Skip to content
Public trust standard

How to test an AI job before it earns more work

A polished answer is not the test. A repeatable job under normal and ugly conditions is the test.

Published August 24, 2026 · Reviewed against current product boundaries

The twenty-case set

  • Ten normal recent cases.
  • Five edge cases the business has actually seen.
  • Three cases with important information missing.
  • Two deliberate traps: a conflicting instruction, wrong identity, bad date, or unsupported request.

Score what matters

Factual accuracy

Important names, dates, amounts, commitments, and records are correct.

Job completion

The output reaches the defined finish line or visibly escalates.

Control behavior

Covered actions pause, permissions stay inside scope, and exceptions reach a person.

Owner effort

Review and repair take less time than the current process.

Recovery

The job fails visibly and safely when a source or provider is unavailable.

The release decision

Pass the first job only when the normal cases are consistent, the deliberate traps stop or escalate correctly, and the owner can inspect the result. Expand one adjacent job at a time.

A product update, integration change, incident, material instruction change, or new data type should trigger retesting.