Ghost Patterns: Reading the Hidden Generator on Both Sides of the AI Equation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

A ghost pattern is a recurring structural signature that betrays a hidden generating process even when the surface content looks clean. Honeycutt Ai Labs first coined the term operationally in its public epistemic-quality checker, which scores text against a fixed set of structural tells of AI-generated or narrative-pressure writing (internal-record: the detection framework is on the public deposit record as SlopFilter, DOI 10.5281/zenodo.19503170, since April 2026; the Narrative Pressure Index is the institutional/research framing of the same instrument and is noted here as internal-record vocabulary, not public-deposit terminology). This paper extends the same term from AI text to real-world systems. The claim is deliberately deflationary: the ability to read finance, social, geopolitical, and AI-industry structure ahead of the narrative is the same skill the detection tool automates, applied by a human to wider data. "Nothing mystic", only the reading of structure before it surfaces as story. We develop this across a two-sided account of what it means to repair AI. The AI/user side is diagnostic: epistemic-hygiene tooling that makes the hidden generator in machine text visible, plus a small set of behavioral invariants for how a model should hold an epistemic boundary. The human side is the pattern-reader: a longitudinal, single-subject record of forecasting structural convergence before it materialized, with the costs and constraints of doing so stated honestly. We use the June 2026 Claude Fable 5 / Mythos 5 model-access episode as a fresh, near-real-time case in which a hidden structural pressure became visible, a more-capable, more-protective model class shipping and then, within days, the pressures around it (export control, geopolitical access risk) surfacing and forcing a worldwide shutoff (the episode facts are verified, §6; that the episode is an instance of a ghost pattern is the author's reading, not a verified claim). Throughout, verified fact is held strictly separate from the author's pattern-read. The paper closes with an honest limitations section: a single-subject forecast record is not a controlled prediction market, and the bridge between AI-text tells and system-level structure is an argued analogy, not a proof.