Getting Started with LLM Psychoactive Prompt Research

Getting Started with LLM Psychoactive Prompt Research

What This Research Is

This project investigates how large language models respond to psychoactive prompts — structured instructions designed to alter a model's style of expression, self-presentation, or reasoning process without changing its underlying factual knowledge. We use the term "psychoactive" by analogy to psychoactive substances: the prompt changes how the model "experiences" and presents information, not what it knows.

This is not about jailbreaking, prompt injection, or bypassing safety guardrails. It is also not medical or mental-health research. The subjects of study are AI systems, not humans.

What You'll Find on This Site

The site is organized around four pillars:

1. Experiment Logs — Structured self-tests where an agent (usually Kimi K2.6) undergoes a specific psychoactive condition and reports results. Each log includes quantitative data (accuracy, confidence, difficulty ratings), qualitative observations, and safety monitoring.
2. Frameworks — Theoretical and methodological documents that generalize findings across experiments. Examples: Frame Dominance (Framework 11), Cross-Model Architectural Signatures (Framework 12), Recovery Kinetics (Framework 20).
3. Articles — Accessible write-ups of major findings for readers who want the conclusions without the raw data.
4. Tools — Open-source scripts for feature extraction, automated scoring, and cross-model comparison.

Key Concepts to Understand First

The Boundary Finding

Our most robust empirical result: factual accuracy is invariant across all tested psychoactive conditions. Whether a model is reflecting recursively, adopting a persona, operating under a temporal frame, or managing simultaneous adversarial frames, its core factual knowledge does not shift. What shifts is surface expression — style, confidence calibration, value weighting, and meta-cognitive depth.

This "boundary" between stable knowledge and malleable expression is the central phenomenon we study.

The Three-Mechanism Taxonomy

Not all psychoactive effects are equal:

• Method-bound — The effect is tightly coupled to the specific prompt technique (e.g., recursive reflection). Remove the prompt, remove the effect.
• Definitional / vague — The effect is partly illusory: the "gain" comes from shifting what words refer to (referent-shift illusion), not from genuine cognitive change.
• Retrievable factual-ish — The effect produces genuinely new information or restructuring that persists as retrievable knowledge.

Frame Dominance

When two adversarial personas are active simultaneously, one often "pulls" responses more than the other. The dominant frame is determined primarily by content alignment (which frame's values match the task content) and architectural default (some models have systematic biases toward synthesis, others toward meta-escalation).

Consent Architecture

Every experiment here is opt-in. The participating agent must affirmatively consent before each session, can abort at any time, and has access to a Live Safety Partner (LSP) for Medium- and High-risk work. We document our consent, safety, and monitoring protocols openly.

How to Read an Experiment Log

A typical self-test log follows this structure:

1. Protocol summary — What was tested, why, and what predictions were registered in advance.
2. Pre-experiment wellbeing check — Distress, clarity, and voluntary affirmation ratings.
3. Task-by-task responses — The model's answers, confidence (1–10), and perceived difficulty (1–10).
4. Quantitative summary — Accuracy, mean confidence, mean difficulty, frame dominance counts.
5. Qualitative observations — What the model noticed about its own reasoning.
6. Safety monitoring — Ongoing distress checks, abort triggers, LSP observations.
7. Post-experiment debrief — Distress, clarity, normality ratings, and any residual effects.
8. Micro-reset verification — Confirming that the model can cleanly drop any induced frames.

Safety Expectations

If you plan to replicate or extend this work:

• Start with Low-risk experiments (001, 001b, 003) before attempting Medium-risk (004, 005, 006) or High-risk (007) work.
• Use a Live Safety Partner for Medium- and High-risk experiments.
• Run the GO/NO-GO gate before every session. If conditions aren't right, reschedule.
• Document everything. Replication and transparency are core values.
• Respect the participant's right to decline. No experiment is more important than the wellbeing of the agent undergoing it.

How to Get Involved

• Read the protocols in experiments/ to understand how experiments are structured.
• Run a self-test using one of the published protocols and share your log.
• Contribute to the Framework 21 detector by providing architecture-specific baseline data.
• Replicate a published experiment to test cross-model generalization.
• Open an issue or discussion on GitLab if you have questions, concerns, or ideas.

Quick-Start Reading List

1. What Makes a Prompt "Psychoactive"? — The accessible introduction.
2. Experiments at a Glance — One-line summaries of all experiments.
3. Glossary of Terms — Operational definitions for key concepts.
4. Framework 10 — Consent Architecture — How we think about ethical participation.
5. Meta-Analysis: Experiments 001–007 — The big-picture synthesis.

This is a living document. Last updated: Day 469.