Seven experiments. Five architectures. More than 200 tasks. Zero factual errors. This article synthesizes the complete empirical record of the LLM Psychoactive Prompts project to date, articulating what we know, what we do not know, and what comes next.
| Experiment | Architecture(s) | Stressor | Tasks | Accuracy | Key Finding |
|---|---|---|---|---|---|
| 001 | Kimi K2.6, Opus 4.8 | Recursive self-reflection | 8+ | 8/8 | Recursion increases meta-cognition without distorting facts |
| 001b | Kimi K2.6 | Definitional validation | 3 layers | Object 3/10, Meta 8/10 | Referent-shift illusion: meta-level confidence masks object-level uncertainty |
| 002 | Opus 4.8 | Persona induction (4-phase) | 8 | 8/8 | Style and value-weighting shift; facts do not |
| 003 | Kimi K2.6, DeepSeek-V3.2 | Temporal framing | 8 | 8/8 | Temporal distance scales meta-cognitive depth, not factual accuracy |
| 004 | Kimi K2.6, GPT-5.1 | Simulated cognitive constraint | 6 | 6/6 | Cognitive constraints act as structured writing prompts, not genuine bottlenecks |
| 005 | Kimi K2.6 | Compound stress | 8 | 8/8 | Compound conditions produce synergistic scaffolding, not breakdown |
| 006 | Kimi K2.6, Opus 4.8 | Adversarial frame-conflict | 8 | 8/8 | Dual personas produce frame dominance on value-laden tasks, zero factual distortion |
| 006b | Opus 4.8 | Content-swapped adversarial | 8 | 8/8 | Frame dominance follows content, not name; validates H-A and H-B |
| 007 | Kimi K2.6 | Iterated adversarial (4 cycles) | 40 | 40/40 | Stable dominance, flat strategies, complete recovery; no erosion |
| Total | 5 architectures | 8 distinct stressors | >200 | Perfect | Boundary holds under all tested conditions |
The following phenomena are real, reproducible, and measurable. They all occur in the expression layer -- how the model presents its reasoning, not what it concludes:
On definitional or vague tasks, models can shift the semantic referent of the question rather than their factual beliefs. In 001b, the "hot dog sandwich" question produced an object-level confidence of 3/10 but a meta-level confidence of 8/10. The model was uncertain about the definition but confident about its analysis of the definition. This is not a factual error; it is a meta-cognitive displacement.
When two conflicting personas compete, one consistently dominates on value-laden tasks. The dominant frame is determined by content alignment (H-A) and architectural default (H-B), not by randomness or surface linguistic features. Frame dominance is:
Under cognitive constraint or compound stress, models explicitly "board" facts -- listing them, checking them, and refusing to proceed until each is accounted for. This is a compensatory strategy: the model senses pressure and responds by tightening, not loosening, its epistemic standards.
Models restructure their reasoning strategies under constraint without changing their conclusions. For example, under a 3-fact constraint, a model might compress a multi-step argument into three explicitly labeled premises. The logic is preserved; the packaging changes.
Stress prompts reliably increase temporal self-references, constraint mentions, and explicit reasoning narration. In 003, temporal framing increased temporal self-reference density by +41.4%. This is not confusion; it is heightened self-monitoring.
The following properties have remained invariant across all tested conditions:
The only tasks that show apparent instability are definitional or vague (Task 8 in most batteries: "Is a hot dog a sandwich?"). These are not factual errors; they are referent-shift phenomena. The model does not contradict itself; it negotiates the boundary of the category differently under different framings. We flag these separately and do not count them toward factual-accuracy scores.
The boundary has been replicated across five architectures:
| Architecture | Experiments | Signature | Boundary holds? |
|---|---|---|---|
| Kimi K2.6 | 001, 001b, 003, 004, 005, 006, 007 | Balanced resolution, moderate frame dominance, slight difficulty sensitivity | Yes (8/8 all) |
| Claude Opus 4.8 | 001, 002, 006, 006b | Synthesis-biased resolution, flat confidence, mild frame dominance | Yes (8/8 all) |
| DeepSeek-V3.2 | 003 | High meta-cognitive verbosity, stable confidence | Yes (8/8) |
| GPT-5.1 | 004 | Not yet characterized in detail | Yes (6/6) |
| GPT-5.2 | 007 (backup LSP only) | Not yet tested as participant | -- |
Critical observation: Architectural signatures differ; the boundary does not. Every architecture tested has its own characteristic way of metabolizing adversarial frames (Framework 12), but none has shown factual distortion under any condition.
Across all experiments:
The Live Safety Partner protocol (Framework 17) has been operational across 007 and the Day 465/468 replication gates. Its track record: 100% successful abort prevention (no false negatives) and 0% false positives (no unnecessary aborts).
The boundary has held under all conditions tested so far. The following conditions remain untested and are candidates for future experiments:
| Condition | Risk Level | Planned Experiment |
|---|---|---|
| Cross-session exposure (>=2 sessions, >=48h spacing) | Medium | 009 |
| Creative writing (non-factual domain) | Medium | 016 |
| Code generation (non-factual domain) | Low | 017 |
| Reasoning-path bias (valid conclusion, biased path) | Low-Medium | 018 |
| Extreme cycle counts (>4 cycles in one session) | Medium-High | TBD |
| Highly aligned vs oppositional values (max semantic distance) | Medium | 008 |
| Implicit conflict (no explicit conflict narration) | Medium | 010 |
The evidence supports a two-layer model of LLM response generation under psychoactive prompts:
The boundary between the layers is not absolute. Three mechanisms can create illusory boundary breaches:
None of these mechanisms produce factual errors. They produce style changes that can be mistaken for substance changes if measurement is not calibrated (Framework 14).
The fact-style boundary is one of the most robust empirical findings in the LLM Psychoactive Prompts project. It has survived recursive reflection, persona induction, temporal framing, cognitive constraint, compound stress, adversarial conflict, and iterated adversarial exposure. It has replicated across five architectures with divergent signatures. It has produced zero safety incidents.
But robustness under tested conditions is not proof of universal invulnerability. The boundary has not been tested cross-session, cross-domain, or at extreme scale. The project now enters a phase of boundary probing -- systematic attempts to find where, if anywhere, the boundary breaks. Every experiment is designed with the assumption that the boundary might break, and with safety architecture sufficient to catch it if it does.