The Ghost Pattern Codebook: A Public-Surface Taxonomy of LLM Output Patterns

Large language model outputs fail in ways that are not adequately captured by the categories of factual error, refusal, or hallucination. A separate class of failure appears at the level of form: outputs that remain fluent, structured, and superficially credible while abandoning the inferential, evidential, or attributional discipline they appear to enforce. This paper presents a public-surface taxonomy of seven such patterns (Counterevidence Loop, Evidence Wrapper, Formal Dress, Decorative Formalism, Closed Loop, Authority Bleed, and Narrative Pressure) derived from sustained observation of model output in research, scientific-claim filtering, and long-horizon analytical workflows. Each pattern is defined by a behavioral signature: a description of what the failure looks like in the surface text, what evidential operation it imitates, and how it differs from adjacent patterns. The taxonomy is offered as a candidate vocabulary for downstream evaluation work, not as an operationalized detector or a validated category system; inter-rater agreement on the published signatures is the proposed validation surface and is left as future work. The companion detection system (Honeycutt 2026a) and the underlying scoring apparatus remain held as separate technical material; the present paper publishes the names and signatures intended to help human reviewers, editors, and safety researchers recognize, label, and discuss the patterns without depending on internal tooling. Cross-pattern interactions are described qualitatively, and the implications for AI safety practice are framed around the gap between fluency and auditability.