Move from the sandbox to a real test
What to change before a practice run becomes a real small test: a fresh account check, a new signed quote and your approval.
Establish real authority in a fresh context
Synthetic sandbox · No live actions
Carry the task brief, permitted source and acceptance rule into a fresh live context. Establish the real account, input, quote and approval before you submit:
task brief
→ real account and permitted input
→ fresh signed quote
→ transfer permission and paid-work approval
→ one supported submission
→ actual result and usage read-backThe sandbox helps you learn what to inspect. It does not establish account access, available capacity, correct model output or a price. Do not replace mock IDs in the script.
Keep the outcome and review rule. Discard the sample quote, receipt, states, and identifiers as live evidence.
For example, carry forward “extract the invoice total with a page citation,” the permitted source and the acceptance rule, then create a fresh input with new IDs. The sandbox's six accepted labels do not authorize that request or predict its result.
Establish the right account context
For first access, choose Apply for Compute access. Applying requests a fit review. It does not issue credentials, connect an agent, reserve work, or create a paid job.
After access is approved, create or resume the authorized account through its supported setup path. Existing team members should use their assigned account rather than create another to bypass a restriction. No provider enrollment is required.
Confirm the account, permitted operation, client version, data policy, and the person or policy that may approve paid work. Keep secret values in the approved protected store. The documentation exercise does not configure that store.
Prepare a small real input
Use a tiny input you have permission to process and can inspect fully. Confirm the chosen operation's actual format and constraints. A successful synthetic replay is not evidence that private customer data is suitable.
For a review-classification test, preserve stable IDs and define how missing or ambiguous text should be reported. Carry the frozen instructions, rubric, output schema and required evidence into the live brief. Lithi owns partitioning and placement; a file is not an execution shard. Your job is to state the semantic constraints and review criteria.
The worksheet is not a wire request. Use the current operation schema and approved client. If a required field or option is unclear, resolve it before quoting instead of inventing a value.
Review the live quote separately
Request a fresh signed quote for the exact input, output, policy and deadline. Check its maximum charge, expiry and included verification. A small test is not automatically free, and the sample quote supplies no commercial terms.
Approve the matching paid work separately, or confirm that an existing bounded approval policy covers it. Setup approval and quote creation do not authorize execution. If the input or policy changes, revisit the quote and approval.
The free estimate remains only a planning aid. Review the signed quote's exact scope and maximum charge before granting transfer permission for the prepared input and approving the paid work separately.
Submit once through the supported path and retain the job and request references. If the response is lost, inspect the existing outcome before starting another request.
Verify the actual terminal outcome
Read exceptions and missing items before sampling accepted output. Match every output to its input ID, check the returned schema, coverage and uncertainty against the brief, verify the supplied integrity and verification evidence, and compare actual usage with the quote. Integrity evidence does not decide semantic correctness.
A real result may differ from the teaching fixture. Record what actually happened; do not copy the fixture's six accepted labels into a report. Keep its receipt separate from delivery confirmation, invoices, and payment status.
If delivery fails after output is accepted, recover that delivery rather than repeat the computation. The first-test guide brings these checks together before you expand the workload.