Two tools with complete JSON schemas and basic descriptions, but significant gaps in parameter documentation and output schema clarity. Tool names lack action verbs (evaluate_* is weak). Parameter descriptions are minimal or absent. No error handling guidance. Output schemas are not documented, LLMs cannot predict response structure. No tool annotations (readOnlyHint, etc.). Batch tool has severely underdocumented nested array structure.

• Rename tools to action-verb form: 'compute_dpo_loss' (for single pair) and 'compute_batch_dpo_loss' (for batch). This clarifies intent and matches LLM naming conventions.
• Add full parameter descriptions to inputSchema. Example: 'pi_w': {'type': 'number', 'description': 'Policy log probability for chosen response (typically negative, e.g., -2.5)'}. Do this for all four log-prob params and beta.
• Document output schema in tool description. Add: 'Returns object with fields: reward_chosen (float), reward_rejected (float), reward_margin (float), preference_probability (float), dpo_loss (float), accuracy (0 or 1).'
• For evaluate_batch, define the 'pairs' array item schema explicitly: items: {type: 'object', properties: {pi_w: {type: 'number'}, ref_w: {type: 'number'}, pi_l: {type: 'number'}, ref_l: {type: 'number'}}, required: ['pi_w', 'ref_w', 'pi_l', 'ref_l']}.
• Add input validation and error messages. Reject non-negative log probabilities with: 'Invalid pi_w: {value}. Log probabilities must be negative (e.g., -2.5). Got {value}.' This enables LLM self-correction.