The original project citation used in this article is no longer available at its linked address. We could not verify the dependent claims from that source, so we have removed those claims and the broken link. What follows is our proposed implementation checklist, not a report of that project's results or a verified incident.
Define the work and its boundaries
Choose a bounded task and write down what a correct completion looks like. List the information and permissions required, the actions outside scope and the person responsible for escalation. For spending or destructive actions, agree an enforceable authorization rule before granting access.
Test observable behavior
Use representative examples, incomplete inputs, an unavailable dependency and a request that exceeds the permitted task. Preserve the expected result and inspect the actual system state. Do not rely solely on an agent saying it succeeded, or assign human emotions to explain an unexpected output.
Evaluate the whole configuration
Record the model, prompts, source information, tools and limits used in the test. Include human review, maintenance and error recovery in the cost comparison. For code changes, inspect the diff and run checks that demonstrate the intended behavior. Avoid attributing a result to model size or a framework unless the comparison actually isolates that variable.
Decide what the evidence supports
Keep unresolved failures in a review log. Agree which findings block expansion and what further evidence is needed. A successful bounded test supports a decision about that configuration; it is not proof of universal reliability or an estimate of a law firm's revenue impact.
Editorial correction: unsupported or unavailable-source claims were revised. Original publication date retained; this correction does not certify every other article in the archive.