Task Boundary Auditor

Problem

LLM boundary violations often do not surface as errors. Instead, the model returns a well-formatted, coherent answer that is still fundamentally wrong. This is the “sophisticated hallucination” risk: counterfactual reasoning, zero-omission auditing, precise verification, real-time closed-loop control, or physical intuition can all produce polished outputs without the required causal model, formal tools, or human accountability. The skill performs a pre-execution audit to prevent high-risk subtasks from being assigned to the LLM as the primary executor.

How It Works

The skill classifies tasks into Safe Zone, Restricted Zone, and Prohibited Zone, then audits them using detection signatures:
- Prohibited Zone: counterfactual reasoning, precise verification, real-time closed-loop, physical intuition, conceptual extrapolation, ethical adjudication, zero-omission audit, and extreme prediction.
- Safe Zone: text generation, translation and conversion, information extraction, summarization, and formatting.
- Restricted Zone: short-chain causality, numerical estimation, draft proposals, and known pattern matching.
For mixed tasks, it decomposes the request into subtasks, classifies each one separately, and outputs a violation block, restricted routing, or safe clearance result.

Boundaries

This skill is useful for auditing high-risk requests, decomposing complex tasks, and labeling what the LLM can and cannot own. It does not replace professional tools, rule engines, statistical models, or human judgment. In legal, medical, financial, or other professional domains, human review should still be recommended even when the task is technically feasible. Audit conclusions must not be downgraded because of user urgency, vague phrasing, or pressure for a final answer.