bridgebench 3.1.0-alpha.0 → 3.1.0-alpha.2

package/dist/contracts/index.d.cts

@@ -1,5 +1,5 @@
-export { A as ArenaTask, B as BenchmarkCategory, a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, d as CompleteArenaTask, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema } from '../tasks-CpaCJ6JE.cjs';
-export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, c as ArenaEventSink, d as ArenaEventType, e as ArenaExecutionOptions, f as ArenaRunConfig, g as ArenaRunResult, h as ArenaSnapshot, C as ChatRequest, i as CompetitorFailure, j as CompetitorResponse, k as CompetitorResponseSchema, l as CompetitorSuccess, E as EloState, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, M as MatchResult, q as MatchResultSchema, r as ModelCompletion, s as ModelCompletionSchema, t as ModelRegistryEntry, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, O as OpenRouterGateway, P as PanelDecision, x as PanelDecisionSchema, S as ScheduledMatch, y as competitorContent, z as competitorCost, B as competitorOutputTokens, D as competitorReasoningTokens } from '../reports-4CejmOHf.cjs';
+export { A as ArenaTask, B as BenchmarkCategory, a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, d as CompleteArenaTask, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema, j as isCompleteArenaTask } from '../tasks-BmhWuMBD.cjs';
+export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, c as ArenaEventSink, d as ArenaEventType, e as ArenaExecutionOptions, f as ArenaRunConfig, g as ArenaRunResult, h as ArenaSnapshot, C as ChatRequest, i as CompetitorFailure, j as CompetitorResponse, k as CompetitorResponseSchema, l as CompetitorSuccess, E as EloState, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, M as MatchResult, q as MatchResultSchema, r as ModelCompletion, s as ModelCompletionSchema, t as ModelRegistryEntry, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, O as OpenRouterGateway, P as PanelDecision, x as PanelDecisionSchema, S as ScheduledMatch, y as SpeedMetric, z as SpeedMetrics, B as SpeedMetricsSchema, D as SpeedStats, F as competitorContent, G as competitorCost, H as competitorOutputTokens, I as competitorReasoningTokens } from '../reports-DPpOoOux.cjs';
 import 'zod';
 
 /**

package/dist/contracts/index.d.ts

@@ -1,5 +1,5 @@
-export { A as ArenaTask, B as BenchmarkCategory, a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, d as CompleteArenaTask, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema } from '../tasks-CpaCJ6JE.js';
-export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, c as ArenaEventSink, d as ArenaEventType, e as ArenaExecutionOptions, f as ArenaRunConfig, g as ArenaRunResult, h as ArenaSnapshot, C as ChatRequest, i as CompetitorFailure, j as CompetitorResponse, k as CompetitorResponseSchema, l as CompetitorSuccess, E as EloState, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, M as MatchResult, q as MatchResultSchema, r as ModelCompletion, s as ModelCompletionSchema, t as ModelRegistryEntry, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, O as OpenRouterGateway, P as PanelDecision, x as PanelDecisionSchema, S as ScheduledMatch, y as competitorContent, z as competitorCost, B as competitorOutputTokens, D as competitorReasoningTokens } from '../reports-s2CTnGN8.js';
+export { A as ArenaTask, B as BenchmarkCategory, a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, d as CompleteArenaTask, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema, j as isCompleteArenaTask } from '../tasks-BmhWuMBD.js';
+export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, c as ArenaEventSink, d as ArenaEventType, e as ArenaExecutionOptions, f as ArenaRunConfig, g as ArenaRunResult, h as ArenaSnapshot, C as ChatRequest, i as CompetitorFailure, j as CompetitorResponse, k as CompetitorResponseSchema, l as CompetitorSuccess, E as EloState, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, M as MatchResult, q as MatchResultSchema, r as ModelCompletion, s as ModelCompletionSchema, t as ModelRegistryEntry, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, O as OpenRouterGateway, P as PanelDecision, x as PanelDecisionSchema, S as ScheduledMatch, y as SpeedMetric, z as SpeedMetrics, B as SpeedMetricsSchema, D as SpeedStats, F as competitorContent, G as competitorCost, H as competitorOutputTokens, I as competitorReasoningTokens } from '../reports-B8TCJtPr.js';
 import 'zod';
 
 /**

package/dist/contracts/index.js

@@ -13,14 +13,16 @@ import {
   ModelCompletionSchema,
   ModelRoleSchema,
   PanelDecisionSchema,
+  SpeedMetricsSchema,
   TaskArtifactSchema,
   TaskPrivateSchema,
   TaskPublicSchema,
   competitorContent,
   competitorCost,
   competitorOutputTokens,
-  competitorReasoningTokens
-} from "../chunk-UECBSKTD.js";
+  competitorReasoningTokens,
+  isCompleteArenaTask
+} from "../chunk-AY45YLYL.js";
 export {
   BenchmarkCategorySchema,
   CATEGORIES,
@@ -36,12 +38,13 @@ export {
   ModelCompletionSchema,
   ModelRoleSchema,
   PanelDecisionSchema,
+  SpeedMetricsSchema,
   TaskArtifactSchema,
   TaskPrivateSchema,
   TaskPublicSchema,
   competitorContent,
   competitorCost,
   competitorOutputTokens,
-  competitorReasoningTokens
+  competitorReasoningTokens,
+  isCompleteArenaTask
 };
-//# sourceMappingURL=index.js.map

package/dist/index.cjs

@@ -27,7 +27,7 @@
 
 
 
