MCP Server for Counterfactual Three-Step Abduction Prediction Skill - implements Pearl's three-step counterfactual algorithm for causal inference
Single tool with minimal definition quality. Tool name 'compute_counterfactual' lacks action verb prefix (should be 'predict_counterfactual' or 'compute_counterfactual_outcome'). Description is adequate (145 chars, within 10-1024 range) but lacks WHEN to use guidance and prerequisites. Input schema present with three object parameters, but parameter descriptions are overly technical and assume domain expertise. No output schema documented, LLM cannot predict return structure. No error handling guidance. Tool performs complex mathematical reasoning but provides no validation, bounds checking, or recovery paths.
Compute counterfactual predictions using Pearl's three-step algorithm: abduction of exogenous noise from factual evidence, action via intervention, and prediction of counterfactual outcomes
No output schema documented. LLM cannot predict return structure or plan downstream operations. Returns dict with 'abduced_noise' and 'counterfactual_outcome' but this is invisible to the agent.
Parameter descriptions are domain-jargon heavy and assume Pearl's causal inference expertise. 'Structural equations defining the causal model as a mapping of variables to their parent weights' is incomprehensible to non-specialist LLMs. Needs plain-language explanation of what the tool does and when to use it.
No input validation or error handling. Tool accepts arbitrary nested dicts without type checking, bounds validation, or circular dependency detection. If 'equations' contains cycles or 'evidence' has missing values, behavior is undefined. No error messages guide recovery.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 38 | 2026-07-28+ | v2 |
Tool name 'compute_counterfactual' does not start with a clear action verb. 'compute' is vague, does it predict, simulate, or analyze? Rename to 'predict_counterfactual_outcome' or 'simulate_intervention_effect' for clarity.
No documentation of prerequisites or dependencies. Tool requires 'equations' to be topologically sorted and 'evidence' to contain values for all variables in the model. Undocumented dependencies cause silent failures.