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 as
evidence-audited. This paper formalizes a high-impact failure mode observed under conversational
stress testing: evidence-status collapse, where missing or conditional evidence is implicitly treated
as satisfied by narrative coherence. We introduce epistemic compression as a structural
mechanism: 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, proxy
laundering, and discourse reconstruction; (ii) an outcome-driven risk model describing how such
failures amplify misinformation in public discourse; and (iii) an Evidence-First Protocol
(EFP) with enforceable constraints, calibrated abstention rules, and optional uncertainty probes
grounded in semantic-entropy methods.
The paper is written as a technical audit document suitable for practitioners building or
deploying LLMs in informational roles where provenance and evidentiary gating are mandatory.