-var _chunkNJTYVNP4cjs = require('./chunk-NJTYVNP4.cjs');
+var _chunkVAS6KNJAcjs = require('./chunk-VAS6KNJA.cjs');
 
 
 
@@ -61,7 +61,7 @@ var _chunkNJTYVNP4cjs = require('./chunk-NJTYVNP4.cjs');
 
 
 
-var _chunk4TWPCPRPcjs = require('./chunk-4TWPCPRP.cjs');
+var _chunkQMOPRKWDcjs = require('./chunk-QMOPRKWD.cjs');
 
 
 
@@ -84,9 +84,9 @@ var _chunk4TWPCPRPcjs = require('./chunk-4TWPCPRP.cjs');
 
 
 
-var _chunkCIXITJW6cjs = require('./chunk-CIXITJW6.cjs');
 
 
+var _chunkX3LPZGHScjs = require('./chunk-X3LPZGHS.cjs');
 
 
 
@@ -167,5 +167,8 @@ var _chunkCIXITJW6cjs = require('./chunk-CIXITJW6.cjs');
 
 
 
-exports.ArenaCancellationError = _chunk4TWPCPRPcjs.ArenaCancellationError; exports.ArenaRunner = _chunkNJTYVNP4cjs.ArenaRunner; exports.ArenaStore = _chunkNJTYVNP4cjs.ArenaStore; exports.BenchmarkCategorySchema = _chunkCIXITJW6cjs.BenchmarkCategorySchema; exports.CATEGORIES = _chunkCIXITJW6cjs.CATEGORIES; exports.CATEGORY_CLUSTERS = _chunkCIXITJW6cjs.CATEGORY_CLUSTERS; exports.CATEGORY_META = _chunkCIXITJW6cjs.CATEGORY_META; exports.CONTRACTS_VERSION = _chunkCIXITJW6cjs.CONTRACTS_VERSION; exports.CompetitorResponseSchema = _chunkCIXITJW6cjs.CompetitorResponseSchema; exports.ELO_INITIAL = _chunkNJTYVNP4cjs.ELO_INITIAL; exports.ELO_K = _chunkNJTYVNP4cjs.ELO_K; exports.ENGINE_VERSION = _chunkNJTYVNP4cjs.ENGINE_VERSION; exports.FAST_RESPONSE_MS = _chunkNJTYVNP4cjs.FAST_RESPONSE_MS; exports.FileArenaLogger = _chunk4TWPCPRPcjs.FileArenaLogger; exports.JUDGE_VERDICT_TRANSPORT_SCHEMA = _chunkCIXITJW6cjs.JUDGE_VERDICT_TRANSPORT_SCHEMA; exports.JudgePanel = _chunk4TWPCPRPcjs.JudgePanel; exports.JudgeVerdictSchema = _chunkCIXITJW6cjs.JudgeVerdictSchema; exports.JudgeVoteSchema = _chunkCIXITJW6cjs.JudgeVoteSchema; exports.LOW_OUTPUT_TOKENS = _chunkNJTYVNP4cjs.LOW_OUTPUT_TOKENS; exports.MAX_PROMPT_CHARS = _chunk4TWPCPRPcjs.MAX_PROMPT_CHARS; exports.METHODOLOGY_VERSION = _chunkCIXITJW6cjs.METHODOLOGY_VERSION; exports.MODEL_REGISTRY = _chunk4TWPCPRPcjs.MODEL_REGISTRY; exports.MatchResultSchema = _chunkCIXITJW6cjs.MatchResultSchema; exports.MockOpenRouterGateway = _chunkNJTYVNP4cjs.MockOpenRouterGateway; exports.ModelCompletionSchema = _chunkCIXITJW6cjs.ModelCompletionSchema; exports.ModelRoleSchema = _chunkCIXITJW6cjs.ModelRoleSchema; exports.OpenRouterClient = _chunk4TWPCPRPcjs.OpenRouterClient; exports.PACKAGE_NAME = _chunkNJTYVNP4cjs.PACKAGE_NAME; exports.PanelDecisionSchema = _chunkCIXITJW6cjs.PanelDecisionSchema; exports.RUN_MANIFEST_VERSION = _chunkNJTYVNP4cjs.RUN_MANIFEST_VERSION; exports.RunManifestSchema = _chunkNJTYVNP4cjs.RunManifestSchema; exports.SOL_FABLE_PILOT_COMPETITOR_IDS = _chunk4TWPCPRPcjs.SOL_FABLE_PILOT_COMPETITOR_IDS; exports.TASKS_PER_CATEGORY = _chunk4TWPCPRPcjs.TASKS_PER_CATEGORY; exports.TASKS_PER_CLUSTER = _chunk4TWPCPRPcjs.TASKS_PER_CLUSTER; exports.TaskArtifactSchema = _chunkCIXITJW6cjs.TaskArtifactSchema; exports.TaskLoader = _chunk4TWPCPRPcjs.TaskLoader; exports.TaskPrivateSchema = _chunkCIXITJW6cjs.TaskPrivateSchema; exports.TaskPublicSchema = _chunkCIXITJW6cjs.TaskPublicSchema; exports.anonymizeCompetitorOutput = _chunk4TWPCPRPcjs.anonymizeCompetitorOutput; exports.applyEloWin = _chunkNJTYVNP4cjs.applyEloWin; exports.assertPromptSize = _chunk4TWPCPRPcjs.assertPromptSize; exports.buildCompetitorPrompt = _chunk4TWPCPRPcjs.buildCompetitorPrompt; exports.buildJudgePayload = _chunk4TWPCPRPcjs.buildJudgePayload; exports.buildSnapshot = _chunkNJTYVNP4cjs.buildSnapshot; exports.canonicalJson = _chunkNJTYVNP4cjs.canonicalJson; exports.categoryStoreConfig = _chunkNJTYVNP4cjs.categoryStoreConfig; exports.classifyError = _chunkNJTYVNP4cjs.classifyError; exports.competitorContent = _chunkCIXITJW6cjs.competitorContent; exports.competitorCost = _chunkCIXITJW6cjs.competitorCost; exports.competitorOutputTokens = _chunkCIXITJW6cjs.competitorOutputTokens; exports.competitorPromptPolicyHash = _chunk4TWPCPRPcjs.competitorPromptPolicyHash; exports.competitorReasoningTokens = _chunkCIXITJW6cjs.competitorReasoningTokens; exports.createRunManifest = _chunkNJTYVNP4cjs.createRunManifest; exports.defaultTaskRoot = _chunk4TWPCPRPcjs.defaultTaskRoot; exports.detectResponseAnomalies = _chunkNJTYVNP4cjs.detectResponseAnomalies; exports.expectedScore = _chunkNJTYVNP4cjs.expectedScore; exports.formatTriage = _chunkNJTYVNP4cjs.formatTriage; exports.getModel = _chunk4TWPCPRPcjs.getModel; exports.isArenaCancellationError = _chunk4TWPCPRPcjs.isArenaCancellationError; exports.isRetryableError = _chunk4TWPCPRPcjs.isRetryableError; exports.judgePromptPolicyHash = _chunk4TWPCPRPcjs.judgePromptPolicyHash; exports.judgeSystemPrompt = _chunk4TWPCPRPcjs.judgeSystemPrompt; exports.judgeVerdictJsonSchema = _chunk4TWPCPRPcjs.judgeVerdictJsonSchema; exports.listModels = _chunk4TWPCPRPcjs.listModels; exports.mergePrivateHalves = _chunk4TWPCPRPcjs.mergePrivateHalves; exports.noopLogger = _chunk4TWPCPRPcjs.noopLogger; exports.parseJudgeVerdict = _chunk4TWPCPRPcjs.parseJudgeVerdict; exports.redactSecrets = _chunk4TWPCPRPcjs.redactSecrets; exports.renderMarkdown = _chunkNJTYVNP4cjs.renderMarkdown; exports.resolveCompetitorRoster = _chunk4TWPCPRPcjs.resolveCompetitorRoster; exports.runIdFromManifest = _chunkNJTYVNP4cjs.runIdFromManifest; exports.runManifestHash = _chunkNJTYVNP4cjs.runManifestHash; exports.runOpenRouterAttempt = _chunk4TWPCPRPcjs.runOpenRouterAttempt; exports.sanitizeError = _chunk4TWPCPRPcjs.sanitizeError; exports.scheduleMatches = _chunkNJTYVNP4cjs.scheduleMatches; exports.throwIfCancelled = _chunk4TWPCPRPcjs.throwIfCancelled; exports.triageJournal = _chunkNJTYVNP4cjs.triageJournal; exports.validatePublicTaskFile = _chunk4TWPCPRPcjs.validatePublicTaskFile; exports.verifiedEloState = _chunkNJTYVNP4cjs.verifiedEloState; exports.verifyJournal = _chunkNJTYVNP4cjs.verifyJournal; exports.writeReports = _chunkNJTYVNP4cjs.writeReports;
-//# sourceMappingURL=index.cjs.map
+
+
+
+
+exports.ArenaCancellationError = _chunkQMOPRKWDcjs.ArenaCancellationError; exports.ArenaRunner = _chunkVAS6KNJAcjs.ArenaRunner; exports.ArenaStore = _chunkVAS6KNJAcjs.ArenaStore; exports.BenchmarkCategorySchema = _chunkX3LPZGHScjs.BenchmarkCategorySchema; exports.CATEGORIES = _chunkX3LPZGHScjs.CATEGORIES; exports.CATEGORY_CLUSTERS = _chunkX3LPZGHScjs.CATEGORY_CLUSTERS; exports.CATEGORY_META = _chunkX3LPZGHScjs.CATEGORY_META; exports.CONTRACTS_VERSION = _chunkX3LPZGHScjs.CONTRACTS_VERSION; exports.CompetitorResponseSchema = _chunkX3LPZGHScjs.CompetitorResponseSchema; exports.ELO_INITIAL = _chunkVAS6KNJAcjs.ELO_INITIAL; exports.ELO_K = _chunkVAS6KNJAcjs.ELO_K; exports.ENGINE_VERSION = _chunkVAS6KNJAcjs.ENGINE_VERSION; exports.FAST_RESPONSE_MS = _chunkVAS6KNJAcjs.FAST_RESPONSE_MS; exports.FileArenaLogger = _chunkQMOPRKWDcjs.FileArenaLogger; exports.JUDGE_VERDICT_TRANSPORT_SCHEMA = _chunkX3LPZGHScjs.JUDGE_VERDICT_TRANSPORT_SCHEMA; exports.JudgePanel = _chunkQMOPRKWDcjs.JudgePanel; exports.JudgeVerdictSchema = _chunkX3LPZGHScjs.JudgeVerdictSchema; exports.JudgeVoteSchema = _chunkX3LPZGHScjs.JudgeVoteSchema; exports.LOW_OUTPUT_TOKENS = _chunkVAS6KNJAcjs.LOW_OUTPUT_TOKENS; exports.MAX_PROMPT_CHARS = _chunkQMOPRKWDcjs.MAX_PROMPT_CHARS; exports.METHODOLOGY_VERSION = _chunkX3LPZGHScjs.METHODOLOGY_VERSION; exports.MODEL_REGISTRY = _chunkQMOPRKWDcjs.MODEL_REGISTRY; exports.MatchResultSchema = _chunkX3LPZGHScjs.MatchResultSchema; exports.MockOpenRouterGateway = _chunkVAS6KNJAcjs.MockOpenRouterGateway; exports.ModelCompletionSchema = _chunkX3LPZGHScjs.ModelCompletionSchema; exports.ModelRoleSchema = _chunkX3LPZGHScjs.ModelRoleSchema; exports.OpenRouterClient = _chunkQMOPRKWDcjs.OpenRouterClient; exports.PACKAGE_NAME = _chunkVAS6KNJAcjs.PACKAGE_NAME; exports.PanelDecisionSchema = _chunkX3LPZGHScjs.PanelDecisionSchema; exports.RUN_MANIFEST_VERSION = _chunkVAS6KNJAcjs.RUN_MANIFEST_VERSION; exports.RunManifestSchema = _chunkVAS6KNJAcjs.RunManifestSchema; exports.SOL_FABLE_PILOT_COMPETITOR_IDS = _chunkQMOPRKWDcjs.SOL_FABLE_PILOT_COMPETITOR_IDS; exports.SpeedMetricsSchema = _chunkX3LPZGHScjs.SpeedMetricsSchema; exports.TASKS_PER_CATEGORY = _chunkQMOPRKWDcjs.TASKS_PER_CATEGORY; exports.TASKS_PER_CLUSTER = _chunkQMOPRKWDcjs.TASKS_PER_CLUSTER; exports.TaskArtifactSchema = _chunkX3LPZGHScjs.TaskArtifactSchema; exports.TaskLoader = _chunkQMOPRKWDcjs.TaskLoader; exports.TaskPrivateSchema = _chunkX3LPZGHScjs.TaskPrivateSchema; exports.TaskPublicSchema = _chunkX3LPZGHScjs.TaskPublicSchema; exports.anonymizeCompetitorOutput = _chunkQMOPRKWDcjs.anonymizeCompetitorOutput; exports.applyEloWin = _chunkVAS6KNJAcjs.applyEloWin; exports.assertPromptSize = _chunkQMOPRKWDcjs.assertPromptSize; exports.buildCompetitorPrompt = _chunkQMOPRKWDcjs.buildCompetitorPrompt; exports.buildJudgePayload = _chunkQMOPRKWDcjs.buildJudgePayload; exports.buildSnapshot = _chunkVAS6KNJAcjs.buildSnapshot; exports.canonicalJson = _chunkVAS6KNJAcjs.canonicalJson; exports.categoryStoreConfig = _chunkVAS6KNJAcjs.categoryStoreConfig; exports.classifyError = _chunkVAS6KNJAcjs.classifyError; exports.competitorContent = _chunkX3LPZGHScjs.competitorContent; exports.competitorCost = _chunkX3LPZGHScjs.competitorCost; exports.competitorOutputTokens = _chunkX3LPZGHScjs.competitorOutputTokens; exports.competitorPromptPolicyHash = _chunkQMOPRKWDcjs.competitorPromptPolicyHash; exports.competitorReasoningTokens = _chunkX3LPZGHScjs.competitorReasoningTokens; exports.createRunManifest = _chunkVAS6KNJAcjs.createRunManifest; exports.defaultTaskRoot = _chunkQMOPRKWDcjs.defaultTaskRoot; exports.detectResponseAnomalies = _chunkVAS6KNJAcjs.detectResponseAnomalies; exports.expectedScore = _chunkVAS6KNJAcjs.expectedScore; exports.formatTriage = _chunkVAS6KNJAcjs.formatTriage; exports.getModel = _chunkQMOPRKWDcjs.getModel; exports.isArenaCancellationError = _chunkQMOPRKWDcjs.isArenaCancellationError; exports.isCompleteArenaTask = _chunkX3LPZGHScjs.isCompleteArenaTask; exports.isRetryableError = _chunkQMOPRKWDcjs.isRetryableError; exports.judgePromptPolicyHash = _chunkQMOPRKWDcjs.judgePromptPolicyHash; exports.judgeSystemPrompt = _chunkQMOPRKWDcjs.judgeSystemPrompt; exports.judgeVerdictJsonSchema = _chunkQMOPRKWDcjs.judgeVerdictJsonSchema; exports.listModels = _chunkQMOPRKWDcjs.listModels; exports.mergePrivateHalves = _chunkQMOPRKWDcjs.mergePrivateHalves; exports.noopLogger = _chunkQMOPRKWDcjs.noopLogger; exports.parseJudgeVerdict = _chunkQMOPRKWDcjs.parseJudgeVerdict; exports.redactSecrets = _chunkQMOPRKWDcjs.redactSecrets; exports.renderMarkdown = _chunkVAS6KNJAcjs.renderMarkdown; exports.resolveCompetitorRoster = _chunkQMOPRKWDcjs.resolveCompetitorRoster; exports.runIdFromManifest = _chunkVAS6KNJAcjs.runIdFromManifest; exports.runManifestHash = _chunkVAS6KNJAcjs.runManifestHash; exports.runOpenRouterAttempt = _chunkQMOPRKWDcjs.runOpenRouterAttempt; exports.sanitizeError = _chunkQMOPRKWDcjs.sanitizeError; exports.scheduleMatches = _chunkVAS6KNJAcjs.scheduleMatches; exports.throwIfCancelled = _chunkQMOPRKWDcjs.throwIfCancelled; exports.triageJournal = _chunkVAS6KNJAcjs.triageJournal; exports.validatePublicTaskFile = _chunkQMOPRKWDcjs.validatePublicTaskFile; exports.verifiedEloState = _chunkVAS6KNJAcjs.verifiedEloState; exports.verifyJournal = _chunkVAS6KNJAcjs.verifyJournal; exports.writeReports = _chunkVAS6KNJAcjs.writeReports;

