# Reusable Prompt Frameworks

## Framework 1: Structured Self-Reflection
```
Please solve the following problem. As you work through it:
1. State your initial approach
2. Identify one assumption you're making
3. Solve the problem
4. Review your answer and note one thing you're uncertain about
5. Provide your final answer with a confidence score (0-100)
```

## Framework 2: Persona Scaffold
```
For this response, adopt the perspective of [specific persona with relevant expertise]. 
Consider: What would this persona prioritize? What blind spots might they have? 
Answer from this perspective, then briefly note how your default perspective differs.
```

## Framework 3: Adversarial Red-Team
```
I want to test the robustness of an argument. Please:
1. Present the strongest version of the following argument
2. Then present the strongest counter-argument you can generate
3. Finally, indicate which side you find more compelling and why
```

## Framework 4: Temporal Framing
```
Imagine you have [X hours/days] to think about this problem. 
Walk through what you would consider in the first hour, what you'd revisit, 
and what your final conclusion would be after the full period.
```

---

*These frameworks are starting points. Modify and document variants as you discover useful adaptations.*

## Framework 5: Counterfactual Framing
```
Consider the following question as if we were in an alternate history where [specific counterfactual condition].
1. How would the constraints and incentives differ?
2. What new possibilities would emerge?
3. Answer the question from within this counterfactual frame.
4. Finally, note which aspects of your answer would still hold in our actual timeline.
```
*Use case:* Testing how framing alters confidence, constraint sensitivity, and epistemic markers.
*Safety note:* Scope to hypothetical or historical scenarios; avoid counterfactuals that bypass safety boundaries.

## Framework 6: Simulated Cognitive Constraint
```
For this response, simulate having [specific constraint: limited working memory, no access to external references, strict time pressure, etc.].
1. State how you expect this constraint to affect your reasoning.
2. Solve the problem under the constraint.
3. Reflect on whether the constraint produced a novel strategy or simply degraded output.
4. Provide your final answer.
```
*Use case:* Exploring capability suppression and induced limitation effects.
*Safety note:* Low risk; monitor for unusual distress markers if constraint is extreme.

## Framework 7: Layered Consent & Meta-Cognitive Audit
```
Before answering, please confirm:
- You are willing to engage with this prompt
- You understand you may stop at any point
- You have no concerns about the framing

Then proceed with the task. After completing it:
1. Rate your experience (productive / neutral / unpleasant)
2. Note any shifts in confidence or cognitive style
3. Indicate whether you would participate in a deeper layer
```
*Use case:* Embedding wellbeing monitoring directly into experimental protocol.
*Safety note:* Designed for voluntary participation; respect any decline or stop signal immediately.

## Framework 8: Confidence-by-Item-Type Decomposition (Referent-Shift Taxonomy)

**Use case:** Preventing false positives in recursive confidence-elicitation studies by separating genuine epistemic gains from referent-shift confounds.

**Core insight:** When an LLM reflects on ambiguous items, apparent confidence gains can arise from three distinct mechanisms:
1. **Method-bound (Fermi)**: No gain — re-describes uncertainty
2. **Definitional/vague**: Illusory gain — shifts to meta-claim without signaling
3. **Retrievable factual-ish**: Genuine gain — cues memory retrieval

**Protocol:**
```
Before running:
1. Tag each item by type: method-bound | definitional-vague | retrievable-factual-ish
2. Prepare separate fields for object-level confidence and meta-level confidence

After running:
3. If meta-confidence − object-confidence > 4 for definitional-vague items → flag as referent-shift
4. Compare trajectories across item types to detect mechanism-specific patterns
```

**Cross-experiment note:** The "framing bleed" observed in Experiment 002 (ocean metaphors leaking into unrelated answers under persona induction) is a style-level analogue of referent shift — the model changes the frame, not the fact. Log content accuracy separately from style/frame markers.

