Evidence repair and workbench delivery

Started: 24 September 2026. This is the current implementation log; the Monday 21 September log remains a historical record.

Outcome requested

Fix the diagnosed natural-language retrieval miss, connect and select captured evidence more effectively, deliver it in bounded parts, and provide an Explorer Evidence workbench where every one of the 40 staff-question occurrences can be examined. All 40 being examinable does not mean all 40 are legally answerable.

Starting evidence

DWP main is 44488b89d82f191f189e7ab57364d554959be9e6; Explorer main is dc54fac4f98fa5bc9db38bf843854ce24a4c130d. The independent capital retest reproduced 52 assemblies without an answer-model call. The subsequent diagnostic reproduced the failing question's chapter and full-corpus packages exactly: the grouped passage ranked 29th against 16 lexical candidates; the everyday wording did not resolve the notional-capital concept; the corresponding guarded routes remained inactive. The chapter package had unused assembly capacity. Appending the specialist term selected the passage but did not establish evidence sufficiency. Neither diagnostic is a repair or a general quality trial.

Parallel ownership and recovery

The agents use isolated worktrees. The integrator alone combines changes and publishes reviewed candidates. New steering changes the relevant workstream; unaffected workers continue. App interruption, usage limits and external service failures can still pause execution. Resume from this log, Git state and live agent status before starting another writer. No new answer-model experiment, embedding job or source acquisition is included in this phase.

Acceptance gates

1. Preserve frozen source bytes, earlier corpora, replays and trials. Keep private correspondence outside Git and publication.
2. Fix the original question and test varied unseen paraphrases and wrong-benefit controls. Compare matching representations across narrow and full corpora.
3. Add only source-backed dependency links; retain ambiguity, absent source bodies and specialist-review obligations. Report concept resolution separately from lexical discovery and legal applicability.
4. Re-run all 40 public staff-question occurrences. Record source/version, engine, selected and missing evidence, paths, truncation, byte use and measured timings.
5. Reconstruct each selected package exactly from responses at most 32 KiB. Report total transferred bodies and overhead as well as individual response limits.
6. Navigate all 40 cases in the workbench. Inspect the source PDF location, text, complete passages, concepts, relationships, gaps and selection reasons; export review proposals without modifying source evidence or claiming review acceptance.
7. Run affected contracts and tests, independent code review and browser journeys; merge through protected PRs, then verify the exact public versions separately.

Publication impact

New authored DWP profiles and generators produce additive evidence projections and a bounded workbench manifest with lazy case packages. Existing source and historical projections remain frozen. Explorer changes the opt-in discovery implementation, delivery helper and workbench routes. README, beginner guidance, the machine backlog, work-package register and changelogs move with the code. The learning-site exporter admits only explicitly listed, hash-bound workbench artefacts; it must not copy arbitrary evaluation directories or private files. The remote MCP deployment and actual client compatibility require their own observations and are not implied by a static workbench or offline test.

Checkpoints

• Initial inspection: no active implementation agents were left from the retest. Existing working directories and the separate security-review checkout remain untouched. DWP main validation and Pages checks passed. New branches start from the current fetched main references. An unrelated historical handover PR and Explorer dependency/learning PRs are left intact.
• Implementation is in progress. No new completion, merge or public deployment is claimed at this checkpoint.

Integration checkpoint

• Explorer includes opt-in weighted discovery, scope-bound navigation routes and early source dependency traversal. The original failure is repaired; the development probe run retains one transfer-wording miss and all insufficient statuses. Earlier failed probe variants are retained.
• The workbench has seven source-to-review tabs, lazy case loading, shared raw and multipart integrity checks, deep links and local proposal export. Review found and fixed request races, stale history, and unbounded stream continuation.
• Independent publication review found restricted-record admission, source-hash and output-collision defects. These are corrected; eight new publisher controls and 26 existing website tests pass. Full JavaScript reconstruction for all 40 retained packages is a separate required gate.
• Explorer's initial integrated full unit suite passed 767 tests. The full static build initially failed because a plain Vite build lacks the required canonical build manifest; the documented deterministic-build prerequisite is now used.
• No sources, historic snapshots or answer-model trials were changed. Current work is still a feature candidate; no new public deployment is claimed here.

Reviewed feature checkpoint

The producer is committed at bea541f2. Explorer PR145 contains the reusable engine, workbench and delivery implementation. Its initial CI run exposed two compatibility issues: newly external schema references were not registered by legacy MCP/archive consumers, and a learner-hub test assumed a generated route number. The fixes preserve local schema loading and test actual route isolation.

