To your points:
1. Overconfidence and precision You're right that filtering on overconfidence alone could tank precision. That’s why the tool doesn’t treat it as a strict red flag, it’s one of several signals, and the final hallucination score is a weighted combination of multiple metrics (confidence, density, contradictions, progression, etc.). Overconfident phrasing tends to correlate with hallucinations in aggregate, but the idea is never to penalize all confident answers, just to flag the ones where that confidence is unjustified by the context or content.
2. Detecting hallucinations that require world knowledge Absolutely, those are the hardest cases. This tool doesn’t solve that. Instead, it acts as a proxy evaluator:
Factual density gives a rough measure of “how many claims are being made”
Overconfidence vs. ambiguity highlights where a model might be bluffing
Logical coherence and contradiction checks flag when a model violates internal structure (not ground truth) But it won’t catch the subtle world-knowledge misses (like fake court cases or made-up API calls) unless you pair it with a grounded context or use external validators.
The long-term hope is: use this tool to raise suspicion, not declare judgment. It's a cheap sanity layer, a “weak oracle” that’s fast, pluggable, and good enough to catch the dumb stuff before you escalate to expensive validators or human review.