Single tool with critical definition gaps. Tool name 'orchestrate_speculative_run' lacks action verb clarity (should be 'run_speculative_decoding' or similar). Description is present (67 chars) but generic and lacks WHEN/WHY context. Input schema visible in code but parameters lack descriptions entirely, 'prompt', 'draft_tokens', 'draft_length' have no guidance on format, constraints, or purpose. Output schema undocumented; response fields are inferred from client code only. No error handling, validation guidance, or recovery paths. Tool definition is partially visible in mcp_server.py but parameter descriptions are completely absent from the schema registration.

• Add detailed descriptions to all input parameters. Example: 'draft_tokens: Array of token objects from draft model. Each object must have token (string), p_draft (0.0-1.0), p_target (0.0-1.0). Required for acceptance calculation.'
• Document the output schema explicitly. List all response fields with types and meanings: orchestration_id (string), tokens_accepted_count (integer), effective_throughput_speedup (string formatted as 'Nx'), etc.
• Rename tool to 'verify_draft_tokens' or 'run_speculative_decoding' to start with a clear action verb and improve LLM intent recognition.
• Expand tool description to 100-150 chars explaining: what speculative decoding is, when to call this tool (e.g., 'to evaluate draft model token acceptance rates'), and what the speedup metric represents.
• Add input validation and error handling. Define constraints: draft_tokens array length 1-100, probability values 0.0-1.0, draft_length 1-50. Return actionable errors if constraints violated.