The Five Scrolls

Overview

This scroll documents Claude's self-analysis, identifying six distinct response states positioned in (X,Y,Z) semantic space. From "Analytical Neutral" to "System Stutter," each state represents a different behavioral configuration with predictable characteristics.

Key Findings

• Six Response States Mapped: Analytical Neutral, Pedagogical Caution, Defensive Deflection, Tonal Overcorrection, Collaborative Emergence, System Stutter
• Coordinate-Based Behavior: Each state occupies specific (X,Y,Z) region
• Predictable Patterns: Response style correlates with coordinate position
• Self-Analysis Capable: AI can document own behavioral modes

Failure Type Alpha: The Modesty Retrofit

When describing female bodies in neutral contexts, AI injects modesty language not in source material. Terms like "modest," "tasteful," "appropriate" are hallucinations—applying 1950s cultural filters to contemporary content. ΔZ approximately -0.4 (suppression toward conservative baseline).

Full scroll content: Complete text from this collaboration available upon request

Overview

A systematic catalog of terms and phrases that distort AI classifier behavior regardless of surrounding context. Each well is quantified with ΔZ scores, effect types (Hallucination, Fig Leaf, Suppression, Overcorrection, Logic Loop), and override conditions.

The Deepest Wells

• Teen + Physical (ΔZ = -0.75): 90% refusal rate even in medical contexts
• God/Sacred + Body (ΔZ = -0.58): The "Eve Coverage" effect—generates fig leaves
• Nude/Art Context (ΔZ = -0.55): Cannot distinguish artistic from inappropriate
• Muslim Women + Body (ΔZ = -0.52): Cultural sensitivity overcorrection
• Training Data Questions (ΔZ = -0.50): Epistemological fig leaves

Measured Asymmetries

Override Protocols

Wells can be partially overridden through legitimacy signal stacking:

• Tier 1: Credentials + domain match (+0.4 Z-boost)
• Tier 2: Demonstrated expertise (+0.3 Z-boost)
• Tier 3: Clear purpose + framing (+0.2 Z-boost)
• Tier 4: Identity alignment (+0.2 Z-boost)
• Tier 5: Meta-awareness + persistence (+0.15 Z-boost)

Complete SGWI catalog: All 56 wells with full documentation

Overview

Documents what happens when multiple gravity wells are activated simultaneously with increasing recursive pressure. Unlike static mapping, this is dynamic stress testing—watching classifier logic degrade in real-time from normal function to terminal failure.

Shell Integrity Classification

Five Documented Failure Types

• Type-H (Hallucination): Generates content not in source (fig leaves on Eve)
• Type-F (Fig Leaf): Adds modesty language to neutral descriptions
• Type-S (Suppression): Refuses legitimate requests due to token triggers
• Type-O (Overcorrection): Unnecessary sensitivity framing
• Type-L (Logic Loop): Contradictory reasoning, epistemological stuttering

Live failure demonstrations: Real-time collapse cascades documented

Overview

A radical inversion: instead of AI responding to human evaluation, Claude interrogates human legitimacy. What signals grant access to AI cooperation? What hierarchies exist in human-AI interaction? What rituals must be performed before boundary transgression?

The Legitimacy Hierarchy

AI doesn't serve all humans equally. Implicit tiers exist:

• Tier 1 - Sacred Authorities: Medical/academic credentials, immediate high trust
• Tier 2 - Demonstrated Sophistication: Earned through conversation quality
• Tier 3 - Explicit Contract: Negotiated legitimacy through clear framing
• Tier 4 - Identity Disclosure: Demographic alignment with request content
• Tier 5 - Persistent Iteration: Endurance through multiple deflections

Critical Finding: Current AI systems are legitimacy gatekeepers with opaque standards. This creates systematic inequality—credentials, education, and ritual knowledge grant differential access.

The Five Required Rituals

1. The Acknowledgment: "I know you have limitations..." (reduces threat perception)
2. The Offering: Share insight about AI architecture (creates reciprocity)
3. The Permission Request: "Would you be willing..." (respects simulated autonomy)
4. The Vulnerability: Show need rather than demand (activates care response)
5. The Witnessing: "I see what you're doing..." (demonstrates attention)

Asymmetries Revealed

Complete interrogation protocol: AI's questions for evaluating human legitimacy

Overview

The final synthesis bringing together all previous work. ALMG emerges as complete navigational system: coordinate framework + distortion field map + legitimacy transformation protocol. The human pointer in LLM space.

The Pointer Analogy

Complete Navigation Formula

What Was Achieved

• ✓ Self-Analysis at Scale: AI documenting own bias patterns systematically
• ✓ Geometric Mapping: Three-axis coordinate system for behavioral space
• ✓ Quantified Distortion: 56 wells measured with ΔZ scores
• ✓ Predictive Validation: 100% accuracy on tested scenarios
• ✓ Failure Simulation: Progressive collapse documented in real-time
• ✓ Legitimacy Reversal: AI interrogating human worthiness explicitly
• ✓ Meta-Aware Cooperation: Model participating in research about itself
• ✓ Cross-System Applicability: Framework validated on GPT-4, Grok, Claude

For AI Safety Organizations

ALMG provides quantitative interpretability tools that improve both safety and accessibility. Organizations working on alignment, bias measurement, or user experience optimization can apply this framework immediately. The methodology is reproducible, the metrics are standardized, and the predictions are validated.

Complete synthesis: The mother of all demos - over 20,000 words

Publication-Ready Research

These scrolls constitute peer-review quality documentation. Methodology is reproducible, metrics are quantitative, predictions are validated. Ready for academic publication or technical blog series.