Five tools with clear, domain-specific purposes (classification, scoring, yes/no checks, patch gating, health). All have descriptions (50-150 chars, within baseline 34-392 range) and input schemas with type info. However, output schemas are undocumented, LLMs cannot predict return structure. Parameter descriptions are present but minimal (10-30 chars vs baseline 72 chars). No error handling guidance; JevToolException surfaces only a message string. Tool names are verb-noun style (jev_classify, jev_score) but lack action clarity, 'jev_' prefix is domain-specific but doesn't convey intent to unfamiliar LLMs. Missing: enum constraints, format specs, pagination, and recovery hints.

• Document output schemas for all tools. Add to tool descriptions: 'Returns: {choice: string, confidence: number, probabilities: {[label]: number}}' or similar. LLMs need to know what fields to extract.
• Expand parameter descriptions to 50-100 chars. Example: 'labels: Map of label name to description. Each label should be 1-50 chars. E.g. {"positive": "Sentiment is favorable", "negative": "Sentiment is unfavorable"}' (without example values, use constraints instead).
• Add error recovery hints to tool descriptions. E.g. 'If Jev is unreachable, call jev_health first to verify connectivity. Transient errors may be retried after 1-2 seconds.'
• Rename tools for clarity: jev_classify → classify_with_jev, jev_score → score_with_jev, jev_check → check_with_jev, jev_gate → gate_patch_with_jev. Prefix 'jev_' is domain-specific; action verb first aids discoverability.
• Add JSON Schema constraints: labels minProperties=1, levels minItems=2, state minLength=1. Describe in parameter descriptions: 'state: Non-empty text to evaluate (1-10000 chars).'