INSTRUMENT 08 · SEEDED-CASE PROTOCOL

How do you keep measuring after validation?

Inject known failure cases into the live queue at a controlled rate. The protocol protects the measurement and the people doing the work — like a mystery-shopper program or a fire drill: known-fake problems are slipped into real work so you can keep checking whether people (and the system) still catch them, without anyone knowing in advance which cases are planted.

Example: at a 3% injection rate on 120 items a week, about 4 planted cases go through per week, mixed in with real ones. If the system's catch rate on those planted cases quietly drops from 92% to 60%, that's an early warning of automation bias or drift — visible weeks before it would show up as a real missed defect.

Injection rate4%Weekly queue volume

Seeds per week: 5 · starter library: 20 cases · typical injection: 3–5%

WEEKLY READING TABLE

System catch rateReviewer catch rateMedian verification time (s)Rejection rate

ETHICS GATE · REQUIRED BEFORE EXPORT

Reviewers are told seeding happens.Results are never used for individual performance management.Quarantine is designed, tested and audited before injection.

LIVE DIAGNOSIS

Working; reduce the rate

Export blocked: affirm all three ethics conditions before creating an operational protocol.