Remaining work and next priorities
Status reviewed on 26 September 2026. This is the current short guide to the work queue; the backlog register retains the detailed tasks and evidence. Older reports retain the results for their own versions.
What works now
- The declared Decision makers' guide (DMG) and Advice for decision making (ADM) capture contains 513 PDFs and 19,090 pages. These are DWP staff manuals, not legislation or a continually updated service.
- Search and Ask OKF can use the source-led corpus. The capital repair connects previously missed passages to source-bound discovery routes and dependencies.
- The Evidence workbench opens all 40 retained staff-question cases. It shows source passages, relationships, why evidence was selected and what is missing. Large retained packages can be reconstructed from checked, small downloads.
- The optional interaction and calculation-inspection views are implemented. The Edge sidebar route has been demonstrated. Calculation inspection explains proposed dependencies; it does not calculate an award.
All 40 packages still say insufficient. This means the system has not
established that every required condition, exception and applicable source is
present. It does not mean that nothing useful was retrieved. Conversely, a
working page or successful download does not make an answer complete.
Recommended order
| Order | Work and backlog IDs | What success would look like | Who can do it |
|---|---|---|---|
| 1 | Review the 28-case passage-boundary increment, tracked in Explorer #150 and DWP #42. BL007.remaining-structural-review | Inspect PDF, extraction and before/after passages, reproduce isolated corrections and review the narrowed parser's impact and fixed-budget question replay. Keep explicit residual defects and uncertain fragments open. | Agent implements and checks; independent reviewers assess source structure and applicability. |
| 2 | Repair the known question failures: capital transfer wording and the new Universal Credit (UC) disabled-child case. Verify the dated legal source before classifying the UC dependency. BL005, BL006, BL007 | The intended passages and conditional dependencies are selected for fixed positive cases and new wordings, while negative cases stay excluded. Gaps and uncertain applicability remain explicit. | Agent prepares and tests; benefits specialist assesses meaning. |
| 3 | Extend concepts and complete declared evidence requirements across the question set. BL005, BL006, BL007, BL009 | Benefit variants, dates, territory, household roles, definitions and exceptions have traceable relationships. Eleven capital dependency groups and the 43 original evidence-closure obligations are examined rather than hidden by larger retrieval counts. | Agent prepares scoped increments; specialist reviews applicability. |
| 4 | Measure selection, package size and delivery separately; restore the original acceptance cases in the newer corpus. BL007, BL008 | Whole relevant passages and qualifications survive bounded assembly; exact small reads reconstruct the selected package. The original imprisonment/hospital checks remain visible. The intended remote client is tested against the admitted source version. | Agent, with account-specific checks where needed. |
| 5 | Evaluate answers after evidence improves. BL010, BL018 | A fixed-source comparison checks each claim and qualification against delivered evidence, with independent review. Actual model identity and usage are recorded before claiming accuracy or affordability. | Agent prepares; specialist reviews. Further answer calls require an agreed usage budget. |
The boundary-review increment proceeds without restarting acquisition or running all 40 questions through more expensive models. An optional defect detector comparison requires an agreed model-call budget first: 12 varied cases, followed by 24 fresh cases only if the fixed criteria pass. The original notional-capital question is repaired, but transfer-to-another-person wording still misses. The UC follow-up selected one relevant passage but missed required conditional guidance. Both are measurable failures with existing source leads.
The local LiteParse pilot, BL007.liteparse-source-pilot, ran 12 precommitted windows and stopped at a failed gate. The reserved 24 windows were not parsed. Any follow-up should test a source-bound spatial sidecar and explicit header/partial-table handling under a new versioned protocol. Keep punctuation alignment separate from lost guidance and from retrieval gains. This deterministic parser experiment is separate from the optional model defect detector above, whose call budget remains a prerequisite.
The original amendment experiment remains preserved: its broad census changed 134 documents without a context replay. The new source review rejected its generic bare-Appendix rule after finding a false sentence split. The successor confines that change to historical amendments, checks all nine previously affected substantive chapters and replays all 40 questions. Read its measured results and residuals; the old branch review describes the earlier candidate, not the successor's acceptance evidence.
What still needs people or a particular environment
- Benefits specialists: review legal meaning, exceptions, effective dates and source completeness. The 203 original obligations include 43 source-closure requirements and 40 each for applicability, legal version, independent review and question scope. Engineering cannot mark these accepted on its own.
- Presenter: rehearse the intended ChatGPT account, microphone, MacBook sound, projection and room PA before the 30 September conference. A text sidebar demonstration does not prove Voice or room-audio operation. BL016.
- Representative users: check keyboard, screen-reader and other assistive technology journeys. Resolve the separately recorded Firefox hosting result. BL019, BL023.
- Rights and maintenance decisions: CPAG substantive reuse remains a separate rights question. A fresh source snapshot must be explicitly scoped and compared with the frozen evidence. BL012, BL020.
The benefits engine, application/change-of-circumstances journeys, external calculator comparison and CASA interface framework remain later work (BL002, BL003, BL014, BL015). They depend on established evidence and service requirements; they are not implied by the inspection prototype.
Repository housekeeping
PR 23 is closed as superseded; its history and branch remain available. The
reviewed implementation is already on main. Local research, failed experiments,
private correspondence and older source snapshots are preserved. The original
amendment branch remains separate; the successor has its own versioned output.
Documentation and backlog corrections from
PR 40 are merged and published.
The service/sidebar follow-up and boundary-review design were merged in
PR 41. The passage-review
implementation is merged in DWP PR 44
and Explorer PR 155.
Its public observation
checks all 28 cases; the partial/unresolved source findings and separate parser
adoption remain open.
Future progress should update the authored backlog JSON, regenerate its work-package page and record the exact tests and release in the changelog. A delivered implementation, a passed test, a public observation and specialist acceptance are separate milestones.