Anti-Hallucination v2 Cognitive Sandbox

Problem

In research writing and code generation, the main risk is not total ignorance. The model can present unverified APIs, citations, dates, or facts as certain. In long sessions, one unchecked error can propagate into later conclusions, while generic disclaimers leave no auditable trail.

How It Works

The skill turns hallucination control into an operational workflow:

• Before output, query the cognitive sandbox with verify_fact.py query against verified_knowledge.json; if a fact is verified, mark it 🟢.
• Annotate numbers, dates, people, organizations, and non-standard library APIs with 🟢🟡🔴 confidence labels, and use uncertainty templates instead of completing unverifiable content.
• In code tasks, add API risk notes after code blocks and run pre-generation checks.
• After self-review, record issues such as fabricated_citation, invented_api, or false_fact through hallucination_logger.py.
• When the user supplies authoritative sources, promote 🟡/🔴 facts to 🟢 with verify_fact.py add, creating a reusable verified-knowledge base across sessions.

Boundaries

It suits technical documentation, research reports, third-party library usage, and tasks requiring traceable facts. The skill does not eliminate hallucination or replace source verification; when authoritative user-supplied evidence is missing, uncertain content should still be explicitly flagged.