Choose a task small enough to judge.
A broad ambition such as “make all our knowledge intelligent” is difficult to test. A narrower question is more useful: can a reviewer locate the right procedure and the recording that explains it? Can a curator trace an exhibition statement to its source?
Start with a permitted sample rather than a complete archive. Choose material that includes a typical item, a difficult item, a misleading near-match and a question with no answer in the sample.
Write the success criteria first.
Before running the new workflow, record the sources an experienced person expects to find. Decide what makes a result relevant and what would make it unsafe to use. Agree who reviews and who can authorize reuse or further work.
Compare the entire task.
Time the current approach and the assisted approach. Count opening sources, checking context, correcting mistakes and resolving uncertainty. A quick answer that takes a long time to repair may not improve the work.
Do not invent a savings percentage from one appealing example. Keep the counts, task descriptions, limits and failures available to interpret the result.
Keep local data handling explicit.
Running on a local workstation avoids making a website upload the starting point, but it does not decide consent, access or retention. Transcripts, previews and indexes are also derived information. Agree where they live and what happens after the session.
Decide without momentum pressure.
The next step can be a second test, a narrower collection, a different tool or a pause. Continued development should follow a concrete need and evidence of usefulness, not the feeling that a demonstration must lead to a purchase.
Selected VAIMS evaluations may be discussed by agreement. They use a configured workstation and confirmed functions; no installation on production systems is needed to begin.
