Computer-work lab: a supplier decision you can verify

Question: can a worker use a screen to produce a useful result that a separate reviewer can check?

This lab opens an isolated browser, compares three fictional supplier quotes and writes a recommendation. The browser interactions, HTTP adapter, dispatch journal, artifact hashes and acceptance checks are real. Worker decisions and the OpenClaw-shaped server response are deterministic fixtures; no live model, provider, wallet, order or settlement is involved.

Run in a few minutes

From the repository root, use the pinned Node setup, install dependencies, then:

npm run build:orchestrator
npx playwright install chromium
npm run demo:computer-work

Linux hosts missing browser system libraries can use npx playwright install --with-deps chromium in an environment where installing those libraries is permitted. The lab listens only on an ephemeral loopback port; it does not need a fixed port or API key.

Expected: accepted: true, passedChecks: 11, totalChecks: 11, browserExecuted: true, liveProvider: false, and settlementApproved: false. The output includes the exact evidence directory under reports/computer-work/<unique-run-id>/.

Stellar is the correct recommendation: Laurentian's cheaper quote misses the deadline. The arithmetic uses integer cents.

Follow the evidence

1. Open desktop.png or mobile.png to see the actual browser after interaction. The worker selects 7 days, clicks Compare quotes, and reads the displayed results.
2. Read comparison.json and recommendation.md. These are the actual text deliverables returned through the Responses adapter.
3. Inspect receipt.json: exact task hash, unique attempt ID, provider response ID, artifact sizes/hashes and required-review status. The synthetic mode comes from operator configuration, not the worker's claims.
4. Open review.json. A separate function recomputes totals and eligibility from quotes.json, checks artifact integrity, and verifies the recommendation. Its eleven checks are arithmetic/structural acceptance rules, not an independent security audit.
5. Inspect browser-evidence.json and accessibility.json. They record the actual interaction sequence, screenshot hash, one provider-shaped dispatch, browser errors and Axe findings. The persistent journal/ prevents a second dispatch for the same admitted job.

Make it fail meaningfully

npm run demo:computer-work -- --inject-error

The worker now recommends Laurentian. The browser still runs, the output is still well-formed and its hashes are valid, but the independent recommendation checks fail: 9 of 11 checks pass, accepted: false, exit status 1. This is the intended result. Compare the two unique report directories; the earlier evidence is retained.

The reviewer uses explicit fixture rules. For a real task, implement acceptance criteria that check authoritative sources or the actual application state; do not merely ask the same model whether it did a good job.

flowchart TD
    Task["Admitted synthetic task"] --> Browser["Select delivery limit and compare"]
    Browser --> Files["Comparison and recommendation"]
    Files --> Review["Recompute from source quotes"]
    Review -->|Correct winner| Accept["Accept fixture evidence"]
    Review -->|Late supplier| Reject["Reject and preserve evidence"]
    Accept --> Boundary["No purchase or settlement"]
    Reject --> Boundary

Move to your own work

Read Computer work: setup, admission, recovery and production boundaries. task.json and worker-profiles.example.json are reviewable templates. The live command is explicit and requires operator admission plus a protected token; the default lab never turns itself into a live run.

The broader One-Box mission demonstrates posting, assignment, submission, review, corrections, disputes and finalization in an offline browser model. This lab adds actual screen interaction and a provider-shaped execution boundary. Neither model claims a production deployment has been commissioned.

Troubleshooting