Team handover: 20 September 2026

For the meeting on Monday 21 September. This is an independent experimental publication. It is not official DWP guidance, a benefits calculator or an individual entitlement service.

Ten-minute Monday demonstration · Beginner learning path

Later status: 21 September 2026

This dated handover preserves its earlier baseline. For the later service 0.6.1 publication and 11-case SDK observation, read the new release record and shared recorded status. The client guide records a corrected native control and the remaining compact-client/Data Agent acceptance gap. The small Staff 012 native check contains no evidence; no new AI answer or Voice acceptance is implied.

What the team can use now

Merged staff increment

DWP PR 12 and Explorer PR 126 are merged. The immutable DWP content commit is 9de52acf1db84b27f8933d80480eaa850e74fa33. The new evidence and model trials below are in the shared repository, not just an agent checkout. Canonical DWP validation passed, and Explorer Pages is deployed. The combined Reader public checks bind actual served bytes and scoped journeys in three browser engines.

Open the combined DMG and ADM Reader. For Monday, load the route and let its loading indicator finish before speaking.

Earlier published baseline

The earlier published baseline is DWP PR 10, merged as 5dab65d2ad21138c47e2ea364e5f74805b4e141f, and Explorer PR 125, merged as 8a38d4bfe07a6797deacc148a2859911d8e2a68e. The DWP main-branch validation passed. Its source inventories, generated outputs, question replays, privacy controls and 147 Python tests are reproducible from the repository.

• Five-minute demonstration: exact replay links, source inspection, gaps and the read-only AI tool instructions.
• Forty staff evidence packs: the supplied 40 question occurrences, including one repeated wording, with 42 shared source excerpts and review checklists. Original questions are retained exactly.
• Verified DMG Reader: benefit, circumstance and topic filters, relationships and source/capture date distinctions. Its published browser observation records exact application and corpus identities.
• Human evidence reader: request a small catalogue and read exact evidence progressively. Source identity is checked on replay; a replay link does not store an AI conversation.
• Actual Claude trials and seven verified compact-tool calls. Failed trials are retained alongside successful mechanical checks.

The captured corpus contains 331 DMG PDFs and 182 ADM PDFs: 19,090 pages in total. Ask OKF searches both captured manuals. The baseline Reader projection contains DMG; the new combined Reader below includes both manuals.

Included in the additive staff increment

• Question/persona/journey matrix: all 40 occurrences, seven primary journeys and six projected personas.
• Semantic guide and results: 43 concepts, 155 source-grounded associations, 60 selected source pages, 40 task profiles and 203 named open obligations. Candidate evidence is found for all 40 occurrences, compared with 12 before the increment. This is a development-set observation.
• Legal reconciliation: observed official provision identities and exact citation mappings, plus six bounded tribunal queries. Legal body evidence and applicability remain separate unfinished work.
• Combined Reader guide: both captured manuals, conceptual facets, source/audit dates, semantic paths and inspectable context packages. Local and public Chrome, Firefox and WebKit checks passed, including narrow-screen keyboard journeys and targeted accessibility checks. See the exact public observations.
• Paired staff trials: five identical-input cases through Claude and Codex, 12 retained attempts, 49 claims and 79 citations. Eight responses pass mechanical checks; two fail exact quotations. Separate model critique finds missing household/date qualifications. No specialist or comparative accuracy claim.
• Source refresh: all 513 attachment identities remain in the fresh listings; fresh PDF and extraction hashes are explicitly unknown.

The earlier baseline links above retain their original versions. New public journey observations and model-trial results have their own exact identities. Service 0.4.0 is live with the new staff corpus; actual SDK checks pass across all three supported source versions.

What remains incomplete, precisely

The 43 full-corpus evaluation cases still report insufficient. More captured pages and a working tool interface do not establish complete answers. The preserved custody example has a narrower, declared legacy-DMG evidence profile; its sufficiency does not transfer to the broader corpus or establish current law.

“Broader semantic modelling” is now represented by two bounded implementation deliveries with explicit remaining coverage and acceptance work:

1. DWP-BL-005: neutral benefit and variant concepts, source-backed definitions, meaningful relationships, aliases, confusions and qualifiers across the supplied question families. The new proposals implement these meanings for the selected staff scope. Literal mention tags are not these relationships.
2. DWP-BL-007: task-specific evidence requirements and paths, including conditions, exceptions, legal references, dates and explicitly missing evidence. All 40 profiles execute; their 203 named obligations remain open. Acquiring and reviewing those missing dependencies is not marked complete.

Further agent work remains under explicit .domain-expansion, .legal-body-evidence, .evidence-closure and .controlled-benchmark packages. These are not marked complete or hidden behind specialist acceptance.

The old aggregate needs_domain_review label obscured unfinished implementation. The work-package ledger now separates delivery from independent review and external permission. Implementation can proceed as clearly unreviewed work; that does not imply specialist acceptance.

The current service browser observation passes the staff evidence journey in Chrome, Firefox and WebKit. All twelve historical functional regression journeys also pass. Chrome and WebKit also pass strict console checks; Firefox retains hosting cookie warnings. BL023 therefore remains open. The earlier failed driver-decoding observation is preserved; the later check reads cloned browser-native response text and verifies rendered package hashes without substituting any response. ChatGPT Voice and the room PA have not been rehearsed.

Work proceeding in dependency order

Parallel research and producers may proceed where they do not depend on unfinished inputs. Their integration and publication still follow these dependencies. The completed BL004/008/009 baseline is preserved and tested, not rewritten. BL011, BL017, BL021 and BL022 continue to apply to every delivery.

CPAG substantive content remains subject to its separate rights gate (BL012). Actual user research, specialist acceptance and the room rehearsal require people or facilities; they are not replaced by agent agreement. P2 operational journeys, the benefits engine, calculators and CASA remain outside this run's implementation.

How to follow progress

Use the backlog, delivery and acceptance ledger, current work log and changelog. Only a merged change with its named checks establishes a delivered increment. An active branch, an agent assignment or a proposal is not a completion claim. This dated handover describes its baseline; later increments have separate evidence.

Private correspondence is ignored and uncommitted. Local browser logs and research scratch files are not part of this handover. No new messages have been sent to staff.