Evidence workbench handover — 24 September 2026
What changed
The diagnosed capital question now finds the complete relevant passage and its linked guidance. The Evidence workbench lets a reviewer examine the evidence selected for each of the 40 public staff-question occurrences. One wording is repeated, so the register contains 39 distinct questions.
This is an independent research publication, 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 paragraph numbers identify guidance and PDF page numbers locate it.
Why the question failed
The text was already captured. The everyday question did not activate its specialist concept, notional capital, and the relevant passage ranked 29th when only 16 initial candidates could enter the evidence package. Increasing that package's final size could not recover an item excluded earlier.
The repair connects ten existing capital passages to the full corpus. It adds explicit, source-bound discovery wording, prioritises the leading passage's dependencies and delivers complete retained text in checked parts. The reusable Explorer engine has no benefit names or paragraph numbers hard-coded into it. The domain's authored navigation remains separate from official guidance.
What has actually been checked
| Check | Recorded result | What it does not establish |
|---|---|---|
| Original capital question | The whole U07 passage, DMG 84861 entry and their direct link were selected in actual Reader → Ask OKF browser use. | All possible exceptions, current-law applicability or a complete benefits answer. |
| Other wording | 5/6 positive development questions and 3/4 separately fixed confirmation questions selected the target; each set selected 0/5 negatives. | General success on unseen questions. Transfer-to-another-person wording still misses. |
| Forty-question delivery | 40 exact packages reconstruct from 673 parts, each no larger than 32,768 bytes. | Text never selected by retrieval is not recovered by delivery. |
| Selection comparison | An 80-assembly same-source comparison preserves gained and lost records and source spans. | Record counts are not relevance or answer-quality scores. |
| Browser inspection | Actual source links, passages, relationships, missing evidence, deep links, local proposal export and narrow layout were exercised. | A review proposal is not specialist approval. |
| Other bundles | The non-DWP study-club check and executed Heritage acceptance passed with prior receipts preserved. | DWP legal completeness. |
All 40 packages still declare insufficient: they do not establish that all evidence needed for a complete answer is present. That status does not mean that every selected passage is irrelevant. The workbench separates selected text, the original review brief, current requirements and remaining gaps so a reviewer can inspect the difference.
The original capital browser result contains 44 records and 68 relationships, using 522,995 of its 524,288-byte allowance in the final public run. The earlier local run used 522,919 bytes; each receipt retains its own URL and binding. Its retrieval and assembly budget omissions remain visible. No new answer-model comparison was run for this repair.
Use it in a meeting
- Open the Evidence workbench.
- Select a question and read its original review brief and current evidence status.
- Choose a passage and open its cited source page. GOV.UK blocks embedded PDFs; the source opens in a separate tab.
- Inspect the extraction, passage boundaries, concepts, relationships and dependencies.
- Show why it was selected and what is missing, then open its machine-readable package or export a local review proposal.
To reproduce the repaired question, open the updated source in Ask OKF and select Ask OKF. Enter:
Can Pension Credit treat capital as still held after someone has got rid of it, and are there exceptions?
Use 64 records, 128 relationships, depth 6 and 524,288 bytes to match the recorded run. Earlier pinned bundle links remain historical snapshots; use this new source descriptor for the repair. The remote MCP service admitted this source in the separately verified 0.7.0 release. Older connected clients may need the metadata refresh.
The beginner guide explains the terms as they arise. Search finds possible records; Ask OKF assembles evidence; the workbench inspects that evidence; an AI answer would be a separate interpretation.
Release identities
- Explorer PR145
merged as
e6084059f9b09633cfb8385be20099b952915e82. All required PR checks passed, including the full Chrome, Firefox and WebKit suite. The merged tree exactly matches the reviewed candidate. Its exact-commit Pages release and public verification passed in run 35969493742. - DWP PR36
merged as
7eeded763042ddd0070f4fed834c6074149e8e2fafter both protected candidate runs passed. Its merged tree exactly matches the reviewed candidate. DWP validation pins the merged Explorer implementation. Exact merged-main validation 35972890182 and Pages deployment 35975520154 passed. - The final public byte check matched 1,141 requests against the exact merged build, including all 1,140 declared outputs. This checks deployment identity; exact package reconstruction is checked separately by the required 40-question delivery gate.
- The final public browser observation
binds the permanent Pages manifest to
7eeded763, verifies all 40 catalogue entries and exercises three representative cases, the repaired capital query, narrow layout and local export. Its deep-link check preserves the selected question, passage and tab. Application console, page and HTTP errors were absent; blocked or aborted source-PDF frames are retained explicitly. Direct PDF links provide the fallback. - The separate public candidate observation
used hosted Explorer and immutable DWP data at
61f8539f. It is labelled separately from DWP's subsequent Pages release.
Reproduce the retained packages
The retained packages come from deterministic, offline replays. Their
binding.index_url uses example.invalid as a virtual replay address; it is
not a public download endpoint. The fingerprint binds the exact corpus bytes.
Use the immutable corpus manifest
and the commands in the beginner guide for reproduction,
or the live descriptor above for a new Ask OKF assembly. The cited official
source URLs and the workbench's package/part URLs are separate from that virtual
replay address.
Remaining work
- Broaden and independently evaluate everyday-wording coverage; retain the transfer-wording miss and the earlier false-positive experiments.
- Review the open dependency and applicability obligations. Captured entry links do not silently expand every cross-reference range or replace missing legal text.
- Use the workbench for specialist review of all 40 questions. The engineering checks do not establish legal acceptance or improved AI answer accuracy.
- Treat remote MCP service admission and observed ChatGPT-client compatibility as separate releases. This browser release does not silently update that service.
The machine-backed backlog separates implementation, public observation and specialist acceptance. Work ran in isolated parallel branches under one integration task; further steering can stay in that task.