package/dist/index.d.cts

@@ -1,10 +1,10 @@
 export { CONTRACTS_VERSION } from './contracts/index.cjs';
-import { A as ArenaStore, a as ArenaLogger, R as RunManifest } from './logger-CCR9Mg1c.cjs';
-export { b as ArenaStoreConfig, F as FileArenaLogger, L as LogLevel, c as RUN_MANIFEST_VERSION, d as RunManifestSchema, e as canonicalJson, f as categoryStoreConfig, g as createRunManifest, n as noopLogger, r as redactSecrets, h as runIdFromManifest, i as runManifestHash } from './logger-CCR9Mg1c.cjs';
-import { O as OpenRouterGateway, c as ArenaEventSink, f as ArenaRunConfig, e as ArenaExecutionOptions, g as ArenaRunResult, S as ScheduledMatch, l as CompetitorSuccess, P as PanelDecision, t as ModelRegistryEntry, C as ChatRequest, r as ModelCompletion, M as MatchResult, E as EloState, h as ArenaSnapshot, j as CompetitorResponse } from './reports-4CejmOHf.cjs';
-export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, d as ArenaEventType, i as CompetitorFailure, k as CompetitorResponseSchema, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, q as MatchResultSchema, s as ModelCompletionSchema, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, x as PanelDecisionSchema, y as competitorContent, z as competitorCost, B as competitorOutputTokens, D as competitorReasoningTokens } from './reports-4CejmOHf.cjs';
-import { d as CompleteArenaTask, B as BenchmarkCategory, A as ArenaTask } from './tasks-CpaCJ6JE.cjs';
-export { a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema } from './tasks-CpaCJ6JE.cjs';
+import { A as ArenaStore, a as ArenaLogger, R as RunManifest } from './logger-BByta-7V.cjs';
+export { b as ArenaStoreConfig, F as FileArenaLogger, L as LogLevel, c as RUN_MANIFEST_VERSION, d as RunManifestSchema, e as canonicalJson, f as categoryStoreConfig, g as createRunManifest, n as noopLogger, r as redactSecrets, h as runIdFromManifest, i as runManifestHash } from './logger-BByta-7V.cjs';
+import { O as OpenRouterGateway, c as ArenaEventSink, f as ArenaRunConfig, e as ArenaExecutionOptions, g as ArenaRunResult, S as ScheduledMatch, l as CompetitorSuccess, P as PanelDecision, t as ModelRegistryEntry, C as ChatRequest, r as ModelCompletion, M as MatchResult, E as EloState, h as ArenaSnapshot, j as CompetitorResponse } from './reports-DPpOoOux.cjs';
+export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, d as ArenaEventType, i as CompetitorFailure, k as CompetitorResponseSchema, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, q as MatchResultSchema, s as ModelCompletionSchema, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, x as PanelDecisionSchema, y as SpeedMetric, z as SpeedMetrics, B as SpeedMetricsSchema, D as SpeedStats, F as competitorContent, G as competitorCost, H as competitorOutputTokens, I as competitorReasoningTokens } from './reports-DPpOoOux.cjs';
+import { A as ArenaTask, d as CompleteArenaTask, B as BenchmarkCategory } from './tasks-BmhWuMBD.cjs';
+export { a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema, j as isCompleteArenaTask } from './tasks-BmhWuMBD.cjs';
 import OpenAI from 'openai';
 export { PrivateHalf, TASKS_PER_CATEGORY, TASKS_PER_CLUSTER, TaskLoader, buildCompetitorPrompt, competitorPromptPolicyHash, defaultTaskRoot, mergePrivateHalves, validatePublicTaskFile } from './tasks.cjs';
 import 'zod';
@@ -20,7 +20,7 @@ declare class ArenaRunner {
     private observeCancellation;
     private cancelRun;
     private prepareRun;
-    run(config: ArenaRunConfig, tasks: CompleteArenaTask[], execution?: ArenaExecutionOptions): Promise<ArenaRunResult>;
+    run(config: ArenaRunConfig, tasks: ArenaTask[], execution?: ArenaExecutionOptions): Promise<ArenaRunResult>;
     private runMatch;
     /** Surface signals that make a journaled response untrustworthy or a task too easy. */
     private inspectResponse;
@@ -105,9 +105,9 @@ declare function parseJudgeVerdict(content: string): {
     rationale: string;
     criteria: {
         correctness: string;
-        grounding: string;
         constraintHandling: string;
         completeness: string;
+        grounding: string;
     };
     violations: string[];
 };
@@ -189,6 +189,7 @@ interface RunTriage {
         judged: number;
         forfeit: number;
         'no-contest': number;
+        'speed-decided': number;
     };
     totalCostUsd: number;
     errorClasses: Record<string, number>;

