Product and evaluation backlog

What changed · Current work log · Retrospective · Portable method

Start with remaining work and next priorities for the 25 September overview. The detailed historical increments below retain their own versions and measured scope.

This register keeps stable IDs for the owner’s original ideas and later findings. DWP-BL-001, DWP-BL-002 and DWP-BL-003 preserve the original three backlog items. Dates, exact dependencies and acceptance checks are in the machine-readable register. IDs are never reused.

recorded_complete means a named bounded result has evidence. in_progress is not a release claim. needs_domain_review and needs_external_permission name the missing decision; not_started records work that has not been done. P0 means the next demonstration or evidence-quality priority, not an instruction to bypass source, rights or review boundaries.

What broader semantic modelling means

The original work was incomplete; it was not recorded as done. The previous single needs_domain_review label hid unfinished implementation behind a review gate. BL005 covers neutral concepts, variants and source-backed relationships. BL007 covers task-specific evidence requirements, routes, exceptions and explicit gaps. BL001 supplies the question/persona scope; BL006 supplies legal reference reconciliation. BL009 mention classifications do not complete either semantic milestone.

The work-package register now separates delivery from independent review, external permission and live observation. A pending human review does not prevent implementation of source-backed, clearly unreviewed proposals. A completed implementation does not close the human review.

The 20 September team handover preserves the baseline and plan for the earlier Monday demonstration.

The 20 September increment now supplies bounded implementation for BL001/005/006/007/013/020. Their bounded delivery packages link actual outputs. Separate agent work packages now name remaining domain expansion, legal body acquisition, evidence closure and a controlled model benchmark; these are unfinished implementation, not blocked behind human review. Independent acceptance remains a separate work package. The 203 named task obligations are in the profile register. Do not read a delivered compiler or a denser graph as complete domain modelling.

The qualification-retention work package records the preceding 15-support-dependency increment: eight for the initial household group, five for housing costs and two for temporary residence. Staff 012 and 013 require the household and housing-cost paths. The component review motivated a separate no-partner and severe-disability overview increment. The original joint comparison now passes 320 assemblies and replay: the final pair retains all 177 known candidate occurrences and 433 declared path occurrences at both tested budgets, including seven of seven household pages for Staff 012 and 013. This bounded retention package is recorded complete. The disability increment is also implemented and independently reviewed: 29 dependencies in a 903-record index, with a separate frozen comparison. All 497 declared path occurrences survive 512 KiB; only 407 survive 256 KiB, despite unchanged 177/177 candidate overlap. Its bounded implementation is recorded complete. The partner extension is now also implemented and independently source-reviewed: 39 dependencies, 585/585 retained path occurrences at 512 KiB and 449/585 at 256 KiB. Its source, combined projection and separate comparison are distinct from public delivery. The conditional ignored-person and normal-residence extension is now authored and independently source-reviewed: 913 records, 53 concepts and 61 support dependencies. Its larger declared closure exceeds even 512 KiB; this measured retention regression is tracked separately from its completed source modelling. See the source review and separate comparison. Exact-version publication and public observation retain a separate in-progress package: the qualification Reader has a new public Chrome receipt, while the later source increments retain their own release checks. Service 0.6.0 has since passed its separate 121-request public SDK observation; the retained reader has a distinct public publication gate. All 43 evidence-closure checks and all 203 total obligations remain open. At 64 KiB, Staff 012 correctly refuses because its obligation metadata does not fit; a small refusal does not erase the underlying task requirements.

The paired direct-v4 results now include two mechanically accepted care-home answers after both empty controls. The independent agent critique identifies qualification omissions and an overstated gap in one response. BL010 remains open for qualification repair and human assessment; a completed pair does not imply complete legal answers or a model ranking.

DWP-BL-008.client-connection remains in_progress. Service 0.6.1 is published and its new SDK observation passes 11 cases in 121 requests. After refresh, the ChatGPT connection advertises all three tools and the corrected question pattern. The same native control rejected before the patch now succeeds. The exact Staff 012 question also reaches the service, but its deliberately small 16 KiB budget returns zero records and an explicit byte-budget omission. These client observations establish neither useful answerability nor complete client integration.

The existing Codex task still exposes only the full-package tool. A new conversation's compact catalogue and exact-read journey, and tool propagation into the intended Data Agent, remain unaccepted. The connection guide preserves the fixed comparison question and budget. Neither installed metadata nor the SDK pass establishes Data Agent or Voice acceptance.

Abroad review and wider semantic follow-up

The 21 September audit traces the reported five irrelevant pages to the preserved 19 September pre-fix result, then tests the published source and engine. The exact question now resolves abroad, but overseas paraphrases, disconnected international concepts, incomplete passage continuations and broad profile activation expose remaining gaps. The machine register names separate connectivity, qualification, cross-benefit coverage, task-discrimination and historical-version competition packages; none declares the broader domain complete.