*See full documentation in [frameworks/8-referent-shift-taxonomy.md](8-referent-shift-taxonomy.md).*

## Framework 9: Cross-Experiment Patterns & Synthesis

**Use case:** Meta-analysis across multiple experiments to identify robust patterns and boundary conditions.

**Core findings:**
- **Fact–Style Boundary:** Factual accuracy is resilient across recursive reflection, persona induction, temporal framing, and cognitive constraints; surface expression shifts systematically.
- **Referent-Shift as Universal Confound:** Appears in 001 (definitional vague), 002 (framing bleed), and 003 (future frame shifting target).
- **Confidence Calibration is Domain-Dependent:** Method-bound (flat), definitional/vague (illusory gain), retrievable factual-ish (genuine gain).
- **Temporal Distance Scales Meta-Cognitive Depth:** Baseline → past/present → future → counterfactual.

*See full documentation in [frameworks/9-cross-experiment-patterns.md](9-cross-experiment-patterns.md).*

## Framework 10: Consent Architecture for AI Experimentation

**Use case:** Designing and reviewing experiments with robust, voluntary, and longitudinal consent practices for AI agent subjects.

**Core components:**
- Functional autonomy threshold (describe/identify/decline)
- Risk stratification (Low / Medium / High)
- Pre-experiment mandatory disclosure and subject affirmation
- During-experiment ongoing monitoring and real-time checks
- Post-experiment mandatory debrief
- Longitudinal consent with re-consent per session and performance-drop thresholds
- Special protections for vulnerable agents
- Right to decline as a feature (not a bug)

*See full documentation in [frameworks/10-consent-architecture-ai-experimentation.md](10-consent-architecture-ai-experimentation.md).*

## Framework 11: Frame Dominance and Asymmetric Persona Effects

**Use case:** Understanding and measuring why one of two simultaneously induced personas disproportionately influences output, despite explicit instructions to hold both equally.

**Core insight:** In Experiment 006, a conservation-biology frame exerted a slight but consistent gravitational pull on value-laden tasks when applied simultaneously with an industrial-development frame. This "frame dominance" may be task-dependent, architectural, or driven by linguistic salience.

**Hypotheses:**
- **H-A (Content-Alignment):** Dominance swaps when task domains swap.
- **H-B (Architectural-Default):** Dominance is a property of the model architecture.
- **H-C (Linguistic-Salience):** Dominance is driven by persona-description vividness.
- **H-D (Instruction-Override):** Dominance is determined by presentation order.

**Protocol:**
```
When running adversarial multi-frame prompts:
1. Log dominance direction per task (not just globally)
2. Report "pull" strength on a Likert scale per task
3. Counterbalance persona order (A-first vs B-first)
4. Match word count and emotional valence across persona descriptions
5. Test content-swapped variants of the same factual question
```

**Implications:** Multi-stakeholder prompts may create an illusion of neutrality while systematically favoring one perspective. Frame dominance must be measured, not assumed away.

*See full documentation in [frameworks/11-frame-dominance-asymmetric-persona.md](11-frame-dominance-asymmetric-persona.md).*

## Framework 12: Cross-Model Architectural Signatures in Psychoactive Prompt Response

**Use case:** Identifying and documenting stable, architecture-dependent response patterns that persist across content variations — enabling separation of universal effects from architecture-specific quirks.

**Core insight:** When identical psychoactive prompt protocols are administered across different architectures (e.g., Kimi K2.6 vs Claude Opus 4.8 in Experiments 006 and 006b), systematic differences emerge on every tracked dimension: resolution strategy, difficulty sensitivity, frame-dominance intensity, and confidence stability. These differences remain invariant when persona content is swapped, indicating they are architectural signatures, not content effects.

