hallucination circuit breaker

The problem

LLMs frequently invent hyper-specific facts, fake citations, and outdated prices with high confidence. These hallucinations are often impossible to spot until a client, lawyer, or customer complains about the inaccuracy.

What it does

• Intercepts every factual assertion before delivery and breaks them into a granular claim inventory.
• Triages claims using a risk-decision table to score them based on falsifiability and consequence.
• Applies a proportional verification ladder, ranging from arithmetic recomputation to primary-source corroboration.
• Appends a hard confidence ledger to every response, categorizing statements as verified, derived, believed, or unverified.
• Identifies signature hallucination patterns like phantom citations, summary invention, and category bleed.

Why this beats prompting it yourself

Standard prompts like "be accurate" or "check your work" are easily ignored by the model during the generation phase. This skill implements a verification-first operating system that forces the agent to audit itself as a hostile fact-checker before shipping any text. It replaces vague confidence with a disciplined, multi-phase verification protocol that the model cannot bypass.

Use cases

• Verifying medical, legal, or financial facts before inclusion in professional reports.
• Auditing product specifications and pricing lists to ensure they match current vendor data.
• Summarizing complex technical papers where compression artifacts often breed invented details.
• Fact-checking competitive analysis to prevent "consensus fabrication" and category bleed.

Known limitations

If browsing tools are unavailable, the skill cannot perform external source retrieval and will default to marking memory-based facts as unverified.