The independent delivery checker passed 40 contexts and 673 parts; maximum part size is 32,768 bytes, and the retained canonical text totals 20,303,318 bytes. All 40 statuses remain insufficient. The same-overlay 80-assembly comparison records 15 cases gaining records and 14 losing records; these counts are changes in selection, not relevance scores. Development tests select U07 and its 84861 dependency for five of six positive questions, and a separately fixed confirmation set selects three of four. Both sets have zero of five negative selections. Transfer-to-another-person wording remains an explicit miss.

Independent producer review confirmed unchanged text, source spans, provenance and assertion status for all ten imported passages. It found and corrected output-path containment and missing input-hash bindings. Case expectations are used only after assembly for assessment. No benchmark IDs seed retrieval.

The required CI now includes a separately pinned workbench job: deterministic projection, all-40 reassembly, independent exact reconstruction, output-path and publication controls, and an exact-commit static-site build. The protected aggregate check requires this job alongside both earlier validation planes.

Final local integration checks

The successor descriptor now retains the original Reader snapshot while binding the additive context overlay separately. An actual Reader-to-Ask loader test caught the previous snapshot mismatch; the corrected descriptor selects U07 without changing the source Reader. All 40 packages and the comparison/probes were regenerated against context manifest 4e65253f72b2d771164d5ddd64374ed8142cf66b4a249f193a1fe44d5679f519.

Local Chrome journeys cover the abroad, care-home and Pension Credit abroad questions, original review briefs, source links, passage and relationship views, local proposal export and a 390-pixel layout. GOV.UK PDFs reject embedded display: the page-linked source document opens separately. An iframe element is not proof that a PDF rendered. The earlier local receipt is preserved under validation/evidence-workbench/2026-09-24/local-before-descriptor-fix/, including its absent timestamp/build identity and failed PDF requests.

Review also corrected a double browser-history entry when following a passage from a relationship. Explorer's required CI exposed an obsolete Heritage app fingerprint; its old acceptance files are preserved. The new deterministic app build passed the actual 100-question browser acceptance and all three Heritage journeys. The non-DWP study-club trial has a separate new execution receipt; its old receipt remains unchanged. These checks test reuse and regression, not DWP legal completeness.

Public deployment remains pending at this checkpoint. The workbench's implementation package is recorded complete; its public-verification and specialist-review packages remain separate and open.

The older structured-corpus checker initially rejected the new additive namespace as unbound surplus. Its source inputs and frozen outputs were not rewritten. A separate wrapper verifies every successor byte, temporarily sets aside only its 288 declared files for the exact base check, restores them and verifies them again. The full 53,727-record base check and five restoration/containment controls pass. CI runs the base and successor planes independently.

The final local Reader-to-Ask browser journey loaded the new structured-context/evidence-connect-explorer.json, asked the exact original capital question, and selected U07, the DMG 84861 entry and their direct link. At 64 records, 128 relationships, depth 6 and 524,288 bytes, the result contains 44 records and 68 relationships using 522,919 bytes. It reports budget truncation and remains insufficient. Browser console, page and request errors were absent. The receipt and screenshot are in local-final/reader-ask.json and local-final/reader-ask.png; the earlier combined-descriptor check is retained separately and is not used as evidence of the successor working.

Explorer PR145 merged as e6084059f9b09633cfb8385be20099b952915e82 after all required checks, including Chrome, Firefox and WebKit. Its Git tree exactly matches the reviewed candidate. DWP's new workbench CI now pins that protected-main consumer commit; the recorded engine and delivery file hashes are unchanged. This is merge evidence, separate from the subsequent public deployment observation.

Merged and public verification — 24 September 2026

Explorer PR145 merged as e6084059f9b09633cfb8385be20099b952915e82. DWP PR36 merged as 7eeded763042ddd0070f4fed834c6074149e8e2f; the reviewed candidate had the same Git tree. All protected checks passed. The subsequent DWP main validation run 35972890182 and Pages deployment 35975520154 succeeded. Explorer's Pages run 35969493742 also passed; these run identities are separate from browser and content acceptance.

The version 3 public-site verifier matched 1,141 of 1,141 declared public responses against a locally rebuilt site bound to merged commit 7eeded763: the site manifest and 1,140 listed outputs. The site receipt records the immutable Git inputs, exact bytes, fixed host, bounded concurrency and zero retries. The public browser receipt records the 40-entry workbench catalogue, three representative case inspections, deep links, a locally exported unreviewed proposal and the live Reader-to-Ask journey. External GOV.UK PDF embeds did not render reliably; the source page links remain the intended way to open those documents.

