Veritas provides 6 well-structured tools with clear naming, comprehensive descriptions, and proper input schemas. All tools follow verb-noun patterns (fetch_, get_, search_, submit_, reject_, done). Descriptions are detailed and contextual (150-300 chars), explaining WHAT the tool does, WHEN to use it, and prerequisites. Input schemas use proper JSON Schema with type declarations and parameter descriptions. However, output schemas are entirely undocumented, the rubric requires documented return types for A+ tools, and this gap prevents a higher score. Error handling is implicit (via tool descriptions) but lacks explicit recovery guidance. Tool composition is excellent, tools chain naturally (fetch_signals → submit_causal_chain) and use consistent naming conventions. No security issues detected (no exposed credentials). The main limitation is the absence of structured output documentation and error handling patterns.

• Document the output schema for each tool. At minimum, specify: fetch_signals returns {results: [{field1, field2, ...}], total_count: number, next_cursor?: string}; get_narrative_context returns {narrative_id, posts: [...], metadata: {...}, temporal_details: {...}}; search_historical_signals returns {signals: [...], domains_searched: [...], total_count, next_cursor}; submit_causal_chain returns {chain_id, status: 'submitted'|'validation_error', validation_errors?: [...]}; reject_correlation returns {status: 'rejected'}; done returns {analysis_complete: true}.
• Add explicit error handling to tool descriptions. Example for fetch_signals: 'Errors: adapter_not_found (try list_adapters first), invalid_date_range (start_date must be before end_date), network_timeout (retry with smaller date range). All errors are retryable.'
• Add pagination parameters to search_historical_signals: limit (1-100, default 50), offset (default 0). Return results with total_count and next_offset so agents can paginate.
• For submit_causal_chain, document validation rules: 'Each chain link must include evidence from actual data retrieved via fetch_signals or search_historical_signals. Submissions with unsubstantiated evidence will be rejected with validation_errors.' This prevents hallucination.
• Add a dry-run parameter to submit_causal_chain (dry_run: boolean, default false) so agents can validate chains before persisting. When dry_run=true, return validation_errors without storing the chain.
• Document which fields are required vs. optional in submit_causal_chain.chain[].*, for example, is timestamp required for every node, or only the root cause?

Inferred effective spec: <=2025-11-25.

• Relies on HTTP+SSE-only transport (deprecated) - migrate to Streamable HTTP