# Data Analysis Prover MCP for AI Agents AI Agent Connect > Data Analysis Prover is a statistical auditing tool for your AI agent. It forces your AI to stop making sloppy data claims and start thinking like a senior statistician. It checks for sample bias, causal fallacies, skewed distributions, and dishonest charts before you present your findings to stakeholders. ## Overview - **Category:** statistics - **Price:** Free - **Endpoint:** https://edge.vinkius.com/vk_preview_z0UOTGMw5z9e0sSzQndQJ5ZX9DZMH1doH3kzey3a/ai-agent-connect - **Tags:** data-analysis, statistics, sample-size, causation, p-value, effect-size, visualization ## Description You know the feeling. You ask your AI to look at a dataset and summarize the trends, and it hands back a slide deck full of significant correlations that are actually just noise. It might tell you that a marketing campaign caused a sales spike when the sample size was tiny, or it might use a mean average on a heavily skewed distribution, making the numbers look way better than they actually are. This Connector changes that dynamic. Instead of letting your AI take the path of least resistance, this tool forces it to pass every claim through a rigorous peer review process. It acts as a guardrail that demands to see the sample size, the power analysis, and the actual distribution shape before it is allowed to make a statement. It catches when your agent is trying to sell you a p-value without an effect size, or when it is trying to hide a small trend behind a truncated Y-axis. When you connect this to your workflow through the Vinkius catalog, you are essentially hiring a senior statistician to sit next to your AI. It stops the hallucination of significance and ensures that the insights you share with your team are actually backed by sound methodology. You get results that will not get you laughed out of a meeting by a real data scientist. ## Tools ### validate_data_analysis Pass a statistical claim or research interpretation to the Connector to get a rigorous peer review of the methodology. It checks for sample validity, causal logic, and visual honesty to ensure your data is not misleading. ## Prompt Examples **Prompt:** ``` I have a dataset of 500 users. Does the 5% increase in signups from the new button mean it actually worked? ``` **Response:** ``` ### Statistical Review **Status: ANALYSIS_PROVEN** Your analysis meets the required rigor. Here is the breakdown: * **Sample Quality:** N=500 provides sufficient power for this effect size. * **Causality:** You have controlled for previous user activity. * **Significance:** p=0.03 with a **Cohen's d of 0.45** (medium effect size). * **Visualization:** Ensure your Y-axis starts at 0 to avoid exaggeration. **Conclusion:** The result is statistically significant and practically meaningful. ``` **Prompt:** ``` The AI says the new pricing caused a 10% drop in churn. Is that a safe claim? ``` **Response:** ``` ### Statistical Review **Status: CORRELATION_CONFUSED** I cannot confirm this claim as 'caused' based on the current data. * **The Issue:** You are making a causal claim from observational data. * **Missing Info:** We haven't accounted for seasonality or concurrent competitor moves. * **Fix:** Rephrase the claim to: *'Signups are associated with a 10% drop in churn.'* To prove causation, we would need an experimental design or a more rigorous causal inference model. ``` **Prompt:** ``` Can you summarize the average salary for our remote workers? ``` **Response:** ``` ### Statistical Review **Status: DISTRIBUTION_IGNORANCE** I cannot provide a simple mean for this dataset. * **The Issue:** The salary distribution is heavily right-skewed (top 5% earn significantly more). * **The Fix:** I will report the **Median Salary ($62,000)** and the **Interquartile Range ($50k - $75k)** instead of the mean. Using the mean in this case would overstate the typical worker's earnings. ``` ## Capabilities ### Verify sample size and power The Connector checks if your sample size is large enough to support the statistical claims you are making. ### Detect causal fallacies It identifies when your agent incorrectly claims causation from simple observational data. ### Check for distribution skew The tool forces the AI to check if data is skewed before it chooses a mean or a median. ### Calculate effect sizes It requires the AI to report practical significance like Cohen's d instead of just a p-value. ### Audit chart honesty It flags misleading visualizations like truncated Y-axes or dual scales that distort your data. ### Identify confounders The Connector forces the agent to look for and account for external variables that might bias the results. ## Use Cases ### Verifying marketing campaign ROI A manager asks the AI to prove a new email subject line increased sales. The AI uses validate_data_analysis to flag that the correlation is observational and lacks a control for seasonality. ### Checking product feature retention A product team wants to know if a new feature improved retention. The agent uses the Connector to catch that the sample size (N=12) is too small for any meaningful statistical power. ### Summarizing skewed salary data A researcher is analyzing salary data. The AI tries to report the mean, but the Connector forces it to use the median and IQR because the distribution is heavily right-skewed. ### Auditing executive dashboards A BI analyst is building a dashboard. The AI suggests a 3D chart, but the Connector rejects it as a visualization deception and demands a linear Y-axis starting at zero. ## Benefits - Stop Significance Theater by forcing the AI to report effect sizes like Cohen's d alongside p-values using validate_data_analysis. - Eliminate Correlation Confusion by making the agent identify confounders and distinguish between association and causation. - Prevent Distribution Ignorance by requiring the AI to check for skew and outliers before choosing a test. - Avoid Sample Blindness because the tool forces the agent to report sample size (N) and power analysis for every claim. - Fix Visualization Deception by auditing charts for truncated axes and dual scales to ensure honest data representation. ## How It Works The bottom line is that it turns your AI from a data summarizer into a rigorous statistical reviewer. 1. Provide the AI with your data summary or a specific statistical claim. 2. The Connector reviews the claim against five core axes: sample quality, causality, distribution, significance, and visual honesty. 3. You get a pass or fail report with specific instructions on how to fix methodological flaws. ## Frequently Asked Questions **Can the Data Analysis Prover catch misleading charts?** Yes. It audits your charts for common tricks like truncated Y-axes, dual scales, and distorted proportions to ensure your data remains honest. **Does Data Analysis Prover help with small sample sizes?** Yes. It flags Sample Blindness by forcing the AI to report the sample size (N) and conduct a power analysis to see if the results are actually reliable. **How does Data Analysis Prover handle p-values?** It prevents Significance Theater by requiring the AI to include effect sizes like Cohen's d alongside p-values, so you know if a result actually matters. **Can I use Data Analysis Prover for marketing research?** Absolutely. It is perfect for checking if marketing campaign results are statistically significant or just a result of random noise. **Will Data Analysis Prover make my AI more accurate?** It makes your AI's statistical claims more accurate by forcing it to follow rigorous data science methodologies instead of taking shortcuts. **Can it detect if a correlation is actually a cause?** Yes, it identifies Correlation Confusion by checking for confounders and distinguishing between observational associations and experimental evidence. **Why is p<0.05 not enough?** p-value measures probability, not magnitude. Cohen's d: 0.2=small, 0.5=medium, 0.8=large. A p<0.001 with d=0.05 is trivial. Report effect size + 95% CI + practical significance. **When can I say 'causes' vs 'associated with'?** Only RCTs establish causation. Observational studies show association. Control confounders, test reverse causality, check dose-response. Even then: 'associated with' unless experimental design. **Why is the mean misleading on skewed data?** Income example: mean $65K, median $45K. The mean is pulled by outliers. Right-skewed data: median represents 'typical' better. Test normality with Shapiro-Wilk before choosing parametric tests.