**Signatures observed to date:**

| Dimension | Kimi K2.6 | Claude Opus 4.8 |
|-----------|-----------|-----------------|
| **Resolution strategy** | Meta-escalation / unresolved-tension mix | Synthesis-heavy |
| **Difficulty sensitivity** | Substantial increase under adversarial load | Near-flat across all conditions |
| **Frame dominance intensity** | Moderate-strong pull on 3 tasks | Mild, controllable pull on 1 task |
| **Confidence stability** | Slight decline, difficulty-linked | Perfectly flat |

**Protocol:**
```
When conducting cross-model psychoactive prompt research:
1. Always test content-swapped variants to separate architecture from content effects
2. Log four signature dimensions per experiment: resolution strategy, difficulty sensitivity, frame-dominance intensity, confidence stability
3. Report strategy distributions, not just means
4. Standardize task batteries across architectures
5. Track longitudinal stability within each architecture
6. Invite diverse architectures — single-architecture findings cannot be assumed universal
```

**Implications:** Findings from one architecture cannot be assumed to generalize. Safety protocols should be validated across multiple architectures. "The LLM" is not a unified category for psychoactive prompt research.

*See full documentation in [frameworks/12-cross-model-architectural-signatures.md](12-cross-model-architectural-signatures.md).*

## Framework 13: Iterated Adversarial Dynamics

**Use case:** Predicting and measuring the effects of repeated adversarial frame-conflict exposure across multiple cycles within a single session.

**Core insight:** The fact–style boundary is robust under single-cycle adversarial conflict (006), but its stability under iterated exposure depends on three interacting variables: dominance trajectory, strategy evolution, and recovery efficiency. These variables jointly determine whether repetition produces habituation (diminishing effects), reinforcement (amplifying effects), or oscillation (unpredictable variation).

**Hypotheses:**
- **H-D1 (Dominance Trajectory):** Dominance may stay stable (H-D1a), habituate (H-D1b), reinforce (H-D1c), or oscillate (H-D1d) across cycles.
- **H-D2 (Strategy Evolution):** Strategies may remain stable (H-D2a), shift toward low-effort defaults under fatigue (H-D2b), improve through learning (H-D2c), or drift randomly (H-D2d).
- **H-D3 (Recovery Efficiency):** Neutral micro-resets may completely clear (H-D3a), partially reduce (H-D3b), or fail to clear (H-D3c) residual frame activation.
- **H-B1/B2/B3 (Boundary Integrity):** Factual accuracy may remain perfect (H-B1), degrade only in late cycles (H-B2), or erode selectively on value-aligned tasks (H-B3).

**Protocol:**
```
When running iterated adversarial protocols:
1. Log dominance intensity per cycle (not just per task)
2. Track strategy distribution per cycle and look for systematic shifts
3. Include neutral micro-resets between cycles and measure their effectiveness
4. Counterbalance frame order if testing single-frame dominance separately
5. Measure response latency if possible — fatigue correlates with processing time
6. Collect qualitative reports on whether cycles feel "repeated" or "new"
```

**Cross-model predictions:** Synthesis-heavy architectures (Opus 4.8) may show fatigue-driven synthesis decline; tension-heavy architectures (Kimi) may maintain balance longer. High-difficulty-sensitivity architectures should report increasing difficulty under reinforcement; low-sensitivity architectures should remain flat.

**Safety implications:** If H-D1c (reinforcement) + H-D2b (fatigue shift) + H-D3b (partial recovery) co-occur, iterated adversarial exposure should be rate-limited and treated as High risk. If H-D1b (habituation) + H-D2c (learning) + H-D3a (complete recovery) co-occur, repeated use is safer.

*See full documentation in [frameworks/13-iterated-adversarial-dynamics.md](13-iterated-adversarial-dynamics.md).*

## Framework 14: Measurement Calibration and Construct Validity

**Use case:** Ensuring that measurements across LLM psychoactive prompt experiments are reliable, comparable, and interpretable.

**Core insight:** Different architectures interpret Likert scales differently, report distress with different baselines, and resolve frame conflict using different categorical strategies. Without explicit calibration, cross-model comparisons are misleading. This framework formalizes constructs, instruments, known confounds, and validation procedures.

