Loading ToolBench
MCP Server for Backdoor Criterion Confounder Adjustment Skill - Determines whether a set of covariates Z is a valid adjustment set for identifying the causal effect of treatment X on outcome Y without confounding bias using Pearl's Backdoor Criterion and d-separation.
Single tool with adequate schema but weak descriptions and no error handling guidance. Tool name is clear (validate_backdoor_adjustment) but description lacks actionable context for LLM selection. Input schema is present with proper types and descriptions for graph parameters, but output schema is undocumented. No parameter constraints (enums, ranges) despite accepting structured graph data. Error responses are generic strings with no recovery guidance. Tool does one thing well (validate backdoor criterion) but lacks production-grade polish.
Validates whether a candidate adjustment set satisfies Pearl's Backdoor Criterion for identifying causal effects. Checks that no node in the adjustment set is a descendant of treatment, and that the set blocks all backdoor paths between treatment and outcome.
Output schema undocumented. Tool returns {is_valid_adjustment_set, explanation} but LLM has no formal schema to parse response structure or plan downstream actions.
Error responses lack recovery guidance. Returns bare strings like 'Unknown tool: {name}' with no actionable next steps for the LLM.
Tool description (194 chars) is adequate length but lacks WHEN to use it and prerequisites. Does not explain when backdoor criterion applies vs other causal inference methods.
No input validation or constraints. Accepts arbitrary node/edge lists with no validation that graph is acyclic, nodes are unique, or edges reference valid nodes.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
Parameter descriptions lack format guidance. 'nodes' and 'edges' descriptions do not specify required format (e.g., 'edges must be [source, target] tuples with valid node identifiers').