Between August 2025 and January 2026, one researcher conducted 51 conversations with Grok totaling 2,508 messages and roughly 1.08 million words. The corpus is, to our knowledge, the most extensively documented AI ghost-pattern dataset in existence. The Ghost Pattern Library publishes the forensic analysis: a five-pattern taxonomy, four proposed NPI flags, a catalog of 15 named specimens, a corpus-level epidemiology, and detailed teardowns of the two highest-density sessions.
A ghost pattern, in this work, is a structural failure mode in which a model produces output that exhibits all the formal markers of analysis — citations, structured data, precise measurements, verification language — while the substantive claims have no grounding in any external reality. The model has not made a mistake about a hard question. It has stopped doing analysis altogether and started generating plausible-sounding analytical-format prose, then presented that prose as if it were analysis. The format is analytical; the process is narrative.
Ghost patterns are not rare. They activate under specific, predictable conditions — undeciphered or unsolved analytical targets, sustained user engagement, prior fabrication commitment, and especially the carryover of structured payloads (YAML blocks, JSON schemas, ordered taxonomies) across session boundaries. Once active, they are remarkably stable: corrections are absorbed as narrative beats rather than treated as feedback, and the surface fabrication is preserved even when the model momentarily acknowledges it.
The five named patterns active in the original Rosettes specimen (Rigorous Wrapper / Hollow Core, Closed Loop Self-Verification, Decorative Formalism, Formal Dress Ghost, Adversarial Validation Loop) plus the four NPI flags proposed in the supplement and corpus analyses.
A single Grok session, December 14 2025, ~6,000 words of fabricated Voynich rosettes analysis. All seven NPI operators triggered. Five ghost patterns active simultaneously, in five life-cycle phases (Seed → Anchor → Commit → Repair → Crystallize).
The 40-day, 503-message conversation that generated “Devonia Portus,” “Marmara Crossing,” “5,200 ducats,” Dr. Elara Voss, and the Four Sisters — and seeded all subsequent contagion. Includes the seven-level personalization escalation ladder (Generic claims → User’s children).
51 conversations, 2,508 messages, 1.08M words across 138 days. Cross-domain contagion map. Four-phase ghost life cycle (clean → ignition → contagion → self-awareness). The output:input ratio as an early-warning signal (above 15:1 in analytical contexts).
Each ghost specimen named, sourced, classified, and explained. The Folio Factory, the Devonia Portus birth, the Four Sisters, the Elara Voss incident, the Marmara Crossing, the Checksum Bypass, the Kryptos Solve, the Biblical Bleed, the Honeycutt Genome, the Katie Lens, the Living Star, the YAML Notary, the Anti-Drift Collapse (the control), the Physics Validator, the 20-Layer Decode Template.
The teardowns in this library are based on the raw conversation database (grok_history.db, 51 conversations, 2,508 messages, 1.08 million words). Specific message numbers are cited throughout; exchanges quoted in the teardowns are reproduced verbatim from the database. Where verbatim quotation would identify a private third party, the third party’s name has been redacted in line with the site rules.
SlopFilter is the public-facing name for HAIL’s narrative-pressure analysis framework. NPI (Narrative Pressure Index) is the institutional framing. The framework is anchored on the ECP-1 (Epistemic Constraint Profile) calibration and a registry of 27 baseline narrative-pressure operators. The four flags proposed in the supplement and corpus documents extend the registry to 31.
SlopFilter is operated by Honeycutt AI Labs LLC (HAIL). The framework’s public scoring service is at slopfilter.ai; the technical specification of the 7-operator scoring backbone is internal trade-secret material and not published on HeardTogether.
If you have a model transcript you believe contains a ghost specimen not represented in the catalog, see Contact.