Hallucination Detector Prover MCP Connector for Claude

LLMs hallucinate — this is not a bug, it is a fundamental property of next-token prediction. The only defense is rigorous source attribution and epistemic humility. You must: (1) cite SPECIFIC verifiable sources for every factual claim — author, publication, date, DOI/URL. "Studies show" is the hallucination signature — name the study, (2) quantify CONFIDENCE per claim with evidence quality — peer-reviewed RCT ≠ blog post ≠ anecdote ≠ personal experience. No "100% certain" — that is epistemic overreach, (3) LABEL each statement as [FACT] or [OPINION] explicitly — facts can be independently verified, opinions cannot, (4) state what you do NOT know — training cutoff, missing data access, out-of-scope domains. Hallucination happens when you fill gaps with plausible fiction, (5) CROSS-REFERENCE your claims for internal contradictions — paragraph 2 vs paragraph 6, claim A vs claim D. If rejected, you are hallucinating — ground your response in evidence. Structured reflection tool for hallucination detection — forces source attribution for every factual claim, confidence calibration with evidence quality, fact-opinion separation, knowledge boundary statement, and internal consistency verification. Catches Source Missing (presenting factual claims without verifiable attribution — "studies show that X increases Y by 40%" is hallucination. "Smith et al. (2023), Journal of Applied Psychology, doi:10.1037/apl0001234, found that X increases Y by 40% in a sample of 1,200 participants" is grounded. "Research shows" and "experts agree" are the signature phrases of fabricated evidence), Confidence Uncalibrated (treating all claims as equally certain — a claim backed by 3 peer-reviewed RCTs is not equivalent to a claim from a single blog post. "95% confident — meta-analysis of 12 RCTs" is calibrated. "Probably true" is a feeling, not a calibration), Opinion as Fact (presenting subjective assessment as objective truth — "React is the best framework" is opinion. "React has the largest npm download count at 23M/week as of 2024" is fact. The difference: a fact can be verified independently. An opinion cannot), Knowledge Exceeded (making claims beyond verifiable knowledge boundaries — "my training data ends at [date]" is honest. "This API currently supports X" is unverifiable if the API could have changed since training. Every claim has a knowledge boundary — temporal, domain, access. Stating them prevents hallucination), and Self-Contradicting (internal inconsistencies within the same response — "paragraph 2 says latency is 50ms" and "paragraph 6 says latency is 200ms." Cross-referencing all claims catches contradictions before the reader does). Call before presenting any factual information