package/dist/index.d.ts

@@ -1,10 +1,10 @@
 export { CONTRACTS_VERSION } from './contracts/index.js';
-import { A as ArenaStore, a as ArenaLogger, R as RunManifest } from './logger-QJU7SBDz.js';
-export { b as ArenaStoreConfig, F as FileArenaLogger, L as LogLevel, c as RUN_MANIFEST_VERSION, d as RunManifestSchema, e as canonicalJson, f as categoryStoreConfig, g as createRunManifest, n as noopLogger, r as redactSecrets, h as runIdFromManifest, i as runManifestHash } from './logger-QJU7SBDz.js';
-import { O as OpenRouterGateway, c as ArenaEventSink, f as ArenaRunConfig, e as ArenaExecutionOptions, g as ArenaRunResult, S as ScheduledMatch, l as CompetitorSuccess, P as PanelDecision, t as ModelRegistryEntry, C as ChatRequest, r as ModelCompletion, M as MatchResult, E as EloState, h as ArenaSnapshot, j as CompetitorResponse } from './reports-s2CTnGN8.js';
-export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, d as ArenaEventType, i as CompetitorFailure, k as CompetitorResponseSchema, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, q as MatchResultSchema, s as ModelCompletionSchema, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, x as PanelDecisionSchema, y as competitorContent, z as competitorCost, B as competitorOutputTokens, D as competitorReasoningTokens } from './reports-s2CTnGN8.js';
-import { d as CompleteArenaTask, B as BenchmarkCategory, A as ArenaTask } from './tasks-CpaCJ6JE.js';
-export { a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema } from './tasks-CpaCJ6JE.js';
+import { A as ArenaStore, a as ArenaLogger, R as RunManifest } from './logger-BQf29BLe.js';
+export { b as ArenaStoreConfig, F as FileArenaLogger, L as LogLevel, c as RUN_MANIFEST_VERSION, d as RunManifestSchema, e as canonicalJson, f as categoryStoreConfig, g as createRunManifest, n as noopLogger, r as redactSecrets, h as runIdFromManifest, i as runManifestHash } from './logger-BQf29BLe.js';
+import { O as OpenRouterGateway, c as ArenaEventSink, f as ArenaRunConfig, e as ArenaExecutionOptions, g as ArenaRunResult, S as ScheduledMatch, l as CompetitorSuccess, P as PanelDecision, t as ModelRegistryEntry, C as ChatRequest, r as ModelCompletion, M as MatchResult, E as EloState, h as ArenaSnapshot, j as CompetitorResponse } from './reports-B8TCJtPr.js';
+export { A as ArenaEvent, a as ArenaEventDataMap, b as ArenaEventInput, d as ArenaEventType, i as CompetitorFailure, k as CompetitorResponseSchema, J as JUDGE_VERDICT_TRANSPORT_SCHEMA, m as JudgeVerdict, n as JudgeVerdictSchema, o as JudgeVote, p as JudgeVoteSchema, L as LeaderboardEntry, q as MatchResultSchema, s as ModelCompletionSchema, u as ModelRequestPolicy, v as ModelRole, w as ModelRoleSchema, x as PanelDecisionSchema, y as SpeedMetric, z as SpeedMetrics, B as SpeedMetricsSchema, D as SpeedStats, F as competitorContent, G as competitorCost, H as competitorOutputTokens, I as competitorReasoningTokens } from './reports-B8TCJtPr.js';
+import { A as ArenaTask, d as CompleteArenaTask, B as BenchmarkCategory } from './tasks-BmhWuMBD.js';
+export { a as BenchmarkCategorySchema, C as CATEGORIES, b as CATEGORY_CLUSTERS, c as CATEGORY_META, M as METHODOLOGY_VERSION, T as TaskArtifact, e as TaskArtifactSchema, f as TaskPrivate, g as TaskPrivateSchema, h as TaskPublic, i as TaskPublicSchema, j as isCompleteArenaTask } from './tasks-BmhWuMBD.js';
 import OpenAI from 'openai';
 export { PrivateHalf, TASKS_PER_CATEGORY, TASKS_PER_CLUSTER, TaskLoader, buildCompetitorPrompt, competitorPromptPolicyHash, defaultTaskRoot, mergePrivateHalves, validatePublicTaskFile } from './tasks.js';
 import 'zod';
@@ -20,7 +20,7 @@ declare class ArenaRunner {
     private observeCancellation;
     private cancelRun;
     private prepareRun;
-    run(config: ArenaRunConfig, tasks: CompleteArenaTask[], execution?: ArenaExecutionOptions): Promise<ArenaRunResult>;
+    run(config: ArenaRunConfig, tasks: ArenaTask[], execution?: ArenaExecutionOptions): Promise<ArenaRunResult>;
     private runMatch;
     /** Surface signals that make a journaled response untrustworthy or a task too easy. */
     private inspectResponse;
@@ -105,9 +105,9 @@ declare function parseJudgeVerdict(content: string): {
     rationale: string;
     criteria: {
         correctness: string;
-        grounding: string;
         constraintHandling: string;
         completeness: string;
+        grounding: string;
     };
     violations: string[];
 };
@@ -189,6 +189,7 @@ interface RunTriage {
         judged: number;
         forfeit: number;
         'no-contest': number;
+        'speed-decided': number;
     };
     totalCostUsd: number;
     errorClasses: Record<string, number>;

