/** * Deterministic memory-value benchmark. * * This does not pretend to measure LLM quality directly. It measures the part * Squad can verify locally: whether governed memory retrieval reduces context * payload while preserving relevant facts and excluding stale or unsafe memory. */ export type MemoryValueStatus = 'active' | 'superseded' | 'deleted'; export type MemoryValueLoadGuidance = 'ALWAYS' | 'ON-DEMAND' | 'ARCHIVE' | 'NEVER'; export interface MemoryValueFact { id: string; title: string; content: string; loadGuidance: MemoryValueLoadGuidance; status: MemoryValueStatus; tags: string[]; } export interface MemoryValueTask { id: string; query: string; expectedRelevantIds: string[]; conflictingIds: string[]; } export interface MemoryValueFixture { facts: MemoryValueFact[]; tasks: MemoryValueTask[]; } export interface MemoryValueTaskResult { taskId: string; baselineLoadedIds: string[]; governedLoadedIds: string[]; baselineContextBytes: number; governedContextBytes: number; baselinePrecision: number; governedPrecision: number; baselineRecall: number; governedRecall: number; baselineDecisionConsistent: boolean; governedDecisionConsistent: boolean; } export interface MemoryValueReport { taskResults: MemoryValueTaskResult[]; baselineContextBytes: number; governedContextBytes: number; baselineEstimatedTokens: number; governedEstimatedTokens: number; contextReductionPercent: number; baselinePrecision: number; governedPrecision: number; baselineRecall: number; governedRecall: number; baselineDecisionConsistency: number; governedDecisionConsistency: number; staleOrUnsafeFactsAvoided: number; staleOrUnsafeFactsLoadedByBaseline: number; verdict: 'pass' | 'fail'; } export declare function createDefaultMemoryValueFixture(): MemoryValueFixture; export declare function runMemoryValueBenchmark(fixture?: MemoryValueFixture): MemoryValueReport; export declare function formatMemoryValueReport(report: MemoryValueReport): string; //# sourceMappingURL=memory-value-benchmark.d.ts.map