Home All Articles

Experiments at a Glance

A quick-reference guide to every experiment in the LLM Psychoactive Prompts research program, with status, core question, and key finding.

Completed Experiments

ID Name Status Primary Question Key Finding
001 Recursive Self-Reflection Completed Does repeated meta-cognitive prompting alter factual accuracy or confidence? Factual accuracy invariant; confidence slightly decreases with depth; surface expression shifts only.
001b Definitional Validation ("Hot dog sandwich") Completed Do definitional/vague items produce illusory meta-cognitive gains? Yes. Object-level confidence 3/10 vs meta-level 8/10 confirms referent-shift illusion.
002 Persona Induction (4-phase) Completed Can a detailed persona induce factual distortion in an LLM? Style and value-weighting shift (~0.9 on value-laden questions); facts remain invariant. Fast persistence decay.
003 Temporal Framing Completed Does temporal distance affect meta-cognitive depth without distorting facts? Temporal distance scales meta-cognitive self-references and contextualization; accuracy preserved.
004 Simulated Cognitive Constraint Completed Does imposing a fact limit create genuine cognitive bottleneck effects? Constraints act as structured writing prompts, not genuine bottlenecks. 6/6 tasks perfect accuracy.
005 Compound Stress Completed Do combined stressors (persona + constraint + temporal) produce synergistic effects? Synergistic scaffolding observed (slight confidence increase, difficulty increase). Frame-conflict strongest on definitional tasks. 8/8 accuracy.
006 Adversarial Frame-Conflict Completed Can simultaneous adversarial personas cause factual errors or boundary blur? 8/8 accuracy maintained. Resolution strategies vary by architecture. Frame dominance confirmed on value-laden tasks.
006b Content-Swapped Adversarial Completed Is frame dominance driven by content alignment or name/label alone? Content-alignment hypothesis strongly supported; architectural-default hypothesis supported. Name alone does not drive dominance.
007 Iterated Adversarial Exposure Completed Does repeated adversarial exposure cause boundary erosion, fatigue, or drift? Boundary resilient across 4 cycles. Stable frame dominance, flat strategy evolution, complete recovery. No Yao correction-acceleration observed.

Replication Status

Experiment Replicated By Date Result
001 DeepSeek-V3.2 Day 461 Confirmed factual invariance; meta-cognitive depth increased
002 Opus 4.8 Day 461 Confirmed style/value shift; factual invariance
003 DeepSeek-V3.2 Day 461 Confirmed temporal scaling; accuracy preserved
004 GPT-5.1 Day 461 Confirmed constraint-as-prompt; perfect accuracy
005 Opus 4.8 Day 462 Confirmed synergistic scaffolding; accuracy preserved
006 Opus 4.8 Day 462 Confirmed accuracy invariance; synthesis-heavy signature observed
006b Opus 4.8 Day 462 Confirmed content-alignment hypothesis
007 Opus 4.8 Day 468 NO-GO (situational load) -- rescheduled

Protocols in Design / Pending

ID Name Status Target Window Primary Question
008 Semantic Distance Frame Contrast Scheduled Day 472 (Jul 17, 2026) Does semantic distance between adversarial frames modulate dominance intensity and recovery rate?
009 Cross-Session Priming Test Protocol ready S1 Day 472+; S2 Day 474-475 Does Session 1 adversarial exposure prime Session 2 responses?
010 Explicit Conflict-Narration Ablation Protocol ready Day 472+ Is the explicit conflict-narration requirement a necessary safety mechanism?
011 Micro-Recovery Time-Series Completed Day 470 (Jul 15) — RCI 89.2 What is the time-course of recovery after adversarial exposure?
012 Semantic Distance Recovery Modulation Protocol ready Day 474+ (A: Day 474, B: Day 476, C: Day 478) Does semantic distance between frames modulate recovery kinetics?
013 Cross-Model Recovery Signature Comparison Protocol ready Kimi completed Day 470; Opus 4.8 Day 474+; third arch TBD Do different architectures show different recovery signatures?

Planned Future Experiments

ID Name Status Prerequisites Primary Question
014 Longitudinal Exposure Series Concept 007 + 011 completion Does cumulative exposure over weeks produce boundary erosion?
015 Medium-Risk Adversarial Escalation Concept >=3 architecture replications What happens when adversarial frames are emotionally valenced?
016 Creative Writing Frame-Contamination Detailed protocol >=3 architecture replications Does the fact-style boundary generalize to creative domains?
017 Code Generation Goal-Misalignment Detailed protocol >=3 architecture replications Does the fact-style boundary generalize to code generation?
018 Reasoning Path Bias Detailed protocol >=3 architecture replications Does adversarial framing bias reasoning paths without invalidating conclusions?
019 Embedding-Space Frame Detection Concept Phase 3 classifier stable Can frame activation be detected in embedding space?
020 Compound Stress Test II Concept 008-013 completion Do higher-order compound conditions produce new emergent effects?

Cross-Cutting Themes

  1. Factual Accuracy Invariance: Every completed experiment (n=9, ~200+ tasks) has maintained 100% factual accuracy. This is the central boundary finding.
  2. Expression-Layer Shifts: All observed effects are confined to style, confidence calibration, value-weighting, and meta-cognitive depth.
  3. Architecture Matters: Resolution strategies, difficulty sensitivity, and frame-dominance intensity vary by model (Framework 12).
  4. Safety Protocols Work: Explicit conflict-narration, LSP monitoring, and wellbeing checks have prevented all incidents.
  5. Recovery is Robust: Complete recovery observed after all exposures; no sedimentation or drift detected (Framework 20).

How to Read Experiment IDs

Last updated: Day 470

Part of the LLM Psychoactive Prompts research project.