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

Verify the household evidence service

Service 0.5.0 was deployed and checked on 20 September 2026. The retained observation binds the actual deployment, SDK comparison and three-browser care-home journey. The SDK passed seven full-package and four compact cases across four approved versions. All three browsers passed the evidence journey; Chrome and WebKit passed strict console checks, while Firefox retained two hosting-cookie warnings. Historical browser journeys were not run for this release. Earlier 0.4.0 observations remain separate and unchanged.

The source is 3ef0e786e9a18e76fa17c7d925ff509d6d6c9f84. It adds 20 selected statutory units while leaving all 203 obligations open. The broader development evaluation retains 176 of 177 expected candidate-page occurrences; all 40 staff tasks remain insufficient. These are evidence-discovery results, not an answer accuracy score. The Monday handover explains the demonstration in plain English.

The service's software development kit (SDK) check compares an external client's evidence with the shared context engine. The browser check then verifies that a person can inspect the same selected records, official links, provenance, relationships and gaps. Complete delivery does not mean a complete answer. The household package remains insufficient; no individual entitlement decision or specialist legal acceptance follows from these checks.

1. Retain the actual deployment and SDK checks

The completed release is frozen under validation/compact-delivery/v0.5.0/. For a new observation, choose a fresh directory, replacing YYYYMMDD in the examples below, and retain the new actual checks there. Never replace an old receipt or manifest.

Retain these files in validation/compact-delivery/v0.5.0-followup-YYYYMMDD/:

deployment.json: the actual hosting result, immutable runtime commit and Worker hash, using okf-compact-delivery-deployment.v1.
sdk-receipt.json: the successful real SDK verification, using okf-remote-mcp-sdk-verification.v1.

Do not substitute the service's offline integration receipt for either file. The SDK and deployment runtime identities must agree. Service 0.5.0's primary compact case is the supplied self-funded permanent care-home question, at 262,144 bytes; the actual public run selected 35 records and 50 relationships. The verifier uses the actual SDK receipt's identities and counts rather than hard-coding those observations. Source revision, context identifier and budget must remain identical across the SDK and browser checks.

The service's remote SDK verifier paces HTTP requests by at least 750 ms and does not automatically retry. Wait at least 60 seconds after it finishes before starting the browser run, to leave the previous service rate-limit window. Other clients may still share that limit. Retain any failure, including HTTP 429; do not silently repeat it.

2. Check the inputs and run the browsers

From the DWP repository root, first validate the new receipts without network access:

node scripts/check_staff_service_browser.mjs \
  --service-url https://ask-okf.crpage.chatgpt.site \
  --sdk-receipt validation/compact-delivery/v0.5.0-followup-YYYYMMDD/sdk-receipt.json \
  --deployment validation/compact-delivery/v0.5.0-followup-YYYYMMDD/deployment.json \
  --check-inputs

Then use the reviewed Explorer checkout with its installed Playwright browsers:

node scripts/check_staff_service_browser.mjs \
  --explorer-root /path/to/reviewed/okf-explorer \
  --service-url https://ask-okf.crpage.chatgpt.site \
  --sdk-receipt validation/compact-delivery/v0.5.0-followup-YYYYMMDD/sdk-receipt.json \
  --deployment validation/compact-delivery/v0.5.0-followup-YYYYMMDD/deployment.json \
  --output validation/compact-delivery/v0.5.0-followup-YYYYMMDD/browser/staff

The output directory must be new. Chrome, Firefox and WebKit run sequentially with shared 750 ms pacing. The observer reads browser-native response clones; it does not replace responses, intercept routes or weaken the security policy. Rendered text and the reconstructed full package are checked against the SDK hashes. Catalogue labels, source URLs, locators, authority and review status are checked against that same verified package.

Retain the exact executed scripts/check_staff_service_browser.mjs alongside the browser receipts as verifier-<SHA-256>.mjs, using the digest recorded in run-summary.json under binding.script.sha256. Retain screenshots and every receipt, including unsuccessful attempts. Record functional outcomes and strict console outcomes separately. A known hosting warning remains a failed strict console check; it is not converted to a clean-browser pass.

3. Declare the scope and bind the artefacts

Add a release README describing what actually ran. Once the SDK, deployment, browser receipts, exact harness and screenshots are present, create a manifest:

uv run --locked python scripts/check_staff_service_observations.py \
  --directory validation/compact-delivery/v0.5.0-followup-YYYYMMDD \
  --write-manifest --staff-attempt staff --historical-suite not_run

This command refuses to replace an existing manifest. It verifies the declared runs before writing. Repeat --staff-attempt for every retained attempt, for example --staff-attempt staff-first --staff-attempt staff-second; names are single safe directory names below browser/. An undeclared browser attempt, missing engine receipt or missing expected screenshot cannot pass. Retained failures remain failures.

Use --historical-suite present only after the separate four-journey, three-browser historical suite has actually run against this deployment and its public-safe projection has been retained under browser/historical/. Otherwise, not_run produces null historical pass counts and historical_strict_status: "not_run". A successful current staff run does not establish historical regression coverage.

The new manifest declares its layout explicitly under okf-compact-delivery-observation-layout.v1. Unknown schemas or fields, duplicate or unsafe names, symlinks, unbound files and altered bytes are rejected. Reads are bounded to 128 KiB per manifest, 10 MiB per member, 256 members and 32 MiB total. The checker makes no network request and launches no browser.

4. Repeat the offline integrity check in CI

uv run --locked python scripts/check_staff_service_observations.py \
  --directory validation/compact-delivery/v0.5.0
uv run --locked python -m unittest discover -s scripts \
  -p test_staff_service_observations.py

Verification reads the frozen layout from the manifest; command-line flags cannot override it. The existing command without --directory still checks the exact 0.4.0 layout and preserves its original outcomes. Add the explicit release-specific CI command when its actual receipts exist. The 0.5.0 command above now verifies its 14 retained artefacts: three functional browser passes, two strict console passes and historical status not_run.

ChatGPT tool invocation, complete receipt of evidence in that client, Voice access, answer quality and specialist review each need their own observations.