Label the origin transparently
Separate operational calls, instructed roleplay, scripted readings, and synthetic conversations. Describe the instructions and scenarios used for staged material when available. A natural-sounding recording is not sufficient evidence of authentic business interaction, and buyers should not have to infer origin from the way a speaker performs.
Match origin to the task
Roleplay can exercise specific scenarios or rare conversational branches, but it may not reproduce authentic pacing, uncertainty, and business context. State why each subset is useful. If deployment concerns real calls, avoid presenting performance on staged conversations as proof of performance across operational telephone traffic.
Keep evaluation slices distinct
Report results by origin, domain, speaker grouping, and recording setup. Prevent repeated scripts or scenario variants from appearing on both sides of an evaluation split when that would undermine independence. Inspect whether a model can identify recurring scenario wording instead of understanding the intended conversational behavior.
Discuss a documented mix
Tell Fiund whether authentic calls are required and where staged supplements are acceptable. Request origin metadata, scenario descriptions, and source-review documentation for the particular project. A representative pilot should expose the limits of each subset so decisions about broader sourcing are based on fit rather than an undifferentiated hour total. If a staged scenario asks every participant to repeat the same cancellation request, compare its wording with authentic examples where available and reviewed. Record the difference as a coverage limitation, not a defect to hide. The point is to understand what behavior each subset can test and where the apparent diversity comes from.
Review checklist
- Label origin for each subset
- Explain the purpose of staged examples
- Separate evaluation results by origin
- Review script and speaker overlap across splits