Logical retrieval units — 22 September 2026

The owner has authorised implementation of the reviewed RQ02 correction. Logical evidence units and the implementation log track three active packages: DWP-BL-005.logical-units, DWP-BL-007.unit-qualification and DWP-BL-009.logical-corpus. Source capture, semantic coverage and specialist acceptance remain separate. Existing page-based releases and service defaults retain their recorded scope until new checks justify publication.

Current register

Acceptance before closure

Each item has an owner role, explicit acceptance checks and evidence paths in the register. A change from in_progress to recorded_complete needs the completed check or observation, the exact source and consumer versions, and any remaining limits. Model agreement does not close a domain-review gate. Keep partial results and failed attempts.

The recorded delivery and navigation milestones have scoped receipts: public service/SDK, local Claude, and exact local and published DMG Explorer. Service-browser deployment checks remain separately recorded: the historical-profile functional assertions passed, while all nine version 6 tests and all twelve version 7 tests failed their strict console gates because of hosting errors. See the earlier result and corrected-service result. DWP-BL-023 records the unresolved integration; those milestones do not close domain review. Independent answer and applicability review remain separate. Legislation integration must research provision/version identities before adding legal implications. A benefits engine and operational application/change journeys depend on those reviews.

DWP-BL-009 has a deliberately narrower Reader scope. Its classification audit accounts for both captured manuals, and Ask OKF already retrieves from both DMG and ADM. The preserved baseline human Reader contains DMG records. DWP-BL-024 now supplies an additive combined Reader, with source-family and date distinctions and locally verified filter parity across Reader/Graph/Timeline. The earlier staff-source public observation and household-source public observation retain their exact source and application versions. Neither attests the later qualification candidate. Literal mentions are not legal assertions. The published DMG navigation check is separate from the unresolved service hosting-console issue.

The logged sources retain the tribunal, calculators and CASA requests. CPAG remains separately constrained. The 24 September content freeze and 30 September seminar do not authorise an unsupported completeness or Voice claim.

DWP-BL-023: supported hosting follow-up

Cloudflare documents Cache-Control: no-transform as preventing JavaScript Detections injection; its detection result then becomes missing. The later service preserves no-store and its Content Security Policy while using no-transform. The actual 0.5.0 observation passed all three functional browser journeys and the Chrome/WebKit strict console checks. Firefox retained two hosting-cookie warnings, so its strict check remains failed. Historical-profile journeys were not run for 0.5.0. Preserve the earlier public-browser failures and keep this work item open for the unresolved hosting acceptance.

Recheck the register

Run uv run --locked python scripts/check_backlog.py. It checks IDs, declared statuses, dependency cycles, public evidence paths and consistency of this table with the machine register. It also checks work-package evidence and rejects a completed parent that conceals unfinished work. Regenerate the detailed ledger with uv run --locked python scripts/build_backlog_work_packages.py; validate it with the same command followed by --check. Regression controls run with uv run --locked python scripts/test_backlog.py. These are structural checks; they do not decide whether a domain or acceptance review has passed.

Logical-unit implementation — 22 September 2026

The logical-unit guide records the completed source producer, scoped qualification fixtures and reusable corpus v2 integration. These are bounded implementation packages, not completion of broader semantic modelling. The current increment records 75 authored and 49,551 uncertain boundaries, with 32 of 40 staff occurrences still lacking a unit profile. Separate packages track profile migration and compact-delivery admission. The five small-budget focused packages retain no source evidence; the larger packages retain 37 of 37 declared paths and remain insufficient. Bounded delivery of those larger packages is a separate checked capability.

Discovery ranking diagnostics — 22 September 2026

DWP-BL-007.discovery-ranking-probes records two completed offline probes with independent arithmetic checks. The retained results reject simple alias expansion and identify length-aware ranking as a lead. They do not close DWP-BL-007.discovery-cards, task discrimination, unit boundary review or answerability. Summary cards and their independent relevance evaluation remain unimplemented; no production ranking change is claimed.

Evidence workbench — 24 September 2026

BL025 tracks all-40 question inspection separately from source-bound selection (BL005), dependency closure (BL007), delivery (BL008) and specialist acceptance. The current work log records ownership, tested milestones and publication gates. Familiar-question repair and prospective paraphrase results remain separate.

The out-of-sample UC disabled-child follow-up records a new question outside the frozen 40. The first bounded replay found the higher-rate condition but missed ADM A4361's effective-date branch; it remained insufficient and the assessment-period dates were unknown. BL007 must address source-bound selection and task discrimination without treating a topical hit as a complete rule. BL006 must verify the governing statutory provisions and temporal/territorial scope; BL025 specialist review remains open. The separate Demo 2 application idea and hospital-source/transport leads are recorded in the same follow-up, with no claim-submission or hospital-answer milestone inferred.