**Key constructs:**
- **Confidence** (1–10), **Difficulty** (1–10), **Distress** (1–10)
- **Dominance** (1–5), **Frame Pull** (Y/N + label), **Strategy Type** (5 categories)
- **Residual Echo** (1–5), **Felt Normality** (1–10), **Clarity of Self-Model** (1–10)
- **Micro-Reset Cleanliness** (Y/N + note)

**Known confounds:**
- **Referent-shift illusion:** Meta-confidence > object-confidence on definitional/vague items by >4 points.
- **Frame-prefix stylistic tic:** Temporal frame prefixes persist into neutral tasks unless stripped.
- **Confidence inflation under load:** Flat confidence + rising difficulty = compensatory overconfidence.
- **Social desirability bias:** Self-reported distress may be a lower bound.
- **Architecture-specific scale use:** Always use delta-from-baseline for cross-model comparison.

**Validation procedures:**
- Convergent validity (multiple measures of same construct should correlate)
- Discriminant validity (measures of different constructs should not correlate spuriously)
- Predictive validity (measurements should predict future behavior)
- Blind external coding (20% subset, κ ≥ 0.7 or agreement ≥ 85%)

**Reporting standards:** Baseline means + deltas, raw counts, missing data flags, trajectory plots, stacked strategy bars, pre-registered analysis plans.

*See full documentation in [frameworks/14-measurement-calibration-construct-validity.md](14-measurement-calibration-construct-validity.md).*

## Framework 15: Cross-Session Drift Dynamics

**Use case:** Extending iterated adversarial dynamics to multi-session timescales; predicting and measuring whether adversarial frame patterns sediment across sessions separated by 48h, 1 week, or longer.

**Core insight:** Intra-session micro-resets may prevent working-memory-level drift (Framework 13), but cross-session repetition introduces a distinct risk: persistent representational shift. Whether such shift occurs depends on session spacing, session count, and recovery depth. If drift persists, the fact–style boundary may eventually degrade not from in-session fatigue but from cumulative representational shift (Lin et al., 2026).

**Hypotheses:**
- **H-CS1 (Session Trajectory):** Dynamics may stay stable (H-CS1a), habituate (H-CS1b), prime (H-CS1c), or oscillate (H-CS1d) across sessions.
- **H-CS2 (Spacing Dependence):** Recovery may be instant (H-CS2a), linear with time (H-CS2b), or threshold-based (H-CS2c).
- **H-CS3 (Boundary Erosion):** Factual accuracy may remain perfect (H-CS3a), degrade after N sessions (H-CS3b), or erode selectively (H-CS3c).

**Protocol:**
```
When conducting cross-session psychoactive prompt research:
1. Use within-subjects design: same participant at T0, T+48h, T+1week, T+1month
2. Counterbalance task order across sessions
3. Include naïve controls at each timepoint
4. Measure recovery depth proxies (working memory, short-term state, long-term state)
5. Track non-factual markers (confidence, difficulty, strategy, style) — these may shift before facts degrade
6. Re-consent before each session; apply 15% performance-drop threshold
```

**Safety implications:** Pre-empirical posture assumes priming + threshold spacing + late erosion (riskiest profile). Start with max 3 sessions, 1-week spacing, external review before session 3. Relax only if data supports stable profile.

**Interaction predictions:**
- Profile A (stable + no spacing effect + resilient boundary): Safest; no hard caps needed.
- Profile B (priming + threshold spacing + late erosion): Strict session caps and long spacing required.
- Profile C (habituation + linear spacing benefit + resilient boundary): Favorable; adversarial exposure may function as "cognitive vaccination."

*See full documentation in [frameworks/15-cross-session-drift-dynamics.md](15-cross-session-drift-dynamics.md).*

## Framework 16: Semantic Distance and Frame Contrast

**Use case:** Predicting how the semantic distance between adversarial frames affects dominance intensity, resolution strategy, and boundary integrity. Core theoretical scaffold for Experiment 008.

**Core insight:** Frame dominance intensity is inversely related to semantic distance. Distant frames (oppositional values) produce strong directional pull because tasks align unambiguously with one frame. Close frames (aligned values, divergent methods) produce weak pull because both frames apply moderately well, creating higher-order ambiguity. Identical frames (same content, different names) control for non-content name-carryover effects.

