Issue a common response schedule with rows for data curation, labels, baseline optimisation, adaptation runs, evaluation, serving integration, security review, monitoring, on-call support, retraining and transition. For each row, ask “included price / usage assumption / buyer owner / supplier owner / excluded work / acceptance evidence”. Keep unknown prices visible.
Example, explicitly hypothetical: Proposal A quotes $30,000 for training; Proposal B quotes $55,000 for training plus integration and evaluation. It is invalid to call A cheaper overall until the missing work is costed. Ask A to price the same scope, or add a transparent buyer estimate with a separate confidence label. Do not silently fill exclusions with zero.
Score only after hard gates. A stronger average cannot compensate for unavailable data rights, a failed critical quality requirement or no acceptable exit. Suggested weighted criteria appear in scorecard.md; weights are a purchasing policy, not a research finding.