Peer Info

Routing Model

Renegade ML Predictor

Renegade is a zero-configuration k-nearest-neighbours model (it auto-selects K and learns which features matter). It predicts per-peer, per-contract outcomes from four features (peer, contract location, distance, time); the router blends its prediction with the distance-based estimate (a weighted average, Renegade's share growing with data up to half) to catch patterns distance alone misses, such as a peer that drops requests for specific contracts.

Brier score: 0 = perfect · <0.05 excellent · <0.1 good · <0.15 fair · 0.25 = random coin flip

Outcomes vs Distance

Actual observed outcomes (dots) against ring distance to the contract, with the isotonic fit overlaid (the All tab is the aggregate the router uses; per-op tabs just break it down and are not consulted separately). "Peer-adjusted" applies this peer's running EWMA correction to that fit — a multiplicative factor for response time, an additive offset for failure and transfer rate. How tightly the dots hug a monotonic curve shows how well distance alone predicts the outcome. A separate Renegade model is blended into the final estimate; its accuracy is in the Prediction Accuracy panel below.

Actual outcomes Isotonic fit Peer-adjusted Peer location Extrapolated

Prediction Accuracy

How well the Renegade predictor's recent predictions matched reality (this scores the Renegade k-NN layer, not the distance-only fit in Outcomes vs Distance above). On the dashed diagonal predictions are perfect: for failure, predicted probability equals the observed failure rate (calibration); for the timing models, predicted equals actual.

Prediction at Peer Location