What must be decided before any score is calculated?

Define the task, document families, approved outputs and operating boundary. A service-deficiency quotation and a detailed commercial tender should not share an undifferentiated score.

Agree:

• The same authorised input set and held-back cases for each candidate.
• A manual baseline covering preparation, review and corrections.
• Which errors are critical and who decides.
• The scope of price, labour, regulatory and engineering review.
• Whether the trial is offline, staged or connected.
• The workload used for cost comparisons.
• Who owns the decision and declares conflicts of interest.

The NIST generative AI profile provides risk-management context; the actual thresholds belong to the buyer.