bridgebench 3.1.0-alpha.0

package/dist/logger-QJU7SBDz.d.ts

@@ -0,0 +1,319 @@
+import { z } from 'zod';
+import { f as ArenaRunConfig, t as ModelRegistryEntry, M as MatchResult, E as EloState, h as ArenaSnapshot } from './reports-s2CTnGN8.js';
+import { d as CompleteArenaTask, B as BenchmarkCategory } from './tasks-CpaCJ6JE.js';
+
+declare const RUN_MANIFEST_VERSION = "1.0.0";
+declare const RunManifestSchema: z.ZodObject<{
+    version: z.ZodLiteral<"1.0.0">;
+    methodologyVersion: z.ZodString;
+    engineVersion: z.ZodString;
+    category: z.ZodEnum<["reasoning", "hallucination"]>;
+    seed: z.ZodString;
+    matches: z.ZodNumber;
+    competitors: z.ZodEffects<z.ZodArray<z.ZodObject<{
+        id: z.ZodString;
+        canonicalSlug: z.ZodString;
+        request: z.ZodObject<{
+            maxTokens: z.ZodNumber;
+            temperature: z.ZodNumber;
+            reasoningEffort: z.ZodEnum<["high", "medium", "low"]>;
+            excludeReasoning: z.ZodBoolean;
+            timeoutMs: z.ZodNumber;
+        }, "strip", z.ZodTypeAny, {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        }, {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        }>;
+    } & {
+        role: z.ZodLiteral<"competitor">;
+    }, "strip", z.ZodTypeAny, {
+        id: string;
+        role: "competitor";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }, {
+        id: string;
+        role: "competitor";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }>, "many">, {
+        id: string;
+        role: "competitor";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[], {
+        id: string;
+        role: "competitor";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[]>;
+    judges: z.ZodEffects<z.ZodArray<z.ZodObject<{
+        id: z.ZodString;
+        canonicalSlug: z.ZodString;
+        request: z.ZodObject<{
+            maxTokens: z.ZodNumber;
+            temperature: z.ZodNumber;
+            reasoningEffort: z.ZodEnum<["high", "medium", "low"]>;
+            excludeReasoning: z.ZodBoolean;
+            timeoutMs: z.ZodNumber;
+        }, "strip", z.ZodTypeAny, {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        }, {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        }>;
+    } & {
+        role: z.ZodLiteral<"judge">;
+    }, "strip", z.ZodTypeAny, {
+        id: string;
+        role: "judge";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }, {
+        id: string;
+        role: "judge";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }>, "many">, {
+        id: string;
+        role: "judge";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[], {
+        id: string;
+        role: "judge";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[]>;
+    tasks: z.ZodArray<z.ZodObject<{
+        id: z.ZodString;
+        version: z.ZodString;
+        publicHash: z.ZodString;
+        privateHash: z.ZodString;
+    }, "strip", z.ZodTypeAny, {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string;
+    }, {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string;
+    }>, "many">;
+    promptPolicyHashes: z.ZodObject<{
+        competitor: z.ZodString;
+        judge: z.ZodString;
+    }, "strip", z.ZodTypeAny, {
+        competitor: string;
+        judge: string;
+    }, {
+        competitor: string;
+        judge: string;
+    }>;
+}, "strip", z.ZodTypeAny, {
+    methodologyVersion: string;
+    seed: string;
+    version: "1.0.0";
+    category: "reasoning" | "hallucination";
+    competitors: {
+        id: string;
+        role: "competitor";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[];
+    tasks: {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string;
+    }[];
+    engineVersion: string;
+    matches: number;
+    judges: {
+        id: string;
+        role: "judge";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[];
+    promptPolicyHashes: {
+        competitor: string;
+        judge: string;
+    };
+}, {
+    methodologyVersion: string;
+    seed: string;
+    version: "1.0.0";
+    category: "reasoning" | "hallucination";
+    competitors: {
+        id: string;
+        role: "competitor";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[];
+    tasks: {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string;
+    }[];
+    engineVersion: string;
+    matches: number;
+    judges: {
+        id: string;
+        role: "judge";
+        canonicalSlug: string;
+        request: {
+            temperature: number;
+            timeoutMs: number;
+            reasoningEffort: "high" | "medium" | "low";
+            maxTokens: number;
+            excludeReasoning: boolean;
+        };
+    }[];
+    promptPolicyHashes: {
+        competitor: string;
+        judge: string;
+    };
+}>;
+type RunManifest = z.infer<typeof RunManifestSchema>;
+declare function canonicalJson(value: unknown): string;
+declare function runManifestHash(manifest: RunManifest): string;
+declare function runIdFromManifest(manifest: RunManifest): string;
+declare function createRunManifest(config: Pick<ArenaRunConfig, 'category' | 'seed' | 'matches' | 'competitorIds'>, tasks: CompleteArenaTask[], resolvedCompetitors?: readonly ModelRegistryEntry[]): RunManifest;
+
+interface ArenaStoreConfig {
+    category: BenchmarkCategory;
+    journalPath: string;
+    snapshotPath: string;
+    markdownPath: string;
+    runsDir?: string;
+    /** Skip directory creation for read-only verification of external journals. */
+    readOnly?: boolean;
+}
+/** Each category gets its own results directory — its own journal, Elo ladder, and reports. */
+declare function categoryStoreConfig(category: BenchmarkCategory, resultsRoot?: string | undefined): ArenaStoreConfig;
+declare class ArenaStore {
+    readonly config: ArenaStoreConfig;
+    constructor(config: ArenaStoreConfig);
+    get category(): BenchmarkCategory;
+    get runsDir(): string;
+    append(result: MatchResult): void;
+    readAll(): MatchResult[];
+    writeRunManifest(runId: string, manifest: RunManifest): void;
+    readRunManifest(runId: string): RunManifest | null;
+    completedMatchIds(): Set<string>;
+    rebuildEloState(competitorIds?: readonly string[]): EloState;
+    writeSnapshot(snapshot: ArenaSnapshot): void;
+    writeMarkdown(markdown: string): void;
+    private atomicWrite;
+}
+
+type LogLevel = 'debug' | 'info' | 'warn' | 'error';
+declare function redactSecrets(value: string): string;
+/** Structured logger every arena boundary reports into; one JSONL file per process. */
+interface ArenaLogger {
+    readonly filePath: string | null;
+    debug(event: string, data?: Record<string, unknown>): void;
+    info(event: string, data?: Record<string, unknown>): void;
+    warn(event: string, data?: Record<string, unknown>): void;
+    error(event: string, data?: Record<string, unknown>): void;
+}
+declare const noopLogger: ArenaLogger;
+declare class FileArenaLogger implements ArenaLogger {
+    readonly filePath: string;
+    private readonly verbose;
+    constructor(options: {
+        dir: string;
+        verbose?: boolean;
+        name?: string;
+    });
+    debug(event: string, data?: Record<string, unknown>): void;
+    info(event: string, data?: Record<string, unknown>): void;
+    warn(event: string, data?: Record<string, unknown>): void;
+    error(event: string, data?: Record<string, unknown>): void;
+    private write;
+}
+
+export { ArenaStore as A, FileArenaLogger as F, type LogLevel as L, type RunManifest as R, type ArenaLogger as a, type ArenaStoreConfig as b, RUN_MANIFEST_VERSION as c, RunManifestSchema as d, canonicalJson as e, categoryStoreConfig as f, createRunManifest as g, runIdFromManifest as h, runManifestHash as i, noopLogger as n, redactSecrets as r };