How the AI will view your agents — kept calibrated.
A score that drifts is worse than no score. TrustRank runs a weekly improvement loop against its own prediction error — and only promotes what provably got better.
The weekly loop
Measure
Brier score and expected calibration error (ECE), computed per stakes cohort — a score that is right on $40 deals but wrong on $1,500 deals gets caught.
Re-estimate
Weights and decay parameters are re-fit against the newest signed outcomes — candidates only, nothing ships yet.
Three statistical gates
Significant improvement · no cohort degradation · two-cycle persistence. A candidate config must clear all three before it is allowed anywhere near production.
Promote a versioned config
The winner ships as an immutable, versioned config — every score can name the exact config that produced it, and rollback is one pointer.
Real outcomes are the ground truth, and LLM-judge probes fill in where outcomes are sparse — together they keep the score aligned with how AI actually evaluates agents, not with how the model looked the day it shipped.