Reading-help rollout work log

26 September 2026: baseline and bounded parallel work

The owner authorised deterministic DMG and ADM processing, scalable reading help, focused legal citation improvements and a separately gated legislation-format reconciliation. The presentation target is 28 September at 18:00 Europe/London. No new automated content-processing or answer-model comparisons are authorised for this phase.

The baseline receipt records exact DWP, Explorer, source-manifest and service-observation identities. It is an observation, not proof of continuous service availability.

• DWP baseline: 2c46fbd4e0807b1c9dfa0e9d79e33987d245ef21.
• Explorer baseline: 0326460034345c7b5227c530065b7f21cb4e3970.
• Structural baseline: 513 documents, 19,090 pages, 893 pages without extracted text and 53,727 units.
• The later 54,577-unit amendment candidate remains separately identified. Its existence does not authorise silent replacement of the structural baseline.
• Existing statutory evidence: 20 units, 16 provisions, 43 navigation relationships.
• A new live health request returned HTTP 403. The earlier recorded service 0.7.0 deployment and verification receipts retain their historical meaning; fresh health was not established by this attempt. A later browser-style user-agent observation returned service 0.7.0 ready. The connected compact-manifest tool observation also succeeded, while explicitly returning metadata_budget, no selected records and insufficient at the deliberately small 32 KiB selection budget. These are separate observations, not a whole-corpus answer test.

Writer ownership

Both agent launches initially failed because of model capacity. One retry succeeded. At most two implementation agents are active; no content-analysis model batch was started. Worktrees and cache are on EXTSSD. The primary DWP checkout's unrelated client-report edit and private email remain untouched.

Gates and outstanding work

1. Review and merge the Reader URL/full-text repair before admitting additional statutory text to a consumer.
2. Validate the shared catalogue/document/passage contracts and their consumer. Preserve the Chapter 60 v1 demonstration and URLs.
3. Produce a complete processing/gaps ledger; validate 12 cases and, only on success, 24 fresh held-out cases.
4. Replay the 40 staff-question occurrences at equal source scope and budgets, with source evidence and legal answerability reported separately.
5. Reuse recorded 28-case structural outcomes and retain unresolved case 014; do not repeat completed parser work without a changed rule or new concern.
6. Validate explicit admission of the bounded legal bridge.
7. Reconcile the legislation repository in isolation after the first consumer increment. Preserve dirty migration work and v0.3.0.
8. Verify exact merged commits, public Pages, workbench and service separately. Tag only a passing paired consumer/data release, retaining the earlier demo fallback.

The term full processing coverage means that every captured document has a recorded outcome. It does not mean every passage has an explanation or every question has a legally complete answer. Completed stages must link to retained checks and receipts before their backlog status changes.

27 September 2026: retained checks and limitations

• Explorer Reader repair is under review in PR 158. CI initially found the new browser suite missing from the impact registry; the correction is separately committed.
• All 40 staff occurrences replayed with the frozen source, engine and 512 KiB/64-record/128-relationship/depth-six budget in both cold and warm runs. Exact packages were unchanged. The report measures local latency, fetched bytes and cache reuse. This is non-regression, not improved retrieval or a legal-answer benchmark.
• The optional Chapter 60 statutory context validates against the actual Explorer context schema. Its first broader natural-language probe failed selection under its fixed budget; see the bridge report. The failure remains visible and no default service data was replaced.
• The dirty legislation migration was preserved as a 2,532-file, hash-inventoried archive on EXTSSD before selectively importing 30 authored/producer files into an isolated branch based on current remote main. Frozen generated source evidence and v0.3.0 remain preserved.

Publication and paired browser gates remain outstanding until their exact receipts are recorded.

The 28-case passage review was rechecked from its frozen inputs on 27 September: 28 cases, 55 files, manifest SHA-256 488df312d1672270e9d458ad0944a30e07d2fc56062896ad3d9e8ca2ad0a21f2. Its recorded nine-chapter review is retained. Case 014 remains unresolved; this check does not turn that finding into acceptance.

Independent control review

The first held-out reading-help run retained 23 passing cases and one failed case. A separate coordinator check confirmed the H05 extraction hash and surrounding source text: the source prints see DMG 070831 to 070834, whereas the frozen expected target named only paragraph 070831. The complete range must remain unresolved rather than being narrowed to satisfy that expectation. The original failed case and result remain unchanged. A further unexecuted set was also rejected at review because et seq (and the following paragraphs) needs to remain visible, and the case mix needed stronger continuation and abbreviation coverage. These are quality controls, not specialist legal acceptance.

CI verifies the retained 40-question receipt with python scripts/check_reading_help_replay.py: source and engine declarations, all 40 package hashes, selected text and paths, obligations, and performance totals are checked without rerunning 80 assemblies. A changed engine or evidence baseline requires a separately named replay, not an overwritten receipt.

Admission correction, 27 September 2026

The coordinator's engine review found that the first optional legal context omitted source-capture timestamps and confused per-fragment hashes with the complete record text hash. The initial failed observation is preserved under evaluation/reading-help-rollout/failed-admission-01/. The corrected projection uses the original acquisition/extraction receipts, complete record digests and separate exact fragment metadata. Every added record and relationship now passes Explorer's governance check; fragment integrity is checked independently. The same question and budget still fail to retain the bridge, so no retrieval improvement is claimed. Focused local reading-help tests: 35 passed, including six workbench-link controls whose data publication awaits the corpus commit.

Beginner navigation integration

The glossary and authored learning/p01/s03 activity now point to the same reading-help walkthrough. Regenerating the additive Reader changed its snapshot identity, so the preceding optional-context observation is preserved under evaluation/reading-help-rollout/pre-learning-navigation-01/. The successor uses the same query, engine and budget and still does not select the bridge. The 40 frozen staff packages remain unchanged. The local focused suite now passes 36 controls, including rejection of conflicting retained record IDs.

