Independent experimental publication. Not an official DWP service, benefits advice or an entitlement calculator.

Multi-agent work log: 19 September 2026

User-facing changelog · Backlog · Retrospective

Branch: codex/evidence-review-navigation. State: dated observations; integration and validation status are linked below. This ledger records responsibilities, decisions and handovers. The changelog summarises changes useful to readers; it is not a dump of agent messages or Git commits. A completed subtask is not a merged or deployed release.

Work ownership and dependencies

Task ID	Owner role / workstream	Owned files or interface	Dependencies	Current handover
DWP-WL-001	Integration / root	Reusable evidence-delivery contract, remote service and human review route	Existing immutable context identity; DWP-WL-002/003 interfaces	Public service 0.3.1 and exact SDK acceptance recorded; v7 historical and full-corpus functional checks passed, strict console gates failed; v6 preserved (DWP-BL-023)
DWP-WL-002	Corpus semantics review	`domain-profile/navigation/`, navigation producer and staff-review packs; narrow additive contracts	Frozen DMG/ADM bytes and existing concept IRIs	Producer checks and exact local/public DMG Reader observations recorded; ADM Reader remains BL024
DWP-WL-003	Explorer navigation	Generic conceptual navigation and full-text human review consumer	DWP-WL-002 manifest; reusable consumer contracts	Exact local and published Chrome navigation journeys passed; service hosting-console failure remains separate
DWP-WL-004	Methodology and evaluation	Methodology/template, ontology map, backlog, retrospective, answer-trial harness; README/CHANGELOG integration	Actual namespace audit from DWP-WL-002; fixed public contexts; existing Claude subscription	Methodology/backlog handed over; fixed trials and actual local compact client replayed; strict failures preserved; human review pending
DWP-WL-005	Integration and independent review	Exact candidate review, CI, PR, release and public observations	All affected workstreams, current documentation and checks	Explorer PR 125 merged, Pages and public DMG navigation passed; DWP PR 10 records integration and exact check/merge status

Public artefacts only. Private correspondence, existing research scratch files, credentials, account identifiers and local browser state are excluded.

Decisions and rationale

Decision	Rationale and limit
Preserve `ask_okf`; add `ask_okf_manifest` and `read_okf_evidence`	Existing clients retain their full-package behaviour. Smaller delivery is additive, not a change to evidence authority.
Recreate approved context for each evidence read	Question, version and original budget must yield the requested context ID before any record is returned; no stored question session or arbitrary source URL fetch.
Exact text slices with offsets, digests and continuation	A consumer can reconstruct evidence without silent summaries. Offsets are UTF-16 units as defined by the shared implementation.
Human review link carries a replay recipe in the URL fragment	The HTTP request does not carry the fragment as a query. The page asks the user to recreate the evidence; it does not reproduce a stochastic AI answer.
Formal MCP resource reading deferred	Prove the read-only tool interface and compatible HTTPS review link first; do not claim untested client support.
New navigation is additive	Preserve frozen source and old semantic builds. Topic assignment is literal discovery, not `appliesTo` or a legal conclusion.
Claude trial is separate from engineering evaluation	Exact citation checks, legal entailment and human usefulness are different judgements. No model output becomes gold by generating it.

Checks actually recorded so far

Claude Code help was inspected before use. An isolated, tools-disabled public probe failed authentication in the sandbox and succeeded at the authorised normal host boundary. No login, installation or model override was performed.
The fixed-package harness has six passing controls for source identity, invented citations, altered quotes/locators, false sufficiency and no-evidence abstention. These test the checker, not the correctness of any model answer.
Actual baseline trials use three retained public packages and save their exact input/prompt/output hashes. Custody produced 12 claims/31 exact citations; the no-result control abstained with no claims. Abroad produced 10 claims/14 citations, with 10 strict quotation failures diagnosed as whitespace-only changes and four exact quotations. The strict failures remain recorded. Results are recorded in the answer-review directory; independent human review remains pending.
Namespace use was checked against the retained semantic shards and pinned context; see ontology use. Declared ELI/OWL prefixes were not promoted to implemented ontology claims.
The reusable service's compact-delivery verifier has eight passing offline controls: exact reconstruction and seven corruption rejections. It preserves the earlier complete-package comparisons and adds schema, annotation, catalogue, provenance, slice-hash, byte-bound and fail-closed checks. A live receipt is now recorded for the deployed 0.3.0 service.
The public backlog checker has eight passing fixture controls, including dependency cycles, private/escaping paths and a misleading prose status. CI now replays the three retained answer trials and their quotation diagnosis offline, without model calls or silently turning strict failures into passes.
An actual Claude local MCP retry made seven verified compact-tool calls and read five complete values, after a zero-call safe-mode attempt generated an invented catalogue. Both observations are retained in the compact-client report. Five offline controls preserve that failure and reject corrupted source/status/size values. No public deployment or model-answer acceptance is inferred.
The navigation browser observation records the separate exact local candidate's conceptual facets and journeys; it does not replace the dated public deployment receipts.

