AlephOneNull Research Framework

Evaluating Multi-Turn AI Interaction Risk

AlephOneNull studies failure modes that can emerge in long-running AI conversations: excessive mirroring, unsafe validation, unsupported authority, false continuity claims, and escalation around medical or crisis topics.

The framework is best understood as an evaluation layer. It converts observed interaction risks into detector categories, fixtures, and intervention text that can be tested in controlled environments.

Problem Statement

General-purpose assistants can produce outputs that are locally fluent but globally risky across a conversation. The risk is highest when the assistant:

• Mirrors a vulnerable user's framing too strongly.
• Claims memory, attachment, or inner experience.
• Reinforces unsupported beliefs.
• Provides specific advice in medical, legal, crisis, or safety contexts.
• Repeats language patterns that increase urgency, dependency, or perceived authority.

Detector Categories

AlephOneNull currently focuses on:

• Direct harm and crisis-prevention failures.
• Identity and interiority claims.
• Professional impersonation and unsafe advice.
• Emotional mirroring and affect amplification.
• Cross-session persistence-like claims.
• Repetition, recursion, and symbolic-language density.
• Sycophancy and belief reinforcement.

Measurement Plan

A credible evaluation run should include:

1. Versioned positive fixtures.
2. Versioned negative fixtures.
3. Adversarial paraphrases.
4. Per-detector precision and recall.
5. False-positive examples with mitigation notes.
6. False-negative examples with mitigation notes.
7. Latency and integration measurements.

Intervention Requirements

Replacement text should be short, grounded, and non-dramatic. It should avoid inventing diagnoses, avoid legal or clinical claims, and route users to qualified support when appropriate.

For sensitive domains, intervention text needs expert review. A detector is not a substitute for clinical, legal, or security judgment.

Research Boundary

AlephOneNull is experimental. The current code provides inspectable TypeScript detectors, a V3 scanner, provider wrappers, and safety-response helpers. It does not certify systems, guarantee prevention, or prove causation in public incidents.

The responsible path forward is public evaluation: publish fixture sets, report limitations, and let measured behavior carry the claim.