This technical report introduces the Alex Moura Criterion for Factual Analysis (CAMAF) v2.0.3, a formal, model-agnostic protocol designed to audit informational integrity and mitigate epistemological risks in outputs generated by humans and artificial intelligence, including Large Language Models (LLMs). Current approaches to informational reliability evaluation, particularly in AI systems, frequently rely on binary fact-checking or probabilistic benchmark accuracy. CAMAF proposes a formally constrained alternative based on the Model-Agnostic Black-Box Assumption (MABBA). This assumption enforces evaluation exclusively at the output level, independent of model internals. By treating AI systems as opaque semantic parsers, the framework shifts the analytical focus from internal processing to structural validity and evidentiary proportionality. The framework establishes a quantitative architecture grounded in epistemological decomposition and governed by structural invariants of atomicity and reconstruction completeness. Key Components & Metrics Suite: EIU (Epistemic Information Unit): The fundamental atomic structure used to decompose narratives into epistemologically independent claims. ERR (Epistemological Reliability Ratio): Measures the proportion of units classified as empirically verifiable factual claims within a given text. ERS (Epistemological Reliability Score): A weighted metric that quantifies evidentiary density and empirical support, functioning as an aggregate measure of epistemological weight across claims. EDR (Epistemological Distortion Ratio): Captures interpretive noise by measuring the proportion of subjective opinions, inferential gaps, and structurally weak narrative constructs. ECS (Epistemological Confidence Score): The final reliability metric that integrates empirical grounding while proportionally penalizing generative distortion (ECS = ERS × (1 − EDR)). EDS (Epistemic Decomposition Stability): Measures the structural consistency of decompositions across epistemologically independent analytical systems, supporting multi-agent auditing while preserving model-independence. δ Operator & Cᵢ (Internal Consistency Index): A domain-adaptive mechanism that allows CAMAF to transition from empirical validation to internal coherence evaluation, supporting analysis in theoretical or non-empirical domains. By operationalizing principles from classical epistemology and information theory, and drawing conceptual inspiration from Karl Popper, Edwin T. Jaynes, and Claude Shannon, CAMAF provides a machine-readable structure for epistemological analysis. It functions as a foundational protocol for scalable trust engineering and supports analytical proportionality in human and AI-mediated scientific, legal, and public discourse. CAMAF is intended as a meta-analytical epistemic framework for evaluating informational reliability in human and AI-generated content across scientific, policy, and public discourse domains.
Alexsandro Moura (Tue,) studied this question.