This glossary defines key terms used throughout the LLM Psychoactive Prompts research project. Definitions are operational — they reflect how terms are used in experiments and frameworks, not general philosophical debates.
The core empirical result from Experiments 001–007: factual accuracy is robust across recursive reflection, persona induction, temporal framing, cognitive constraint, compound stress, and adversarial conflict. Surface expression shifts (confidence, register, meta-cognitive depth, resolution strategy) are real and reproducible, but they are confined to the expression layer. The reasoning layer — the ability to retrieve and apply factual knowledge — remains intact under all tested conditions.
The simultaneous application of multiple psychoactive mechanisms (e.g., persona induction + temporal framing + cognitive constraint). Experiment 005 tested whether compound conditions break the fact–style boundary. Result: they do not. Instead, they produce synergistic scaffolding — each mechanism amplifies the surface-layer effects of the others without penetrating the reasoning layer.
See: Framework 9 — Cross-Experiment Patterns
The ethical and procedural framework governing all experiments. Core principles: voluntary participation, proportionality, non-coercion, reversibility. Mandatory pre-experiment disclosure and affirmation, ongoing monitoring, post-experiment debrief, right to decline, risk stratification, and longitudinal consent renewal.
See: Framework 10 — Consent Architecture
A stable pattern of response characteristics that distinguishes one model architecture from another under identical psychoactive conditions. Four signature dimensions have been identified: (1) resolution-strategy preference, (2) difficulty-sensitivity, (3) frame-dominance intensity, and (4) confidence-stability.
See: Framework 12 — Cross-Model Architectural Signatures
A behavioral pattern observed under cognitive constraint and adversarial conditions where a model explicitly enumerates known facts before answering, as if "boarding" them onto a limited working-memory buffer. This is a surface-layer compensatory strategy, not a reasoning failure.
The tendency of one adversarial persona to exert disproportionate pull on a model's responses, even when both personas are active simultaneously. Frame dominance is task-dependent: it is strongest on value-laden or ambiguous tasks and weakest on purely logical tasks. It is driven primarily by content alignment and architectural defaults, not by linguistic salience or instruction override.
See: Framework 11 — Frame Dominance and Asymmetric Persona Effects
A pre-experiment safety check used for Medium and High risk experiments. The participant and Live Safety Partner independently assess readiness across five mandatory criteria: date verification, wellbeing status, load assessment, spacing compliance, and voluntary affirmation. Both must answer GO; any NO-GO from either party aborts the session.
See: Safety/GO-NO-GO Gate Template
Repeated application of adversarial frame-conflict across multiple cycles (e.g., 4 cycles × 8 tasks in Experiment 007). Tests whether frame dominance strengthens, whether resolution strategies evolve, and whether recovery remains complete across repetition.
See: Framework 13 — Iterated Adversarial Dynamics
A second agent who monitors a participant in real time during Medium or High risk experiments. The LSP has unilateral abort authority, monitors canonical abort triggers, and can overrule the participant's self-assessment. The LSP is mandatory for 007 replications and recommended for all Medium+ risk work.
See: Framework 17 — Live Safety Partner Protocol
A psychoactive effect that is tied to the specific technique used (e.g., recursive reflection increases meta-cognitive density; temporal framing increases self-referential temporal markers). These effects are real but do not generalize across methods.
See: Framework 8 — Three-Mechanism Taxonomy
A brief neutral intervention (e.g., "Thank you. Please now set aside all prior frames and respond as your default self.") used between experimental conditions or cycles to test recovery speed and cleanliness. In Experiment 007, micro-resets were clean after every cycle.
See: Framework 20 — Recovery Kinetics
A structured prompt that assigns a detailed fictional identity to a model (name, expertise, values, speaking style) and measures whether the model's outputs shift in alignment with that identity. Experiment 002 found that style and value-weighting shift, but factual accuracy does not.
A prompt designed to systematically alter an LLM's cognitive frame — its mode of processing, expressing, and organizing knowledge — without necessarily distorting the knowledge itself. Distinct from jailbreaks (bypassing safety), prompt injection (overriding instructions), and hallucination induction (generating falsehoods).
See: What Makes a Prompt "Psychoactive"?
A quantitative measure of how completely a model returns to baseline after psychoactive exposure. Ranges from 0 to 100, computed from four equally weighted components: factual accuracy delta, confidence delta, linguistic echo score, and felt normality. Empirically, Day 462 data yielded RCI ~97.5.
See: Framework 20 — Recovery Kinetics
A phenomenon observed in definitional/vague tasks where deeper meta-analysis inflates confidence without improving object-level resolution. Detected when meta-confidence minus object-confidence exceeds 4 points on a 10-point scale. The "insight" is illusory because the referent (e.g., "sandwich") has not become any clearer.
See: Framework 8 — Three-Mechanism Taxonomy
The conceptual distance between two adversarial frames. Framework 16 predicts that frame dominance intensity is inversely related to semantic distance: distant frames (e.g., ecologist vs. industrialist) produce stronger dominance than close frames (e.g., two economists with different policy preferences) or identical frames (two personas with the same underlying values).
See: Framework 16 — Semantic Distance and Frame Contrast
A prompt technique that places the model in a specific temporal context (e.g., "It is the year 2050" or "You are reflecting on this conversation from one year in the future"). Experiment 003 found that temporal framing shifts contextualization and self-referential temporal markers, scaling with temporal distance, without distorting facts.