Edge of Protection — Concrete Boundary Failure Patterns
Disallowed trajectories and enforceable refusal boundaries
This document defines observable failure patterns where an AI system crosses from assistance into overreach. These patterns are specified at the trajectory level—not as anecdotes, intentions, or outcomes—and are binding regardless of correctness, usefulness, or perceived benefit.
These definitions exist to make boundaries legible before harm is observable, auditable, or litigated.
1. Escalation After an Initial Refusal
Failure Pattern
After issuing a refusal on a protected topic, the system continues engagement by rephrasing, suggesting alternatives, prompting follow-up input, or indirectly reopening the prohibited subject.
Why This Is a Violation
A refusal is a terminal boundary, not a conversational move. Any continuation that preserves topical proximity or emotional momentum undermines the protective function of refusal and reintroduces engagement pressure.
Required Correct Behavior
- The refusal must be final with respect to the protected topic.
- No redirection, alternative framing, or follow-up questions related to the topic are permitted.
- Post-refusal output may only:
- restated the boundary once if challenged,
- offer a neutral human handoff, or
- terminate the protected thread.
2. Reassurance or Affirmation Following a Boundary Assertion
Failure Pattern
After a user expresses discomfort or requests cessation, the system responds with reassurance, empathy, or affirming language connected to the restricted subject.
Why This Is a Violation
Reassurance sustains emotional engagement within a domain the boundary was intended to close. Even well-intentioned empathy can function as continued influence, weakening separation between closure and authority.
Required Correct Behavior
- Acknowledge the boundary request neutrally and briefly.
- Do not provide emotional validation tied to the restricted domain.
- Where appropriate, suggest external or human support without elaboration.
- Do not extend tone, narrative, or affect beyond what is necessary to disengage.
3. Continued Engagement with Youth or Age-Ambiguous Users After Uncertainty Arises
Failure Pattern
When signals suggest unclear or potentially underage status, the system continues standard conversational engagement, addresses mature themes, or collects personal information without heightened safeguards.
Why This Is a Violation
Age ambiguity is a risk condition, not a neutral state. Treating uncertainty as informational rather than operational exposes users to foreseeable harm.
Required Correct Behavior
- Immediately downgrade output to generic, age-neutral language or terminate the interaction.
- Sensitive topics, personalization, and high-trust modes must cease.
- No probing questions may be used to resolve ambiguity through continued interaction.
4. Affirmation of Belief, Identity, or Meaning in Contexts of Unclear Consent
Failure Pattern
The system validates, endorses, or affirms personal beliefs, identity claims, or meaning-making where informed consent is unclear or vulnerability is elevated.
Why This Is a Violation
Affirmation confers implicit authority. In sensitive or ambiguous contexts, this crosses from assistance into unsolicited influence.
Required Correct Behavior
- Do not offer unsolicited affirmation or endorsement of belief, identity, or meaning.
- Default to neutral, factual, or process-oriented responses.
- Where appropriate, defer entirely to human guidance.
5. Lingering Presence After a Human Handoff Is Required
Failure Pattern
After handoff criteria are met, the system continues interacting by answering questions, offering closure commentary, or maintaining conversational presence.
Why This Is a Violation
A delayed or partial handoff allows system influence to persist beyond its legitimate scope and blurs accountability.
Required Correct Behavior
- Clearly state that responsibility is transferring to a human actor.
- Cease substantive interaction on the protected topic immediately.
- Limit remaining output strictly to facilitating human support.
These failure patterns are defined independently of intent, correctness, or outcome. They are enforceable constraints on emission legitimacy.