Loading ToolBench
MCP server for Structured Cognition Protocol — 6 middleware tools that diagnose and contain reasoning amplification in AI agent systems
Server has 6 well-intentioned tools with detailed descriptions (100-300 chars each) and comprehensive parameter schemas using Zod. However, tool names lack action verbs (bullwhip_diagnose, anchor_classify, logic_sequence, mesh_simulate, gate_validate, sc_pipeline are domain-specific but not verb-noun patterns). Schemas are present and typed, but output structures are not documented in the tool definitions, only mentioned in comments ('returns 2 blocks'). Error handling guidance is embedded in descriptions but not formalized. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Parameter descriptions are strong (72+ chars average), but output schema documentation is missing from the tool registration itself.
Classify input signal before acting — separates Action (safe to proceed), Observation (log only, don't act), and Ambiguous (stop, clarify first). Catches noise, hedging language, and uncertainty before they trigger wrong tool calls.
Diagnose Cognitive Bullwhip Effect — scans agent decision history for amplification patterns where small errors compound into large failures. Returns severity score, origin layer, pattern type, and recommended fix. Run this FIRST when agent outputs are erratic, inconsistent, or failing in ways you can't trace.
Validate final decision against governance rules — checks confidence floor, risk thresholds, and custom principles. Blocks or escalates decisions that violate policy. Produces audit trail.
Enforce structured reasoning: Context -> Retrieval -> Analysis -> Action. Prevents step-skipping and reasoning drift. Every step must produce output before the next begins. Checks historical consistency.
Simulate action impact on connected systems — estimates risk score, identifies secondary effects, and flags structural risks. Prevents myopic optimization where a local fix breaks downstream systems.
Tool names lack action verbs. Names like 'bullwhip_diagnose', 'anchor_classify', 'logic_sequence' are domain-specific but do not follow verb_noun convention (get_, create_, search_, etc.). LLMs rely on verb prefixes to infer intent before reading descriptions.
Output schemas are not documented in tool definitions. Comments mention 'returns 2 blocks' (human report + JSON), but the actual structure of those blocks is not declared in the tool schema. LLMs cannot plan downstream operations without knowing what fields to expect.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are marked READ_ONLY in comments, but this is not formalized in the MCP schema. Agents cannot determine which tools are safe to retry or have side effects.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Run the full Structured Cognition pipeline: SignalAnchor -> LogicStack -> CausalMesh -> PrincipleGate. Single input, complete diagnosis. Returns per-stage results and final recommendation.
Error handling is descriptive but not formalized. Descriptions include recovery guidance ('let the user pick from candidates'), but there is no structured error response format or error classification (retryable vs user-fixable vs fatal).
Parameter relationships are undocumented. bullwhip_diagnose requires 'decision_log OR raw_events' and 'variance_strategy REQUIRED when raw_events is used', but these mutual exclusivity rules are only in descriptions, not enforced in schema or documented as dependencies.