Head-to-Head Residual Odds

Grounded in BridgeMem: Causal Dyadic Transition Residuals for Temporal Knowledge Graph Forecasting (arXiv:2609.29268). The paper's claim: generic temporal-knowledge-graph forecasters mostly summarize each entity's own history and miss pair-specific transition evidence — how the prior relations between this specific pair shift the odds of the next event between them. BridgeMem retrieves the pair's own past events, encodes their relations & time-lags into a likelihood-ratio residual, and adds it on top of a frozen backbone forecaster's log-scores — with a support-adaptive empirical-Bayes component that trusts exact counts when the pair's history is rich and backs off toward learned/global priors when it's sparse (reported MRR gains of 0.0028–0.0216 over 9 prior methods across 5 benchmarks). Below: the "backbone" is a power-rating win probability; the "residual" is a real recency-weighted log-odds adjustment computed live from the synthetic head-to-head timeline you generate.