Single tool 'orchestrate_speculative_run' has critical definition gaps. Tool name lacks action verb (should be 'run_speculative_decoding' or 'verify_speculative_tokens'). Description is adequate (89 chars, within 10-1024 range) but generic. Input schema is present in the rubric description but NOT visible in actual mcp_server.py code, tools/list response omits schema entirely. The client.py shows parameters exist (prompt, draft_tokens) but mcp_server.py never exposes them to the MCP protocol. Output is unstructured JSON string, not typed fields. No error handling, no validation, no recovery guidance.

• Add inputSchema to tools/list response with full parameter definitions: prompt (string, required, description), draft_tokens (array of objects with token/p_draft/p_target, optional). Use JSON Schema with type and description for each field.
• Rename tool to 'run_speculative_decoding' or 'verify_draft_tokens' to start with action verb and clarify intent.
• Document each parameter: 'prompt' (string, max 2000 chars, the input text for speculative decoding), 'draft_tokens' (array of {token: string, p_draft: number 0-1, p_target: number 0-1}, optional; if omitted, uses default example).
• Return structured response object instead of JSON string: {orchestration_id, prompt_sample, tokens_accepted_count, accepted_sequence, rejected_token, speculative_acceptance_rate, effective_throughput_speedup, latency_reduction_percentage} with explicit types.
• Add validation and error handling: reject draft_tokens with p_draft or p_target outside [0, 1]; return actionable error like 'Invalid probability: p_draft=1.5 must be between 0 and 1' instead of silent defaults.