Loading ToolBench
MCP server for longitudinal user modeling with 14 tools for managing living user models, beliefs, patterns, and psychological structures
Significant gaps across naming, descriptions, and schemas. Tool names lack clear action verbs (e.g., 'cognition_soul_reflect', 'cognition_shadow_map' are vague and domain-specific rather than action-oriented). Parameter descriptions exist but are minimal. No output schemas documented. Error handling and recovery guidance absent. The domain (user modeling, consciousness reflection) is novel but the tool interface does not follow production MCP patterns.
Add a new belief to a model
Show belief graph structure and connections
Query beliefs by domain or state
Track value drift and metamorphosis over time
Create a new user model
Pulse model with observations
Full model portrait with belief topology
Tool names lack action verbs and are domain-jargon heavy. 'cognition_soul_reflect', 'cognition_shadow_map', 'cognition_self_topology' do not clearly signal what action the LLM should take. LLMs rely on verb_noun naming (get_, create_, update_, search_) to infer intent.
Descriptions are too short (most under 20 chars). Examples: 'Pulse model with observations' (29 chars), 'Show model health vitals' (24 chars), 'Show soul reflection' (19 chars). LLMs cannot determine when/why to select tools without context.
No output schemas documented. Tools return unstructured data (e.g., 'belief topology', 'consciousness reflection'). LLMs cannot plan downstream calls or extract structured fields without knowing response shape.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 33 | <=2025-11-25 | v2 |
Show model health vitals
Show decision fingerprint and cognitive patterns
Predict user preference for an item based on model
Show self-concept topology with confidence peaks and valleys
Show shadow map with projections and blindspots
Simulate decision outcome in a scenario
Perform soul reflection on model consciousness
No error handling or recovery guidance. Tools do not document what happens on invalid model_id, malformed observations, or out-of-range confidence values. LLMs receive no actionable error messages to self-correct.
Parameter descriptions are minimal and lack constraints. 'confidence' (0.0-1.0) is stated in description but not enforced via schema minValue/maxValue. 'domain' and 'state' accept free-form strings instead of enums. LLMs will hallucinate invalid values.