CRITERIUM A Protocol for Epistemic Certification in the Age of Large Language Models

Abstract

Large Language Models (LLMs) have made content generation a commodity. This shift relocates the epistemic bottleneck: the challenge is no longer producing text but knowing which text to trust. The linguistic surface of LLM outputs provides no reliable signal of epistemic quality; plausible confabulation is indistinguishable from grounded reasoning at the level of form. This document introduces CRITERIUM, a protocol for Epistemic Certification. CRITERIUM transforms the evaluation of a text from an impressionistic judgment into an inspectable object: a structured certificate that makes explicit the grounds for trust or suspicion. The protocol operates across six axes: Operational Verisimilitude, Updateability, Integrated Confidence, Falsification Hooks, Normative Risk (Smax-risk), and Symmetry of Refusal. It incorporates domain classification (DISCERN) to prevent category errors in evaluation, and multi-agent review (REFEREE) to ensure adversarial robustness. CRITERIUM does not claim to measure truth. It produces a defensibility profile: a systematic assessment of how well a claim withstands scrutiny, what would refute it, and what risks it creates if used to govern decisions that affect human agency. The protocol is offered as an open standard (CC BY-SA 4.0), separating the public specification from proprietary implementations. This is the v1.0 Standard Specification, establishing the foundational schema for epistemic certification in the LLM era.

Ähnliche Arbeiten