Work

Contextual Deviation Information: A Normative Information-Theoretic Framework for Protocol Conformance, Channel-Aware Monitoring, and Semantic Reliability

Year: 2026
Type: preprint
Abstract: Information-theoretic surprise, protocol conformance, and semantic reliability are related but non-identical properties of communication. A message can be statistically common and still violate a rule; it can be procedurally admissible and nevertheless misrepresent the state it reports; and a correct action can arrive corrupted by a physical channel. This article develops Contextual Deviation Information (CDI) as a framework for keeping these questions separate while connecting them through a common probabilistic formalism. At the event level, CDI is the negative base-two logarithm of the probability assigned to an observed output by an explicitly normative, context-conditioned distribution. The paper establishes that this event score is a contextual surprisal rather than, by itself, a proof of misconduct. At the distribution level, the expected CDI is cross-entropy and decomposes exactly into the entropy of the agent's output law plus its Kullback--Leibler divergence from the contextual law. The divergence term, not cross-entropy alone, isolates systematic distributional mismatch. The framework is extended through a stochastic channel model that separates agent behaviour from transmission noise, a carefully qualified truth-conditional extension for semantic reliability, and a multi-agent monitoring scheme. Theoretical results include non-negativity and identifiability properties, temporal chain rules, a data-processing bound for channel-observed deviation, and a likelihood-ratio construction for sequential surveillance. A reproducible binary-protocol experiment illustrates the distinctions among event CDI, cross-entropy burden, and distributional departure. The paper also specifies assumptions, calibration requirements, failure modes, and a validation agenda. CDI is proposed as a normative scoring framework; it does not infer intention, deception, or moral responsibility from an information score alone.
Language: English
Open Access status: green
OpenAlex ID: W7235072713
Cites:
Cited by:
Related to: 10