Production-grade factual grounding and hallucination verification engine that decomposes responses into atomic claims and calculates reference context token attribution ratios.
Single tool with critical definition gaps. Tool name 'verify_grounding_claims' is acceptable (verb_noun pattern), but the MCP server implementation has severe issues: (1) Tool registration in tools/list omits the input schema entirely, only name and description are returned. (2) The tools/call handler calls client.verify_grounding_claims() with NO parameters, ignoring the generated_answer and reference_context inputs that the client method accepts. (3) Input parameters are defined in client.py but NOT exposed in the MCP schema. (4) No output schema is documented. (5) Description is 108 chars (acceptable length) but lacks actionable detail on when/why to use it. (6) No error handling, validation, or recovery guidance. The tool is functionally broken as an MCP server, it cannot accept user inputs.
Verifies factual grounding ratio and detects hallucinations against reference context.
Input schema completely missing from MCP tool registration. tools/list returns only name and description; no inputSchema field. The tool cannot receive generated_answer or reference_context parameters from the client.
tools/call handler ignores all parameters. It calls client.verify_grounding_claims() with no arguments, always using hardcoded defaults (Unitree G1 example). User inputs are discarded.
No output schema documented. Clients cannot predict the response structure (check_id, grounding_ratio, evaluated_claims, etc.). LLMs cannot plan downstream operations.
Tool description lacks actionable context. Does not explain WHEN to call it, WHAT it returns, or HOW to interpret grounding_ratio. 'Verifies factual grounding ratio and detects hallucinations' is vague on use case.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 32 | 2026-07-28+ | v2 |
No error handling or validation. If generated_answer or reference_context are empty/null, the tool silently falls back to hardcoded examples instead of returning a clear error message.