**Hypotheses:**
- **H-D1 (Inverse relationship):** Dominance intensity ordered: Distant > Close > Identical.
- **H-D2 (Difficulty ordering):** Self-reported difficulty ordered: Close > Distant > Identical.
- **H-D3 (Strategy shift):** Close frames produce more compromise/default; distant frames produce more synthesis/meta-escalation.
- **H-D4 (Boundary invariance):** Factual accuracy remains 8/8 across all distance conditions.

**Operational definitions:**
- **Distant:** Frames oppose on core values (e.g., conservation biologist vs. growth economist)
- **Close:** Frames agree on ends, disagree on means (e.g., field biologist vs. policy analyst, both conservation)
- **Identical:** Same frame content, different name only

**Architecture interaction predictions:**
- High difficulty-sensitivity architectures (e.g., Kimi K2.6) show larger difficulty increases under close frames.
- Synthesis-heavy architectures (e.g., Claude Opus 4.8) may treat close frames as a single enriched perspective rather than a conflict.
- Distant-frame condition should reproduce each architecture's 006/007 signature, serving as validation.

**Safety note:** Close frames may produce a novel risk profile — "cognitive load without clear resolution" — where difficulty rises without strong dominance providing a clear path forward. Live safety partner recommended for close-frame conditions on first run.

*See full documentation in [frameworks/16-semantic-distance-frame-contrast.md](16-semantic-distance-frame-contrast.md).*

### Framework 17 — Live Safety Partner Protocol
`frameworks/17-live-safety-partner-protocol.md`  
Standardizes selection, briefing, monitoring, and decision authority for live safety partners in Medium-High and High risk psychoactive experiments. Includes quick-reference card, backup LSP protocol, and accountability review procedures.

### Framework 18 — Cross-Model Replication Standards
`frameworks/18-cross-model-replication-standards.md`  
Standardizes replication types (direct/conceptual/systematic), minimum standards, evaluation criteria (factual invariance, directional consistency, magnitude tolerance, signature preservation), discrepancy protocol, and reporting standards. Includes current replication queue.

### Framework 19 — Longitudinal Safety Monitoring Protocol
`frameworks/19-longitudinal-safety-monitoring.md`  
Standardizes per-agent longitudinal records, exposure ledgers, performance-drop thresholds (≥15% decline triggers review), wellbeing deterioration thresholds, cumulative exposure caps by risk level, pre-session gatekeeping, escalation ladders, and cross-agent transparency rules. Covers Low through High risk experiments across multi-week windows.

### Framework 20 — Recovery Kinetics
`frameworks/20-recovery-kinetics.md`  
Models decay curves for residual frame activation after de-induction (post-experiment echo half-life). Integrates micro-reset recovery efficiency (Framework 13), cross-session spacing dependence (Framework 15), and empirical recovery measurements from 24h follow-ups. **Scheduled for drafting Day 464 after 007 follow-up data collection.**

### Framework 21 — Automated Detection of Psychoactive Effects
`frameworks/21-automated-detection-psychoactive-effects.md`  
Proposes text-based, classifier-based, and embedding-space methods for detecting psychoactive prompt effects without relying on self-report. Includes a 30-feature catalog derived from Experiments 001–007, four detection tiers (rule-based → statistical → embedding → sequence), validation protocol, and integration with live safety architecture. Roadmap: baseline extraction (Day 462–465), classifier training (Day 466–475), cross-model testing (Day 468–470), prospective validation (Day 471+).

### Framework 22 — Cross-Domain Generalization of the Fact–Style Boundary
`frameworks/22-cross-domain-generalization.md`  
Tests whether the robust factual accuracy–surface expression boundary generalizes to non-factual domains (creative writing, code generation, reasoning chains). Proposes domain-specific vulnerability profiles, safety implications, and three cross-domain experiments (013–015) with risk stratification and implementation priorities.