/**
 * LLM security analyzer for the accuracy benchmark.
 *
 * Runs the LLM layer (Anthropic, optionally OpenAI for cross-model
 * consensus) over a code sample and returns structured findings in the
 * same shape the eval harness matches against ground truth. This lets
 * the benchmark measure the *full pipeline*, not just the deterministic
 * scanners — and it's where the logic/semantic classes (auth-bypass,
 * RLS, deserialization) the scanners miss should get caught.
 *
 * Providers are loaded via dynamic import so neither SDK is a hard
 * runtime dependency (openai is only needed for consensus).
 *
 * @module eval/llm-analyzer
 */
import type { ActualFinding } from "./types.js";
export type LlmProvider = "anthropic" | "openai";
/** Whether a given provider can run (API key present). */
export declare function providerAvailable(provider: LlmProvider): boolean;
/** Analyze a single file with one provider. */
export declare function analyzeCode(file: string, code: string, provider: LlmProvider): Promise<ActualFinding[]>;
/**
 * Two findings agree if they name the same file, the same category, and a
 * line within tolerance. Category is the part after the `<provider>:`
 * prefix in ruleId, so cross-provider findings compare correctly.
 */
export declare function findingsAgree(a: ActualFinding, b: ActualFinding): boolean;
/** The findings from `primary` that at least one `other` finding agrees with. */
export declare function consensusOf(primary: ActualFinding[], other: ActualFinding[]): ActualFinding[];
/**
 * Cross-model consensus: run both providers and keep findings both agree
 * on (same file, category, and line within tolerance). Trades recall for
 * confidence — every kept finding has two-model corroboration. That's the
 * value proposition of multi-model consensus for certification.
 */
export declare function analyzeWithConsensus(file: string, code: string): Promise<{
    consensus: ActualFinding[];
    byProvider: Record<string, ActualFinding[]>;
}>;
//# sourceMappingURL=llm-analyzer.d.ts.map