Reading-help release record — 27 September 2026

Status: technically verified; specialist review remains open. This is an independent experimental publication, not official DWP guidance or an entitlement decision. Explorer PR 159 merged as b633038c6173f1405199df9f08e8d8aa2ff51c0b; its exact public Pages deployment and all 16 public browser checks passed. Earlier failed attempts are retained. The final demonstration tag and its release assets bind the subsequent publication of these release notes.

What is implemented

• The deterministic corpus accounts for 513 DMG and ADM documents, 19,090 pages, 53,727 adopted structured units and 893 pages without usable extracted text. It publishes 54,299 bounded passage segments in 1,455 compressed leaves, with 197,564 delivered occurrence candidates. Processing coverage records gaps; it does not mean complete explanations. The later 54,577-unit amendment candidate remains separate.
• The catalogue has SHA-256 6c53b659b519aa4cbf7529e7e10eb4ad3bfe8efcac18055ec2a5ade724998b46 at data commit cbada7bf544106bdceb2445fc379c8b3e5b73ee4; corpus PR 50 merged as 239f974c6cba23a2a37182af722ebba0803504b6. The additive workbench integration is recorded at 2ad33add6d8660fec98bf34fb98480a9d0de9e27 and DWP PR 49 merged as f61d41b689427c27275ae3b7e64b612925b6321e; seven public documentation resources matched that exact Pages artefact.
• Exact unit and source identity links cover 401 of 403 distinct retained workbench records across all 40 cases. The two unmatched capital composites, U06 and U10, remain unlinked. The workbench does not rewrite their packages or evidence requirements.
• The original Chapter 60 v1 reading aid remains unchanged at fallback tag demo-2026-09-30-reading-help-v1 (pinned manifest commit 46a85142e0fe3a73b686b91029d357770b68591d). Separate v1-compatible manifests add 20 body-to-footer occurrence pairs and dated candidate statutory links.

Checks and retained results

• A source-led 12-case gate and a separately frozen fresh 24-case gate passed their stated structural controls. The first held-out attempt remains 23 passed and one failed: H05 expected only DMG 070831 while the frozen source prints a range to 070834, which the producer correctly retained unresolved. A further pre-execution fixture was rejected over et seq and coverage. Later producer regressions replayed unchanged expectations; none is specialist legal review.
• The full consumer compatibility audit found zero shape/size violations across 513 indexes and 1,455 leaves, including the printed 166-abbreviation table. The integrated focused reading-help suite recorded 50 passing controls. The 28-case passage review recheck reproduces all 55 files and passes four controls. The nine previously reviewed substantive chapters retain exact PDF/extraction hashes and every recorded source span. This is a non-regression check of the retained review; case 014 remains unresolved.
• Cold and warm equal-budget replay passed 40/40 package-integrity checks, with all 40 still insufficient and exact retained packages unchanged. The fixed limits were 512 KiB, 64 records, 128 relationships and depth six. Cold fetches totalled 158,837,842 bytes; warm fetches totalled zero bytes. This is a non-regression result, not improved retrieval or answer quality. There were zero content-processing or answer-model calls for this increment.
• The optional legal bridge reuses the existing 20 statutory units, 16 provisions and 43 navigation relationships. It requested three new provision targets and retained four observed bodies because section 70 has distinct England and Wales and Scotland variants. Four source-span-bound full-title mappings and 16 citation bindings passed structural admission; five unsupported citation cards remain unresolved. The optional context adds 10 records and 51 relationships while preserving all 40 original evidence requirements. The unchanged ordinary natural-language probe still misses the bridge at its fixed budget; it is not the default service.

Boundaries and remaining work

The catalogue is machine-proposed and unreviewed. The 893 extraction gaps are not blank-page findings. Mixed-case abbreviations such as WDisP may remain unrecognised, and repeated-phrase proposals are capped at 100 per document. A source-verified literal proves its location and bytes, not the correctness of an expansion, a citation or legal applicability. No new statutory request, OCR, content model or answer model is needed to reproduce these deterministic artefacts.

The public browser receipt binds the exact consumer and frozen data. The previous Chapter 60 fallback tag and every failed control are preserved. Specialist review of candidate meanings, territorial and date variants, unresolved references and case 014 remains open. The separate legislation format migration is not release-ready: current release contracts bind historical consumer versions, and locked reproduction found a documentation ownership defect. Its candidate remains independent; the DWP demonstration uses only the validated bounded citation bridge.

Evidence pointers

• reading-help-corpus/manifest.json — catalogue and processing ledger.
• evaluation/reading-help-rollout/heldout-gate-02.json and heldout-03-producer-replay-03.json — retained failure and current unchanged-expectation controls.
• evaluation/reading-help-rollout/consumer-compatibility-01.json — complete leaf/index contract audit.
• evaluation/reading-help-rollout/replays/reading-help-20260927-01/report.json — fixed-budget cold/warm non-regression.
• evaluation/reading-help-rollout/workbench-link-report.json — 401/403 exact target mapping.
• domain-profile/reading-help-law/admission.json and evaluation/reading-help-rollout/law-context-probe.json — optional legal admission and failed ordinary selection.
• Release manifest — exact hashes, observed identities and explicit remaining gates.

Verify the retained release evidence

Run this offline check from the repository root:

uv run --locked python scripts/check_reading_help_release.py \
  --root . \
  --manifest evaluation/reading-help-rollout/release-20260927/manifest.json

CI runs the same check. It verifies the bound artefacts, the separate 12/24 controls, all 40 unchanged insufficient packages and the exact public verifier. A candidate can be structurally valid while public verification remains false. A technically verified status requires the complete passing public receipt, matching consumer and data identities, and the exact verifier script hash. These checks do not establish legal completeness or specialist acceptance.

Final publication identity

The intended paired tag is demo-2026-09-30-reading-help-v2. After the normal release-record PR and Pages checks pass, its GitHub release assets retain the final documentation observation and paired commit identities. This avoids an impossible self-reference in which committing a receipt changes the commit being observed. Until the tag exists, use the pinned links in the walkthrough; the earlier demo-2026-09-30-reading-help-v1 remains the tested fallback.