Engine-pinned replay: local candidate evidence

Observed on 21 September 2026. Not deployed. Package/service version remains 0.5.0 and the four approved source revisions are unchanged. These receipts do not attest a new public service, model answer, current law or specialist acceptance.

What passed

• Actual integration: eight combinations of four approved immutable DWP versions and two frozen assemblers; 50 source files read from exact Git blobs and hash-verified. Both SDK generations validate the catalogue envelope. Full tool results, ordered catalogue pages and reconstructed complete package slices match direct assembly.
• Every preceding-engine package matches the complete canonical SHA-256 in the unchanged original 0.5.0 receipt. All four engine-unspecified historical replays succeed within the shared limits and retain unknown originating-engine status.
• Actual local Chrome: Chrome 153.0.8010.52, eight tool calls, frozen custody profile. Checks include inert link opening, keyboard submission, explicit engine/context identity on pagination, exact displayed source text, machine-package view, same-tab fragment changes, and visible historical-unavailable/pinned-mismatch refusal without replacement evidence.
• Unit checks cover the two-attempt limit, full-byte ambiguity, incompatible source/engine pairs, shared resource and deadline bounds, exact 8 KiB Unicode delivery, build-input admission and engine manifests. Runtime/build source and observation helpers are hash-bound; fresh output directories are required.

No corpus network, public-service or model calls were made. The local browser uses the actual service handler with a vendored historical profile; it does not claim full-corpus browser coverage, a screen-reader evaluation or ChatGPT/Voice acceptance. The source files and questions used here were already approved public examples; anonymous user questions are not stored by the service.

Exact candidate

Worker SHA-256: 0720239e4657fba2fcaa037a2fdfb6fa7ba16f4adcc9c92db4bb99cfadcff706.

Both observation directories retain their exact build receipt. The current assembler is frozen from c4f2de0a99b7bc2f8b8c8a06a3c715fb56b66d8e; the previous one is frozen from b9a3b68b6dbf222f9a73cc8f450dd53f126e1b55. Manifest hashes identify implementations independently of their human-friendly labels.

The care-home source/budget pair reproduces both outcomes:

Changing the assembler changes selection; preserving historical evidence does not establish that either result is sufficient. The original custody profile retains its bounded historical sufficient result, without extending that finding to current law or the wider corpus.

Measured build increment

Local measurements, using gzip level 9:

The Worker grows by 72,008 raw bytes and 16,821 locally compressed bytes. These measurements include the replay implementation and Reader changes; they are not hosted compression, memory, cold-start or latency measurements. The earlier Worker and Node digests match the retained required-evidence candidate build.

Reproduce safely

From the service directory, use a locked installation and a new output directory:

npm ci --ignore-scripts --no-audit --no-fund
npm run check
npm test
node --experimental-strip-types scripts/verify-versioned-replay.ts \
  --dwp-root /path/to/okf-dwp --out /path/to/new-integration
node --experimental-strip-types scripts/verify-versioned-review.ts \
  --playwright-module /path/to/locked/@playwright/test/index.mjs \
  --out /path/to/new-browser-observation

The browser harness requires installed Chrome and permission to listen on loopback. It does not download browsers or dependencies. Its temporary local server closes after the check. Do not reuse a retained observation directory.

Preserved earlier attempts

The history directories preserve actual earlier files, separately labelled by retrospective classifications:

• Browser attempt 01: sandbox refused loopback listening before Chrome launched. Its early harness did not produce a failure receipt; the classification states this limitation and preserves the executed script.
• Browser attempt 02: real Chrome exposed the pre-existing same-tab fragment issue. The preceding recipe remained active and the negative check timed out. The fix now clears old evidence and reparses the new fragment without a call.
• Browser attempt 03: preliminary pass before the final observation-input admission/provenance improvements.
• Integration attempt 01: preliminary eight-pair pass before the independently identified shared-deadline guard was corrected.

All published 0.4.0/0.5.0 and required-evidence receipts remain byte-for-byte unchanged. A new observation must never relabel those records as this build.