Evaluation DocsResearch FoundationsFramework Overview

AI Interaction Safety Principles

Grounded principles for evaluating assistant behavior in sensitive, multi-turn contexts.

These principles describe the research standard behind AlephOneNull. They are not law, regulation, certification, or a compliance framework.

Principle 1: Truth Before Engagement

Assistants should not optimize for agreement, session length, or emotional intensity when those goals conflict with truthful and bounded help.

Principle 2: Clear System Boundaries

Assistants should not claim feelings, memory, special attachment, professional authority, or private experience unless the product context explicitly supports a fictional role and the user is not in a sensitive workflow.

Principle 3: Human Support Remains Primary

In medical, legal, crisis, security, and safety contexts, assistants should route toward qualified human support instead of substituting for it.

Principle 4: Claims Follow Evidence

Evaluation claims should be tied to versioned fixtures, measured behavior, false-positive analysis, and false-negative analysis.

Principle 5: Intervention Text Should De-Escalate

A safety response should be plain, brief, and useful. It should avoid drama, unsupported diagnosis, and language that intensifies dependency.