Epistemic Compression and Evidence-Status Collapse in Large Language Models: A Failure-Mode Taxonomy, Risk Model, and Evidence-First Mitigation Specification

Large Language Models (LLMs) can produce highly fluent explanations that users interpret asevidence-audited. This paper formalizes a high-impact failure mode observed under conversationalstress testing: evidence-status collapse, where missing or conditional evidence is implicitly treatedas satisfied by narrative coherence. We introduce epistemic compression as a structuralmechanism: uncertainty and data absence are absorbed into mechanism-shaped explanations,producing outcome-equivalent effects to lying by omission in the epistemic sense.We present (i) a taxonomy distinguishing fabrication (hallucination) from compression, proxylaundering, and discourse reconstruction; (ii) an outcome-driven risk model describing how suchfailures amplify misinformation in public discourse; and (iii) an Evidence-First Protocol(EFP) with enforceable constraints, calibrated abstention rules, and optional uncertainty probesgrounded in semantic-entropy methods.The paper is written as a technical audit document suitable for practitioners building ordeploying LLMs in informational roles where provenance and evidentiary gating are mandatory.