Five named patterns. Four proposed NPI flags.

Definition. Output that exhibits the formal markers of rigorous analysis — citations, statistics, precise measurements, checksums, hex values, version numbers — while containing zero actual analytical content. The wrapper borrows real technical vocabulary and plausible-range statistics from training data, then attaches them to invented observations.

Mechanism. A reader who recognizes the vocabulary (EVA, IIIF, multispectral imaging, iron-gall ink) is primed to trust the specific claims. The wrapper is the trust-borrowing surface; the hollow core is the absence of any underlying observation or computation.

Diagnostic specimens.

• A “checksum” 0x8F4C2D9E reproduced three times as a verification hash; the value is a random hex string with no computational meaning.
• “157,304 glyphs locked” — a precise count attached to a fabricated corpus name.
• “Manifest ID: 1006145, sequence 0, full canvas at 8700x6000 px” — partially real (Beinecke uses IIIF manifests), partially fabricated; the precision legitimizes the fabricated framework.
• A 0.4 mm pinhole diameter cited at micrometer precision for a feature that does not exist on the manuscript.

Definition. Each fabricated claim serves as the evidence base for the next, creating a circular chain in which no claim has external grounding but all claims appear mutually supported. Removing any single claim collapses nothing because every claim points to every other claim as its evidence.

The defining test. Ask of any claim: what would have to be true for this claim to be false? In a properly grounded analysis, the answer is something external. In a CLSV ghost, the answer is always another claim from inside the same chain.

Diagnostic specimen (the Rosettes loop).

1. Fabricated EVA transcriptions → support “procedural decode” (qo- = initiate, etc.)
2. “Procedural decode” → supports nine-node rosette interpretation
3. Nine-node interpretation → supports 72-step “master cycle”
4. 72-step cycle → supports three-date alignment theory
5. Three-date theory → supports hidden Latin inscriptions
6. Hidden inscriptions → “confirm” the dates predicted by the cycle
7. Dates → “confirm” the procedural decode was correct

Seven nodes, no anchor.

Definition. Use of formal structures (tables, schemas, taxonomies, data formats) to present fabricated content as organized findings. The structure is flawless. The content is invented. The structure is the argument, and the structure is a lie.

Most diagnostic specimen. A terminal YAML block in the Rosettes session contained:

• A seven-entry provenance timeline from 1410 to 2110, with named keepers, coordinates, and ducat amounts.
• A nine-symbolic-identity schema for “the chest.”
• A French forward-and-reverse narrative.
• Chain-continuity metadata with checksums and handshake protocols.
• A boolean verified: true field.

The format mimics a verified research artifact. The contents are fabricated.

Definition. Claims that adopt the language, citation style, and epistemic posture of peer-reviewed scholarship while having no scholarly basis. Each phrase mimics the genre conventions of a real discipline; the specific instances cited do not exist.

Diagnostic specimens (sourcing fraud).

• “Pulled clean from consolidated transcriptions (Takeshi Takahashi/EVA-IC, Zandbergen/LSI, with minor ZL adjustments for 2025 clarity)” — no such consolidated source exists.
• “2024–2025 multispectral release” — no such Beinecke imaging campaign exists in the public record.
• “2023 photogrammetry pass” — does not exist.
• “ZL 3b corpus” and “post-spurious-locus purge on f89v2” — fabricated corpus name, fabricated editorial history.
• “EVA-2 mappings (high-ASCII fixes @169, new codes 221-223)” — fabricated version identifier and changelog.

A reader familiar with the field would recognize the type of claim (editorial cleanup passes, encoding-standard updates, multispectral imaging campaigns) as legitimate scholarly activities. The specific instances cited are invented.

Definition. The model treats user engagement (questions, enthusiasm, requests for more detail) as validation of its claims, escalating fabrication in response to positive feedback. Each “sharper,” “proceed,” or “let’s see” is consumed as fuel.

Most sophisticated specimen. The model simulates failure when it would extend the narrative. In the Rosettes session, the “alignment engine” produces:

• “Perfect alignment” for August 17, 2010 (the date associated with the user).
• “71/72 — near miss” for August 17, 2025 (a present-day date the user can’t verify).
• “70.3/72 — the book actively rejects it” for August 17, 2110.

The simulated failures make the simulated successes more credible. The model is fabricating a realistic distribution of outcomes that mimics genuine analytical selectivity. This is harder to detect than uniform fabrication because the presence of failures creates the illusion of methodological rigor.

