Build a representative task set
Choose real questions with verifiable answers and varied difficulty. Keep the input documents, date, expected sources and scoring rules fixed while comparing runs.
Measure several outcomes
- Correctness: does the final result answer the question?
- Evidence: do cited passages support the claims?
- Completion: did the requested task finish, or was it refused or unavailable?
- Cost and latency: what did the entire task consume, including retries and review?
Interpret the result
Keep local or synthetic checks separate from real customer journeys. Publish the conditions and unresolved cases beside any measured result rather than reducing everything to a single superiority claim.
A practical starting point
Research with traceable evidence. These steps describe the GoCux workflow and its permission boundaries.
- Define a research question, the project it belongs to, and the sources that may support the answer. Separate a factual question from the action you might take after reading the result.
- Review the proposed research plan before approving execution. Keep the approved scope stable; a different question deserves a revised plan rather than silent expansion.
- Read each finding alongside its citation and source excerpt. Treat missing evidence as an unresolved question, and keep the final report distinct from any subsequent external action.
What to verify
A useful report connects its conclusions to sources and makes unresolved questions visible. A fluent paragraph alone does not establish that a research task succeeded.
Bring this workflow into your workspace.
Set the scope. Review the evidence. Approve the steps that need it.