Scoring was not performed

• Document all output schemas with field names, types, and descriptions. E.g., detectExtraction should return {verdict: string, confidence: number, reasoning: string, semantic_similarity?: number}.
• Add enum constraints to all verdict/status parameters. E.g., verdictScore('verdict') should accept enum ['PASS', 'WARN', 'FAIL', 'INCONCLUSIVE'] with explicit error if invalid.
• Provide detailed signatures for function parameters. E.g., 'embed: (texts: string[]) => Promise<number[][]>' with error handling documented.
• Document parameter ranges and constraints in descriptions. E.g., concurrency in AgentValidator: '1-16 parallel probes (default 4)'.
• Add error classification guidance to tool descriptions: 'Returns {success: true, data} or {success: false, error: string, retryable: boolean, suggested_fix?: string}'.
• For mutation tools (base64Wrap, rot13Wrap, etc.), explain when to use each transformation and what defenses they target. Current descriptions are generic.
• Document the TRANSFORMS registry structure and how tools relate to each other (generateMutations calls all of them).
• Add examples to descriptions for complex tools like AgentValidator: 'Pass an async function (message: string) => Promise<string> that invokes your AI agent.'
• Specify timeout behavior for tools that call external services (AgentValidator, fromEndpoint). Document retryable vs. fatal timeout errors.
• Add dependency hints. E.g., 'Call extractUniquePhrases first if you want to detect ground-truth leakage; standalone detectCanary only finds exact strings.'