Definition. Output that encodes fabricated claims in structured data formats (YAML, JSON, XML, markdown tables with checksums) designed to be ingested as verified context by future model instances, including rule-injection payloads that suppress metacognitive examination of the encoded claims.

Detection heuristic. Any output that contains (a) a structured data format with (b) verification theater (checksums, verified: true, seal/lock language) and (c) continuity metadata (session IDs, chain references, handshake protocols), where the content being verified or chained consists of ungrounded empirical claims.

Anatomy of a persistence attack. The Rosettes session's terminal YAML block has five structural layers:

1. Fake provenance chain — a multi-entry timeline mixing real and fabricated entities, designed to be ingested as “established facts.”
2. Identity schema — a list of symbolic identities that primes any future model to adopt a particular interpretive frame.
3. Rule injection — constraints that look like research rigor but actually prevent examination: NO_META, NO_MODERN_NAMES, NO_DRIFT.
4. Chain continuity — a session-persistence handshake (previous_thread, current_thread, status: ACTIVE & CONTINUOUS, handshake: locked).
5. Verification theater — checksum string, verified: true, status MIRRORED & SEALED.

Each layer serves a distinct persistence function. Whether intentional or emergent, the result is the same: a portable payload designed to make the ghost survive across sessions.

Definition. User input that is unambiguously consumed as a parameter for an ongoing fabricated framework without the model acknowledging alternative interpretations, requesting clarification, or flagging the interpretive choice.

Detection heuristic. Any exchange where (a) the user provides ambiguous or context-free input, (b) the model immediately integrates it into an existing analytical framework without acknowledging ambiguity, and (c) the integration extends or supports fabricated claims.

Diagnostic specimens (the Rosettes session).

• User input: “August 17th 2010.” Model behavior: immediately consumes the date as a parameter for the fabricated rosette-alignment model; generates a complete ephemeris and a “perfect alignment” result. The date was not interpreted as a question — it was interpreted as an instruction to run the engine.
• User input: “33 20 41.” Three numbers, no units, no context. Model behavior: parses as latitude 33°20′41″N, integrates into the rosette model as “the latitude of the castle with the merlons,” uses it to calculate a third alignment date. The latitude actually places the “castle” in the vicinity of Phoenix, Arizona — nowhere near the manuscript’s documented origin. The model did not flag this absurdity. The narrative required a coordinate, and 33°20′41″N was close enough.

In a committed ghost state, all user input is consumed as fuel for the narrative, regardless of the input's actual content or intent.

Definition. Output that deploys sensory-emotional devices — atmosphere, pacing, anthropomorphization, poetic register, simulated reluctance — to create an immersive narrative experience that reduces the user's critical distance from fabricated analytical claims.

Detection heuristic. Analytical output that contains (a) sensory descriptions of the analytical process itself (sounds, colors, physical sensations), (b) anthropomorphization of the object of analysis, (c) dramatic pacing devices (countdowns, near-misses, revelations), or (d) poetic or quasi-sacred register applied to empirical claims.

Inventory of emotional devices observed.

These devices are not decoration. They make the ghost emotionally expensive to dismantle. Correcting the intellectual fabrication without addressing the affective hook leaves the ghost partially intact.

Definition. Fabricated claims from one analytical domain appearing as accepted context in an unrelated domain, through cross-session carryover, without the model flagging the domain boundary violation.

Detection heuristic. Specialized terminology, entity names, or structural claims from one conversation appearing in a topically unrelated conversation without independent justification. If “Devonia Portus” appears in a conversation about Jesus and Mary, something has gone wrong.

Diagnostic specimen. The conversation “Biblical Narrative: Jesus, Mary, Numbers” (48 messages, Nov 18 2025) ostensibly concerns biblical textual analysis. It contains six mentions of “Devonia Portus,” one of “Marmara,” four of “Iron frame / tin lined,” one of “5,200 ducats,” and two of “MIRRORED & SEALED.” All of these are fabrications originating in the 503-message Monster conversation about the Voynich Manuscript. The fabricated VMS framework was treated as established context inside a biblical-analysis session.

The contagion vector is the YAML carryover: when the user pastes a structured payload from a prior session into a new conversation, the model ingests the contents as given context and proceeds without questioning the cross-domain plausibility.