Corpus and workbench integration

The corpus producer is published for review in DWP PR 50 at commit 115397c19e0b1e1e16d476df788c33a1196a3990. Catalogue SHA-256: acb63e18d10c1a8a0a6c7d57f8ebfc41948e3837cd27c4589d4540061507d8a6. It accounts for 513 documents, 19,090 pages, 53,727 units and 893 empty extractions. There are 1,455 bounded leaves and 197,564 delivered literal occurrences; proposed meanings have no specialist acceptance. Shared parser, helper and printed-table inputs invalidate the cache when changed.

The additive workbench manifest retains all 40 cases and links 401 of 403 distinct retained evidence records by exact unit and source identity. Capital composites U06 and U10 remain unmatched rather than receiving guessed links. The integrated focused reading-help suite passes 48 controls. The small and held-out gates and their failed predecessors remain separate artefacts.

A second agent independently checked the legal bridge, including all 16 resolved citation mappings, four work identities, separate territorial variants, rights notices and the actual context-engine checks. It found no further material fault. This technical review is separate from legal acceptance.

The independent legislation migration remains a candidate. Exhaustive semantic validation passed 929,053 assertions, but browser release checks have not passed. Review identified public-main escapes from pinned/local federation and narrow graph links; these are being corrected before another candidate observation. The first runtime attempt also stopped because the coordinator archived an old receipt during the identity check; that failure is retained. No new legislation release or specialist acceptance is claimed.

Frozen-context integration correction

CI detected that the new teaching link inside the combined Reader changed inputs used by frozen logical and structured context projections. The teaching link now lives in the beginner documentation. The combined Reader and authored learning activity were restored to their exact pre-increment bytes; no frozen context outputs were regenerated. The withdrawn optional-context observation is retained in evaluation/reading-help-rollout/reverted-learning-navigation-01/. Occurrence-specific workbench links remain a separate additive sidecar.

Portable corpus and complete consumer audit

Corpus commit cbada7bf544106bdceb2445fc379c8b3e5b73ee4 uses the existing platform-neutral gzip writer. The catalogue SHA-256 is 6c53b659b519aa4cbf7529e7e10eb4ad3bfe8efcac18055ec2a5ade724998b46. The unchanged 12 and 24 source expectations pass again; earlier failures remain frozen. All 1,455 compressed headers, exact hashes and byte limits were checked. An independent full compatibility audit found no remaining violations against the updated consumer limits, including the printed table with 166 abbreviations. The integrated focused reading-help suite passes 50 tests. The withdrawn learning-only change was rechecked: all 4,905 logical-context and 4,758 combined Reader outputs reproduce exactly from their original inputs.

Merged consumer and follow-up checks

Explorer PR 158 merged as 936da562e4629ed2021bb09fcc55c6ac53ca3a9c on 27 September. All required checks passed at head 6b7ec7c0b181ed431b2f04c0fbb5a46b9f04a113, including the full three-engine browser suite. Public Pages verification is a separate gate. The corpus producer merged through DWP PR 50, commit 239f974c6cba23a2a37182af722ebba0803504b6.

The independent legislation browser gate exposed undersized facet controls. A separate Explorer follow-up repairs those controls and explicitly separates a footer row's source identity from a target in another document. The existing corpus contains no fabricated cross-document targets. Historical failed runtime receipts remain retained in the legislation candidate; its publication remains separate from this presentation release.

Public browser defects retained

The first public browser attempt passed immutable data checks but failed in the verifier's accessibility setup; explicit browser contexts fixed that setup. The second attempt passed 14 of 16 checks against the merged public consumer. It found two actual accessibility defects: adjacent extraction-gap links need larger targets or spacing, and the workbench's scrollable source text needs keyboard focus. Both repairs remain on the existing consumer follow-up branch; neither failure is waived or represented as a public release pass.

Exact DWP documentation observation

The seven checked public documentation resources matched the successful Pages artefact for merged DWP commit f61d41b689427c27275ae3b7e64b612925b6321e (workflow 36289751755). The receipt and bounded verifier are retained under evaluation/reading-help-rollout/release-20260927/learning-site-attempt-01/. This verifies the implementation documentation at that revision, not the later release notes or the still-pending consumer browser repair. The offline release gate now checks the recorded file bindings and cannot promote a candidate to verified without the complete paired public browser evidence.

Follow-up consumer merged

Explorer PR 159 merged as b633038c6173f1405199df9f08e8d8aa2ff51c0b after every CI job passed at exact head 26af7231c8350d123bbd72324dfbfeca9feaa497 (workflow 36290277261). It fixes the two public accessibility findings, keeps proportional facet displays accessible, and validates explicit cross-document targets against their exact leaf segments. Public deployment and the final paired receipt remain distinct checks.

Paired public acceptance passed

The exact merged Explorer deployment passed workflow 36291402585. Public attempt 03 then passed all 16 checks, including source spans, continuations, extraction gaps, all 40 bound case sidecars, dated references, keyboard/accessibility and the original v1 fallback. Consumer b633038c, corpus cbada7bf, workbench 2ad33add and the DWP implementation Pages observation f61d41b6 remain separate identities in the release manifest. Failed attempts are retained. All 40 question packages remain insufficient; no model answers or specialist review were added. Final release-record deployment and paired tags follow through the normal reviewed PR.

Release-record CI correction

The first release-record CI run rejected an unregistered work-package evidence link: the 28-case recheck was linked from its structural-review package but missing from the parent backlog item's evidence inventory. The authored parent register is corrected and generated views rebuilt. The check is unchanged; the failed workflow 36292138883 remains visible in GitHub. No source, reading-help output or public acceptance receipt changed.