Production AI Dataset RFP Checklist (2026)

Before you issue an RFP, define the failure modes your eval harness must catch: occlusion, sync drift, glare, or rare defect classes. Harbor programmes document capture briefs, reviewer consensus, and export slices aligned to those scenarios.

What to require in every response

1. Lineage — who captured data, under which programme, with which rubric version. 2. QA evidence — inter-annotator agreement, escalation rates, and filtered sample rates. 3. Eval alignment — held-out slices with scenario pass rates, not only aggregate accuracy on public benchmarks.

Bottom line

Use Harbor comparison pages and benchmark scorecards to align procurement, security, and ML on one rubric before you scale spend.