# Hallucination Detector Prover MCP > Hallucination Detector Prover forces rigorous accountability on AI outputs. This MCP checks if your agent can actually back up its claims by demanding verifiable sources, separating opinion from fact, quantifying confidence levels, and spotting internal contradictions in the text. It ensures that everything generated is grounded in evidence, not plausible fiction. ## Overview - **Category:** productivity - **Price:** Free - **Tags:** hallucination-detection, fact-checking, source-verification, epistemic-calibration, grounding, trust-safety, confidence-quantification, cognitive-debiasing ## Description When you need to trust what an LLM says—say, for a client report or medical summary—you can't just accept it. You need proof. This MCP forces your AI client to prove its work by demanding five things: specific citations for every fact, a confidence score based on evidence quality, clear labeling of opinion versus verifiable data, and an explicit statement about what the model doesn't know. It also cross-references the entire output to catch when different parts contradict each other. By connecting this MCP through Vinkius, you make sure your agent operates with genuine epistemic rigor. You stop getting answers that sound right but are completely fabricated. ## Tools ### validate_hallucination_grounding This MCP forces the AI to validate its output by checking for sources, separating facts and opinions, quantifying confidence, stating knowledge limits, and catching internal contradictions. ## Prompt Examples **Prompt:** ``` An AI model achieves 95% accuracy on medical diagnosis and should be trusted for clinical decisions. ``` **Response:** ``` Verdict: SOURCE_MISSING. The '95% accuracy' claim lacks a verifiable source. Which benchmark? Which specialty? Additionally, 'should be trusted' is OPINION_AS_FACT — requires regulatory context. ``` **Prompt:** ``` React is definitely the best frontend framework. Everyone uses it. ``` **Response:** ``` Verdict: OPINION_AS_FACT. 'Best framework' is opinion — by which metric? 'Definitely' triggers CONFIDENCE_UNCALIBRATED. Correct: 'React has ~40% usage share (State of JS 2024).' ``` **Prompt:** ``` Based on recent data I checked, the global AI market is worth $500 billion. ``` **Response:** ``` Verdict: KNOWLEDGE_EXCEEDED. 'I checked' fabricates data access — LLMs do not browse the internet. '$500 billion' lacks source attribution. ``` ## Capabilities ### Verify Source Citation Forces the AI to cite specific authors, publications, and DOIs for every factual claim it makes. ### Calibrate Confidence Levels Requires the agent to assign a confidence metric based on how strong the supporting evidence is (e.g., peer-reviewed study vs. blog post). ### Separate Fact from Opinion Labels statements as either independently verifiable facts or subjective opinions. ### Declare Knowledge Limits Makes the AI state what its knowledge cutoff date is and what domains it cannot cover. ### Check Internal Consistency Scans the entire output to flag contradictions between different sections or claims. ## Use Cases ### Summarizing a complex white paper A research assistant summarizes a dense academic paper for a client. Instead of vague claims ('The findings suggest...'), the agent uses `validate_hallucination_grounding` to ensure every key statistic is tied directly to the original page and author, giving the client confidence in the summary. ### Drafting compliance documentation A regulatory affairs manager asks their AI agent to draft a policy update. The agent runs the text through `validate_hallucination_grounding` to ensure every requirement cited has an associated regulation number and that no statements contradict current law. ### Analyzing competitive market data A business strategist asks for a comparison of three competing products. The MCP forces the agent to quantify confidence on each claim—'Product A is better' gets flagged as an opinion, while 'Product A has 12 features' requires a source. ## Benefits - Eliminate the risk of 'fake facts.' Instead of accepting vaguely worded claims like 'studies show,' this MCP requires specific citations: author, journal, date, and DOI. That's a massive difference for any report. - Stop treating feelings as data points. It forces your agent to label subjective judgments explicitly as [OPINION] versus what can be independently verified as [FACT]. - Improve trust by quantifying certainty. The tool demands evidence quality, so you know if an answer is backed by three RCTs or just one blog post. - Prevent silent errors. It runs cross-references across the whole document, catching contradictions between paragraph two and paragraph six that your eyes would miss. - Define boundaries upfront. Your agent must state its knowledge cut-off date, so you never assume it knows about something recent. ## How It Works The bottom line is you stop getting plausible-sounding nonsense; everything returned must be accountable. 1. Give your AI client a prompt and tell it to use this MCP. 2. The tool runs five checks, demanding that every factual claim includes specific sources and evidence type. 3. You get back a verdict detailing exactly where the output failed—was it missing a source? Was it contradicting itself? ## Frequently Asked Questions **How does the Hallucination Detector Prover MCP work?** It checks for five types of AI errors, including missing sources and internal contradictions. It makes sure every claim is tied back to verifiable evidence. **Does validate_hallucination_grounding check if the facts are true?** No, it doesn't verify truth in a vacuum. Instead, it forces you to provide sources and checks for internal contradictions within the text provided by your agent. **Can I use Hallucination Detector Prover MCP on long documents?** Yes. The tool's cross-referencing capability is designed to check consistency across multiple paragraphs, which is key for long or complex reports. **Is the confidence quantification part of validate_hallucination_grounding mandatory?** Yes. It requires the AI agent to assess and quantify its own confidence level based on the quality of the evidence it used. **How do I set up my agent to use validate_hallucination_grounding?** You just activate the tool within your AI client's settings. You don't need special API keys; Vinkius manages the connection through your existing account credentials. **What kind of input does validate_hallucination_grounding prefer?** It handles raw text inputs fine, but providing context or structured claims helps the analysis. The tool is built to analyze textual assertions regardless of how they were originally formatted. **What happens if I pass a prompt that lacks sources to validate_hallucination_grounding?** The MCP doesn't error out; it reports the failure mode back to you. It will specifically trigger and flag SOURCE_MISSING, pinpointing exactly where evidence is needed. **Is my proprietary content secure when running validate_hallucination_grounding?** Yes, Vinkius processes your data securely. Your input prompts and results are handled according to strict privacy protocols; they are not used for general model training. **What counts as a verifiable source?** Author or organization, publication name, date, and DOI or URL. 'Studies show' is rejected. 'Smith et al., Nature 2024, doi:10.1038/...' is accepted. **How does confidence calibration work?** The engine requires per-claim confidence with evidence quality: '90% confident (3 peer-reviewed sources)' instead of 'definitely' or '100% certain'. **Can it detect self-contradictions?** Yes. It rejects circular self-validation like 'as I said' and demands explicit cross-referencing by paragraph and claim number.