget omissions. Export a review proposal if something should change; it stays local until separately reviewed.
What the terms mean
| Term | Meaning in this workbench |
|---|---|
| Source | The preserved official document and its exact downloaded version. |
| Extraction | Machine-readable text taken from the document; it can contain extraction errors. |
| Logical unit | A passage grouped because its text belongs together, possibly across PDF pages. |
| Discovery card | A short project-authored description and alternative wording used to find a passage. It is not evidence. |
| Concept | A named idea, such as capital or payment, with a stable identifier. |
| Relationship | A labelled connection between records. Its provenance and review status matter. |
| Dependency | Other text needed to interpret a passage, such as an exception, definition or cross-reference. |
| Provenance | Where an item came from and how it was transformed. |
| Evidence package | The selected text, relationships, explanations, source references and explicit gaps for one question. |
| Checksum | A fingerprint used to detect whether the delivered bytes differ from the declared version. |
Why the capital question failed
The question about treating capital as still held after someone disposed of it had relevant text in the corpus. The complete passage was labelled using terms such as notional capital. The question used ordinary wording instead. Its passage ranked below the 16 lexical candidates admitted by the earlier engine, and the available conceptual route did not activate. A larger final byte budget would not fix that selection failure.
The additive repair connects ten already repaired capital passages to the full DMG/ADM corpus, improves source-bound discovery, and follows the best passage's dependencies before broad graph expansion. It preserves the original PDFs, extractions, earlier releases and failed trial results. Everyday-wording routes are project-authored navigation proposals, not a change to the meaning or legal status of the guidance. Test results distinguish familiar development questions from new phrasings; finding the original question alone is not proof of general retrieval quality.
Small downloads, complete retained text
A package can be up to 512 KiB. It is delivered in separately checked responses of at most 32 KiB and reconstructed exactly. This avoids treating a small response limit as a reason to discard qualifications. It does not recover text which was never selected: selection gaps remain visible in the package.
Only the chosen question's evidence is downloaded. The catalogue lists all 40; private correspondence and unpublished questions are outside it. Source dates are not replaced by the date this catalogue was generated.
What review can and cannot establish
Source-byte integrity, useful selection, complete legal evidence and a good AI answer are different tests. All 40 cases can be examined without asserting that all 40 can be answered. Unresolved applicability, uncaptured legal provisions, ambiguous references and specialist acceptance stay open. No new answer-model trial is part of this repair.
See the implementation work log, the backlog work packages, and the preserved capital source repair.
Demonstration sequence
Choose a staff question, open its source, show the complete passage and its page spans, then show a dependency and the reason it was selected. Finish with a missing requirement and the machine-readable package. This demonstrates an inspectable grounding process, without turning an incomplete package into a complete benefits answer.
Ask a new question against the repair
Open the updated DMG and ADM source in Explorer, then select Ask OKF. This link fixes the source version to the reviewed release. Earlier pinned bundle links keep their original data; they do not silently acquire this repair. The 40-question workbench opens retained results, whereas Ask OKF assembles a new package for the words you enter.
Reproduce the retained evidence
The producer and the Explorer consumer must use the versions recorded in the workbench manifest. With a checkout containing the integrated delivery helper:
uv sync --locked
uv run --locked python scripts/build_evidence_connect.py --check
node scripts/test_evidence_connect_descriptor.mjs --explorer-root /path/to/okf-explorer
node scripts/build_evidence_workbench.mjs --explorer-root /path/to/okf-explorer --check
node scripts/check_evidence_workbench.mjs --explorer-root /path/to/okf-explorer
After dependency setup, these commands check the corpus projection, exercise
Explorer's actual Reader-to-Ask loader, repeat the 40 assemblies and compare
every byte, then independently reconstruct all retained packages without
rerunning retrieval. The loader check also needs Explorer's locked JavaScript
dependencies installed with pnpm install --frozen-lockfile. None calls an answer model.
See the source connection profile and measured tests.
The retained offline replay uses a virtual example.invalid index address.
That address is an identity within the replay, not a public download link; the
manifest's corpus path and checksum bind the committed source. Use the updated
Explorer link above for a new live assembly, or the documented producer commands
for byte-for-byte replay.
The new live Ask OKF entry uses structured-context/evidence-connect-explorer.json
from an immutable GitHub revision. In Ask OKF, the retained-case comparison
uses limits of 64 records, 128 relationships, depth 6 and 524,288 bytes; a
32 KiB delivery part is a different limit. The workbench consumes only its retained
public question packages. Remote MCP service admission and an observed ChatGPT
client journey are separate from this browser release.