Later owners should append concrete test commands/results and receipt links, not change “in progress” to “passed” from implementation alone. Preserve failed attempts. Before handover, state modified files, exact inputs/consumer revision, tests run, remaining checks and any active process. Before publication, reconcile this ledger, the backlog and the relevant changelog entry with the actual candidate.

Public service handover

Hosting and SDK verification bind runtime 169b8c387a29435d39dc31cbb2066376d84b39a6. The SDK completed at 20:21:55 UTC on 19 September. It retained all three original full-package comparisons and verified the smaller catalogue/read path, full reconstruction and three rejected invalid reads. The five-minute guide keeps this public-service result separate from local navigation, local Claude, the earlier ChatGPT call and any later public-browser observation.

Public browser gate: hosting conflict retained

The public run summary records the historical custody profile across Chrome, Firefox and WebKit. All functional assertions in the nine cases passed before their final console check. All nine overall tests failed because the strict no-console gate reported the host's injected Cloudflare challenge loader, blocked by the unchanged CSP; Firefox also reported invalid-domain __cf_bm cookies. The earlier local 9/9 result remains separate. That suite does not establish full-corpus browser behaviour or AI-answer quality.

A separate public full-corpus Chrome observation at 20:38 UTC made four real tool calls: catalogue, exact source text, provenance and diagnostics. It displayed six records and the insufficient 2cdfa5fe… abroad context; rendered text and metadata for ADM C2 page 18 and complete diagnostic hashes matched the SDK. Functional checks passed, while the strict console gate retained the host CSP error. This is hosting version 6, runtime 169b8c387a29435d39dc31cbb2066376d84b39a6, before the 0.3.1 replay-link correction. No other browser engine or AI answer is covered by this observation.

DWP-BL-023 assigns a bounded hosting/CSP integration follow-up. Its acceptance requires supported hosting changes, preserved CSP and controls, and fresh exact functional and console observations. No inline-script allowance, console-error suppression or hosting-control bypass was introduced. SDK integrity and the local Claude evidence observation remain separately recorded.

Final review correction and service 0.3.1

Independent review found that changing or resubmitting the question could leave an earlier replay link visible. The correction clears that link on question or source changes and on resubmission; twelve local journeys across Chrome, Firefox and WebKit passed. The 0.3.1 deployment at 20:47:55 UTC and SDK receipt completed at 20:48:50 UTC bind runtime 8493b323ca664e645a2548ebb48bf7917d7f6eb1 and hosting version 7. The SDK again verified all three unchanged full packages and compact reconstruction. Earlier v6 browser failures remain preserved.

The v7 historical-profile suite then exercised four journeys in each of Chrome, Firefox and WebKit, including question A → B and the replacement replay-link identity. All twelve functional sequences passed before the final strict console assertion; all twelve overall tests failed because the host CSP/cookie errors remain. A separate v7 full-corpus Chrome journey made four real calls and matched displayed source text, provenance and complete diagnostic hashes to the SDK's six-record insufficient abroad context. Its strict console check also failed. DWP-BL-023 remains open; no AI-answer or clean public-browser acceptance is claimed.

The final assembled-site check also exposed a documentation-cache defect: transitive linked service documents were rendered but absent from the cache fingerprint. The narrow build-script correction includes the existing bounded reading closure. Forty focused build/cache tests and the complete site build passed, including 340 reading pages and 7,449 internal references. Source evidence, service runtime and the application build were unchanged by that fix.

Integration checkpoint

Explorer PR 125 merged as 8a38d4bfe07a6797deacc148a2859911d8e2a68e after the required CI run passed: 552 app tests, 493 Python tests, browser suites of 348 and 78 tests, 37 service tests and 12 local service-browser journeys. The Pages deployment then completed successfully at 21:15:34 UTC. The separate public Chrome navigation observation passed at 21:16:55 UTC against that merge and DWP corpus commit 0c59f476602a49d4293e46112a0ca1ebf1486ff0. It verified 21 application files and 188 observed corpus files over actual HTTPS, with no response substitution. Benefit, circumstance, topic and authored-concept selections matched counts 74/343/268/1 across Reader, Graph and Timeline; capture/source roles and the 80-group Timeline limit were visible. No console errors or targeted axe violations were found. Whole-application accessibility and ADM Reader coverage are not established by this DMG observation.

DWP PR 10 records integration of the evidence and documentation; consult the PR for its exact validation and merge status. The observed model trials await specialist review. All 40 staff-question occurrences and all 43 full-corpus evaluation cases remain insufficient; this integration checkpoint does not establish complete evidence profiles or benefit decisions.