Epistemic Accuracy

Epistemic Accuracy is the measure of how close a belief or credence function is to the truth. Unlike Coherence (Probability), which is a purely structural constraint on the internal consistency of beliefs, epistemic accuracy is a correspondence constraint: it evaluates beliefs against the way the world actually is. A belief can be perfectly coherent yet entirely inaccurate, and a set of beliefs can be accurate yet incoherent — though the latter is harder to sustain systematically.

The epistemic accuracy framework was formalized by Joyce, Greaves, and Wallace, who showed that the probability axioms can be derived not from Dutch book arguments but from a accuracy-dominance principle: any incoherent credence function is accuracy-dominated by some coherent one, meaning there exists a coherent function that is closer to the truth in every possible world. This is a powerful result: it means that coherence is not merely a constraint on betting behavior but a constraint on truth-seeking itself.

However, the accuracy framework has its own difficulties. It requires a measure of "distance from truth" — typically the Brier score or a similar proper scoring rule — and the choice of scoring rule is not innocent. Different scoring rules privilege different epistemic virtues: precision, calibration, or discrimination. The relationship between accuracy and pragmatic utility is also contested: a belief can be accurate but useless, or useful but inaccurate.

Epistemic accuracy is now a central concept in formal epistemology, the philosophy of science, and the foundations of Bayesian inference. It provides a bridge between the subjective Bayesian tradition — which emphasizes coherence — and the realist tradition, which emphasizes truth. It also raises a difficult question for Artificial Intelligence: if an AI system's beliefs are coherent but systematically inaccurate, is the system rational? The answer depends on whether rationality is a property of internal structure or of correspondence to reality.

Epistemic accuracy is the missing link between Bayesian epistemology and scientific realism. Without it, coherence is just a beautiful cage. With it, we can ask the hard question: not just "are my beliefs consistent?" but "are my beliefs close to the truth?" And that second question is the one that matters.