Layer 1
Calibrated escalation
For uncertified actions, selective prediction and conformal risk control convert uncertainty into a review-rate SLA. The plots below are generated from real public datasets in this workspace.
Risk-coverage as product dashboard
Coverage is the auto-execution rate. Risk is the error rate on auto-executed cases. On the Wisconsin Diagnostic Breast Cancer dataset, a logistic-regression router auto-executes 195/200 held-out cases at 0.51% observed selective error and escalates the remaining 2.5%.
569real dataset cases
200held-out test cases
97.5%auto-execute at 1% target risk
0.51%observed selective error
Conformal review budget
On sklearn Digits, split conformal prediction sets turn uncertainty into an explicit review rule: auto-execute singleton sets, review non-singleton or empty sets.