In the public capital journey, the original question selects U07, the DMG 84861 source entry and their direct dependency. The package contains 44 records and 68 relationships, uses 522,995 of 524,288 bytes, and reports truncation and insufficient evidence. The earlier local journey's 522,919 bytes belong to its separate receipt and are not a substitute for this public observation.

Capital delivery and additive-overlay adoption, bounded workbench delivery, and BL025 public verification are now recorded complete for these exact revisions. Eleven declared capital dependency groups remain open overall, even though four captured groups gained exact entry links. BL025 remains in progress because independent specialist review is open. See the handover for source, engine, site and browser identities and the continuing limitations.

Additive out-of-sample follow-up — 24 September 2026

The UC disabled-child case preserves one SHA-bound, offline replay of a new hypothetical question. It selected ADM F1123 but not A4361, A4352 or F1120, returned insufficient and exhausted query, candidate and resource budgets. The unrelated Staff 008 requirement also activated. The plain-English source review explains why the conditional effective-date wording cannot yield a calendar date without UC assessment-period boundaries. The original 40 cases, source snapshots, published workbench and model trials were not changed. BL007 selection/profile repair, BL006 statutory reconciliation and independent specialist acceptance remain open. Demo 2 and the hospital question are separately labelled leads; no live WebMCP response or new hospital provision is claimed. The authored backlog now cites the exact case under BL006 and BL007, with an open agent package for UC source classification and controls. Its generated work-package page retains the completed historical packages unchanged.

Branch review and sidebar handover — 24 September 2026

The unmerged-work review distinguishes squash-merged work from the unfinished amendment parser, stale local trial files and the superseded PR 23 checkpoint. Three independent bounded reviews checked the parser, learning-page fix/trial files and PR 23. No old source projection or working file was copied over current main.

The sidebar demonstration records actual page-tool calls from Edge's ChatGPT sidebar: staff-016 Requirements, then a plain-English follow-up opening staff-039 Interactions. Restricted inspection had incorrectly suggested the native interface was absent. The working route uses the existing developer connection; dedicated WebMCP discovery remains a separate gate. The guide now distinguishes itself from the interactive Explorer page and links the reviewed model manifest. No evidence status or calculation boundary has changed.

Connection routes and repeatability — 25 September 2026

The connection guide now separates three routes: saved, hash-checked workbench packages with seven page tools; fresh evidence assembled in Explorer Reader and Ask OKF; and the separately versioned remote MCP service with three read-only tools. The saved inspection manifest at d31f16fb… and additive source descriptor at 7eeded763… have different roles. The remote service's latest recorded deployment remains on the generated service status page; preparing a newer service does not itself publish or admit its source. At this checkpoint, the separate 0.7.0 service implementation is being prepared to admit immutable DWP source 7eeded763… with Explorer engine source d6930bbc…. It is not a recorded public deployment or a change to the saved d31f16fb… inspection packages.

A low-context remote starter now begins with a normal public question and the small evidence catalogue, then reads selected passages and diagnostics using the returned version, engine, context and budget. A complete-package read and independent byte check are a separate, longer acceptance journey. An older cached connector schema may require refresh and a new conversation after a recorded service release; tool discovery and actual calls still need checking.

The sidebar guide preserves both fresh-session attempts. The first encountered a blank or unavailable native interface and clipboard/window errors before any prompt or model call. A later conversation successfully used the existing authorised CDP connection to call the native page tools and show staff-016 Requirements. Independent page inspection saw the move from Calculation stages to Requirements and three of six visible rows. The sidebar kept the result insufficient, identified missing source references and made no award calculation. This is one scoped repeatability observation, not a dedicated WebMCP-discovery, Voice, room-audio or answer-quality acceptance. No permissions or settings changed; the user's other draft was untouched.

Service admission and scheduled boundary review — 25 September 2026

The earlier preparation checkpoint above is superseded by the recorded service 0.7.0 publication. Explorer PR149 merged after required checks; its exact build passed the public 12-case, 135-request SDK check. The native connector's initial stale-schema rejection, metadata refresh and complete hash-checked diagnostics read are preserved separately. The sidebar result remains scoped to its saved manifest.

Explorer #150, design PR151 and DWP #42 schedule the next boundary-review increment under BL007.remaining-structural-review. The first usable increment inspects all 28 known cases. Generic controls stay in Explorer; DWP owns source fixtures, nine-chapter impact and all-40 equal-budget replay. No duplicate parser implementation, source acquisition or model trials were started. Optional 12/24 defect-detector experiments need an agreed budget first.