EARCP Published on arXiv
Ensemble Auto-Regulative Coherence Protocol
EARCP (Ensemble Auto-Regulative Coherence Protocol) is our ensemble learning framework for dynamically weighting multiple expert models. The paper is now available on arXiv under identifier 2603.14651.
The core insight behind EARCP is that ensemble methods traditionally weight models based on historical performance alone. This is a lagging indicator. A model that performed well last week may be poorly calibrated for today's input distribution. EARCP addresses this by incorporating inter-model coherence as a second weighting signal.
How it works
At inference time, EARCP queries all expert models in the ensemble and computes two scores for each: a performance score derived from recent accuracy on held-out data, and a coherence score that measures how closely each model's output aligns with the ensemble consensus. The final prediction is a weighted combination of all expert outputs, with weights updated continuously as new data arrives.
This dual-signal approach makes the ensemble more robust to distribution shift and reduces the risk of a single poorly-calibrated model dominating the final prediction.
Open source
The reference implementation is available on PyPI. The paper includes full experimental details, ablation studies, and comparison against standard ensemble baselines.
