Use a compact evidence ledger rather than a feature popularity contest. The following are proposed measurement definitions; set the acceptance thresholds for your own recruitment task before testing.
| Measure | Suggested recording method | Guardrail |
|---|---|---|
| Additional supported answers | Record the recruitment question, incumbent answer, supplier contribution and supporting evidence | Distinguish new evidence from reinterpretation, confirmation and faster retrieval |
| Usable coverage | Count distinct usable matches against expected completed matches within a frozen scope and date | Report missing video, missing fields and late delivery separately; do not publish a percentage without a denominator |
| Time to an acceptable answer | Record active staff time and elapsed time, including verification and reconciliation | Separate onboarding from repeated use; log assistance and trial order |
| Identity quality | Record wrong matches and unresolved matches separately against reviewed records | Do not treat a plausible name match as verified identity |
| Traceability | Count checked material claims with supporting records and list unsupported claims | A linked source must support the specific statement, not merely mention the player |
| Operational burden | Log manual steps, support requests, exceptions and recurring maintenance | Do not convert every hour released into a cash saving |
Keep results separate for materially different competitions, evidence types and workflows. Do not let strong results in one tested setting conceal an untested priority market.
Label findings verified within this sample, partial, supplier-claimed, unknown or failed. A small diagnostic trial should not become a platform-wide completeness claim.