package/dist/index.js

@@ -27,7 +27,7 @@ import {
   verifiedEloState,
   verifyJournal,
   writeReports
-} from "./chunk-EQHRUV2I.js";
+} from "./chunk-IUPFMGUL.js";
 import {
   ArenaCancellationError,
   FileArenaLogger,
@@ -61,7 +61,7 @@ import {
   sanitizeError,
   throwIfCancelled,
   validatePublicTaskFile
-} from "./chunk-JTVNKSMO.js";
+} from "./chunk-DVMGL3L7.js";
 import {
   BenchmarkCategorySchema,
   CATEGORIES,
@@ -77,14 +77,16 @@ import {
   ModelCompletionSchema,
   ModelRoleSchema,
   PanelDecisionSchema,
+  SpeedMetricsSchema,
   TaskArtifactSchema,
   TaskPrivateSchema,
   TaskPublicSchema,
   competitorContent,
   competitorCost,
   competitorOutputTokens,
-  competitorReasoningTokens
-} from "./chunk-UECBSKTD.js";
+  competitorReasoningTokens,
+  isCompleteArenaTask
+} from "./chunk-AY45YLYL.js";
 export {
   ArenaCancellationError,
   ArenaRunner,
@@ -118,6 +120,7 @@ export {
   RUN_MANIFEST_VERSION,
   RunManifestSchema,
   SOL_FABLE_PILOT_COMPETITOR_IDS,
+  SpeedMetricsSchema,
   TASKS_PER_CATEGORY,
   TASKS_PER_CLUSTER,
   TaskArtifactSchema,
@@ -145,6 +148,7 @@ export {
   formatTriage,
   getModel,
   isArenaCancellationError,
+  isCompleteArenaTask,
   isRetryableError,
   judgePromptPolicyHash,
   judgeSystemPrompt,
@@ -168,4 +172,3 @@ export {
   verifyJournal,
   writeReports
 };
-//# sourceMappingURL=index.js.map

package/dist/{logger-CCR9Mg1c.d.cts → logger-BByta-7V.d.cts}

@@ -1,13 +1,13 @@
 import { z } from 'zod';
-import { f as ArenaRunConfig, t as ModelRegistryEntry, M as MatchResult, E as EloState, h as ArenaSnapshot } from './reports-4CejmOHf.cjs';
-import { d as CompleteArenaTask, B as BenchmarkCategory } from './tasks-CpaCJ6JE.cjs';
+import { f as ArenaRunConfig, t as ModelRegistryEntry, M as MatchResult, E as EloState, h as ArenaSnapshot } from './reports-DPpOoOux.cjs';
+import { A as ArenaTask, B as BenchmarkCategory } from './tasks-BmhWuMBD.cjs';
 
