Objective
Preserve prompt, system, date, output, citations and repeated runs. The controlled asset is the prompt panel. Completion means the firm can inspect the work, reproduce the check and continue without depending on an unexportable vendor view.
When to use this procedure
Run this SOP when model, prompt wording and account context needs a new baseline, a correction or a scheduled review. Do not begin because a dashboard has issued a generic warning. Write the decision the procedure is meant to support. That keeps the work bounded and gives the quality gate a real pass condition.
Assign one operator and one approver for the prompt panel. The same person may fill both roles for a small firm, but the worksheet should still separate the act from the approval. Attorney review is required whenever published text could state law, characterize a charge, describe a defense or expose a client fact.
Inputs
Procedure
- Open the control record
Create the prompt panel under firm ownership. Write the market, practice scope, responsible person and review date before changing anything.
- Capture the starting state
Inspect model, prompt wording and account context. Save the raw export or source copy and note any field that cannot be observed.
- Set the acceptance rule
Define what counts as complete, which value must remain fixed, and who can approve an exception. Put the rule in the worksheet.
- Perform the work
Run repeated dated observations without editing prompts. Make one traceable change at a time when the system allows it, then log who acted and when.
- Test the result
Repeat the observation with the same settings. Compare the output with the acceptance rule and retain output captures and citation list.
- Close or recover
If the gate passes, prepare the handoff. If it fails, restore the last known state where possible and open a blocker record.
Asset-specific record
For AI answer testing SOP, the first inspection is model, prompt wording and account context. Save that state before acting and name the person who can confirm it. The working action is to run repeated dated observations without editing prompts. Put that sentence in the task record so a later operator can distinguish the intended change from incidental edits made during the same session.
The proof file is output captures and citation list. Give it the procedure number and observation date, then store it with the source material used for the decision. If the platform has no export, capture the complete visible settings and account context. A screenshot can show a state; it cannot replace an available raw file used for a calculation.
Asset closeout
The deliverable is answer-test log. Review it against the starting state and acceptance rule before transfer. The firm should receive an editable or exportable copy under its control, along with the next review date. Record any dependency that remains open and the person responsible for resolving it.
Stop the closeout when the model or prompt changes without a new baseline. Preserve the failed state and write the expected result beside the actual one. If a repair changes the observation method, close the old version as inconclusive and open a dated procedure. That keeps the original baseline intelligible.
Quality gate for AI answer testing SOP
Approve the prompt panel only when a second operator can reproduce model, prompt wording and account context and reach the same disposition. The reviewer must locate output captures and citation list, identify who approved it and state the limit that travels with the result. Return the AI answer testing SOP file if a conclusion depends on an unrecorded setting or a vendor-only screen.
Scenario test for AI answer testing SOP
Run a dry test on the prompt panel before declaring this procedure ready. Begin with model, prompt wording and account context and give the test file to someone who did not perform the setup. That reviewer should be able to locate the starting record, identify its date and explain which field the procedure may change. If those answers depend on a verbal explanation from the operator, the AI answer testing SOP package is not ready for repeat use.
Next, ask the reviewer to follow this working instruction: run repeated dated observations without editing prompts. Compare the resulting observation with output captures and citation list. A useful comparison names the unchanged settings as well as the changed field. It also records an absent value as absent; it does not fill a blank from memory or copy a later value back into the baseline. File the review beside the prompt panel so the test travels with the work.
Acceptance example
A passing AI answer testing SOP file produces answer-test log that another firm-authorized operator can open and continue. The package identifies the approval rule, links to output captures and citation list and gives the next review a calendar date. The receiving operator can tell which source controls the fact and which worksheet contains the observation. No private case detail is needed to understand the result.
A failed example begins when the model or prompt changes without a new baseline. Leave that state visible in the prompt panel record. Assign the repair to the person who controls the source, then repeat model, prompt wording and account context under the recorded settings. Do not replace the failed copy with the repaired one. The two files show what changed, and they let the approver decide whether the original acceptance rule still fits the AI answer testing SOP procedure.