cvx-bayes — Bayesian Network Inference for ChronosVector

Provides discrete Bayesian networks for probabilistic reasoning over temporal episode data. Designed to answer questions that vector similarity alone cannot:

• P(success | task_type=clean, region=R5, action=navigate)
• "How confident am I in this prediction?" (posterior variance)
• "What's the most likely action given partial observations?"

Theoretical Foundation

A Bayesian network is a directed acyclic graph (DAG) where:

• Nodes are random variables (task_type, region, action, success)
• Edges encode conditional dependencies
• CPTs (Conditional Probability Tables) store P(X | parents(X))

Inference computes the posterior P(query | evidence) by propagating beliefs through the graph. For small networks (< 20 variables), exact inference via variable elimination is tractable.

References

• Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems
• Koller & Friedman (2009). Probabilistic Graphical Models
• Murphy, K. (2012). Machine Learning: A Probabilistic Perspective