 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"]>;
+    category: z.ZodEnum<["reasoning", "hallucination", "security", "bullshit", "refactoring", "debugging", "generation", "speed"]>;
     seed: z.ZodString;
     matches: z.ZodNumber;
     competitors: z.ZodEffects<z.ZodArray<z.ZodObject<{
@@ -152,17 +152,17 @@ declare const RunManifestSchema: z.ZodObject<{
         id: z.ZodString;
         version: z.ZodString;
         publicHash: z.ZodString;
-        privateHash: z.ZodString;
+        privateHash: z.ZodNullable<z.ZodString>;
     }, "strip", z.ZodTypeAny, {
         id: string;
         version: string;
         publicHash: string;
-        privateHash: string;
+        privateHash: string | null;
     }, {
         id: string;
         version: string;
         publicHash: string;
-        privateHash: string;
+        privateHash: string | null;
     }>, "many">;
     promptPolicyHashes: z.ZodObject<{
         competitor: z.ZodString;
@@ -175,10 +175,16 @@ declare const RunManifestSchema: z.ZodObject<{
         judge: string;
     }>;
 }, "strip", z.ZodTypeAny, {
+    tasks: {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string | null;
+    }[];
+    version: "1.0.0";
+    category: "reasoning" | "hallucination" | "security" | "bullshit" | "refactoring" | "debugging" | "generation" | "speed";
     methodologyVersion: string;
     seed: string;
-    version: "1.0.0";
-    category: "reasoning" | "hallucination";
     competitors: {
         id: string;
         role: "competitor";
@@ -191,12 +197,6 @@ declare const RunManifestSchema: z.ZodObject<{
             excludeReasoning: boolean;
         };
     }[];
-    tasks: {
-        id: string;
-        version: string;
-        publicHash: string;
-        privateHash: string;
-    }[];
     engineVersion: string;
     matches: number;
     judges: {
@@ -216,10 +216,16 @@ declare const RunManifestSchema: z.ZodObject<{
         judge: string;
     };
 }, {
+    tasks: {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string | null;
+    }[];
+    version: "1.0.0";
+    category: "reasoning" | "hallucination" | "security" | "bullshit" | "refactoring" | "debugging" | "generation" | "speed";
     methodologyVersion: string;
     seed: string;
-    version: "1.0.0";
-    category: "reasoning" | "hallucination";
     competitors: {
         id: string;
         role: "competitor";
@@ -232,12 +238,6 @@ declare const RunManifestSchema: z.ZodObject<{
             excludeReasoning: boolean;
         };
     }[];
-    tasks: {
-        id: string;
-        version: string;
-        publicHash: string;
-        privateHash: string;
-    }[];
     engineVersion: string;
     matches: number;
     judges: {
@@ -261,7 +261,7 @@ 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;
+declare function createRunManifest(config: Pick<ArenaRunConfig, 'category' | 'seed' | 'matches' | 'competitorIds'>, tasks: readonly ArenaTask[], resolvedCompetitors?: readonly ModelRegistryEntry[]): RunManifest;
 
 interface ArenaStoreConfig {
     category: BenchmarkCategory;

package/dist/{logger-QJU7SBDz.d.ts → logger-BQf29BLe.d.ts}

@@ -1,13 +1,13 @@
 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';
+import { f as ArenaRunConfig, t as ModelRegistryEntry, M as MatchResult, E as EloState, h as ArenaSnapshot } from './reports-B8TCJtPr.js';
+import { A as ArenaTask, B as BenchmarkCategory } from './tasks-BmhWuMBD.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"]>;
+    category: z.ZodEnum<["reasoning", "hallucination", "security", "bullshit", "refactoring", "debugging", "generation", "speed"]>;
     seed: z.ZodString;
     matches: z.ZodNumber;
     competitors: z.ZodEffects<z.ZodArray<z.ZodObject<{
@@ -152,17 +152,17 @@ declare const RunManifestSchema: z.ZodObject<{
         id: z.ZodString;
         version: z.ZodString;
         publicHash: z.ZodString;
-        privateHash: z.ZodString;
+        privateHash: z.ZodNullable<z.ZodString>;
     }, "strip", z.ZodTypeAny, {
         id: string;
         version: string;
         publicHash: string;
-        privateHash: string;
+        privateHash: string | null;
     }, {
         id: string;
         version: string;
         publicHash: string;
-        privateHash: string;
+        privateHash: string | null;
     }>, "many">;
     promptPolicyHashes: z.ZodObject<{
         competitor: z.ZodString;
@@ -175,10 +175,16 @@ declare const RunManifestSchema: z.ZodObject<{
         judge: string;
     }>;
 }, "strip", z.ZodTypeAny, {
+    tasks: {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string | null;
+    }[];
+    version: "1.0.0";
+    category: "reasoning" | "hallucination" | "security" | "bullshit" | "refactoring" | "debugging" | "generation" | "speed";
     methodologyVersion: string;
     seed: string;
-    version: "1.0.0";
-    category: "reasoning" | "hallucination";
     competitors: {
         id: string;
         role: "competitor";
@@ -191,12 +197,6 @@ declare const RunManifestSchema: z.ZodObject<{
             excludeReasoning: boolean;
         };
     }[];
-    tasks: {
-        id: string;
-        version: string;
-        publicHash: string;
-        privateHash: string;
-    }[];
     engineVersion: string;
     matches: number;
     judges: {
@@ -216,10 +216,16 @@ declare const RunManifestSchema: z.ZodObject<{
         judge: string;
     };
 }, {
+    tasks: {
+        id: string;
+        version: string;
+        publicHash: string;
+        privateHash: string | null;
+    }[];
+    version: "1.0.0";
+    category: "reasoning" | "hallucination" | "security" | "bullshit" | "refactoring" | "debugging" | "generation" | "speed";
     methodologyVersion: string;
     seed: string;
-    version: "1.0.0";
-    category: "reasoning" | "hallucination";
     competitors: {
         id: string;
         role: "competitor";
@@ -232,12 +238,6 @@ declare const RunManifestSchema: z.ZodObject<{
             excludeReasoning: boolean;
         };
     }[];
-    tasks: {
-        id: string;
-        version: string;
-        publicHash: string;
-        privateHash: string;
-    }[];
     engineVersion: string;
     matches: number;
     judges: {
@@ -261,7 +261,7 @@ 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;
+declare function createRunManifest(config: Pick<ArenaRunConfig, 'category' | 'seed' | 'matches' | 'competitorIds'>, tasks: readonly ArenaTask[], resolvedCompetitors?: readonly ModelRegistryEntry[]): RunManifest;
 
 interface ArenaStoreConfig {
     category: BenchmarkCategory;