MCP Server for Bayesian Network Exact Variable Elimination Skill - Pure Python Standard Library implementation of Exact Inference in Bayesian Networks using factor reduction, pointwise multiplication, and marginalization
Two tools with minimal documentation. add_factor and query have basic descriptions (50-80 chars) but lack parameter-level descriptions entirely. Input schemas present but sparse: no type constraints, enums, ranges, or validation guidance. No output schema documented. No error handling guidance. Tool names are somewhat descriptive (verb_noun pattern) but descriptions do not explain WHEN to use each tool or what happens on failure. No mention of idempotency, side effects, or prerequisites. Schemas lack the detail needed for LLM reasoning.
Add a factor (conditional probability table) to the Bayesian Network
Compute P(target_var | evidence) via Variable Elimination algorithm
No parameter descriptions. 'variables', 'table', 'target', 'evidence', 'order' lack any explanation of format, constraints, or valid values. LLMs cannot infer whether 'table' is a dict of tuples or nested objects, or what 'order' should contain.
No output schema documented. query returns {'target': str, 'distribution': dict} but the structure of 'distribution' (keys are variable values, values are probabilities) is not specified. add_factor returns {'status': str, 'total_factors': int} without explanation.
No error handling or recovery guidance. If query fails (invalid target, missing evidence variable, empty elimination order), the response is {'error': 'Unknown tool: ...'} with no actionable next step. No distinction between retryable and fatal errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
Input schemas lack type constraints and validation rules. 'variables' is array but element type not specified (string assumed). 'table' is object but key/value types not documented. 'order' is array with no element type or length constraints. No enums, patterns, or ranges.
Tool descriptions are generic and lack context. 'Add a factor...' and 'Compute P(target_var | evidence)...' do not explain when to call each tool, what prerequisites exist, or what side effects occur. No mention that add_factor modifies state or that query is read-only.