bridgebench 3.1.0-alpha.0 → 3.1.0-alpha.1

package/dist/{chunk-JTVNKSMO.js → chunk-3CX4FIFZ.js}

@@ -4,7 +4,7 @@ import {
   JudgeVerdictSchema,
   TaskPrivateSchema,
   TaskPublicSchema
-} from "./chunk-UECBSKTD.js";
+} from "./chunk-UXFKTSQS.js";
 
 // src/tasks.ts
 import { createHash as createHash2 } from "crypto";
@@ -1093,4 +1093,3 @@ export {
   competitorPromptPolicyHash,
   buildCompetitorPrompt
 };
-//# sourceMappingURL=chunk-JTVNKSMO.js.map

package/dist/{chunk-NJTYVNP4.cjs → chunk-5OZGHLDH.cjs}

@@ -13,7 +13,7 @@
 
 
 
-var _chunk4TWPCPRPcjs = require('./chunk-4TWPCPRP.cjs');
+var _chunkSMG3PORScjs = require('./chunk-SMG3PORS.cjs');
 
 
 
@@ -23,7 +23,7 @@ var _chunk4TWPCPRPcjs = require('./chunk-4TWPCPRP.cjs');
 
 
 
-var _chunkCIXITJW6cjs = require('./chunk-CIXITJW6.cjs');
+var _chunk5PCDG7VAcjs = require('./chunk-5PCDG7VA.cjs');
 
 // src/elo.ts
 var ELO_INITIAL = 1e3;
@@ -49,7 +49,7 @@ var PackageMetadataSchema = _zod.z.object({
   version: _zod.z.string().min(1)
 });
 var metadata = PackageMetadataSchema.parse(
-  JSON.parse(_fs.readFileSync.call(void 0, _path2.default.join(_chunk4TWPCPRPcjs.packageRoot.call(void 0, _chunkCIXITJW6cjs.importMetaUrl), "package.json"), "utf8"))
+  JSON.parse(_fs.readFileSync.call(void 0, _path2.default.join(_chunkSMG3PORScjs.packageRoot.call(void 0, _chunk5PCDG7VAcjs.importMetaUrl), "package.json"), "utf8"))
 );
 var PACKAGE_NAME = metadata.name;
 var ENGINE_VERSION = metadata.version;
@@ -80,7 +80,7 @@ var RunManifestSchema = _zod.z.object({
   version: _zod.z.literal(RUN_MANIFEST_VERSION),
   methodologyVersion: _zod.z.string().min(1),
   engineVersion: _zod.z.string().min(1),
-  category: _chunkCIXITJW6cjs.BenchmarkCategorySchema,
+  category: _chunk5PCDG7VAcjs.BenchmarkCategorySchema,
   seed: _zod.z.string(),
   matches: _zod.z.number().int().positive(),
   competitors: _zod.z.array(ManifestCompetitorSchema).min(2).superRefine((models, context) => {
@@ -138,17 +138,17 @@ function runManifestHash(manifest) {
 function runIdFromManifest(manifest) {
   return `run-${runManifestHash(manifest).slice(0, 20)}`;
 }
-function createRunManifest(config, tasks, resolvedCompetitors = _chunk4TWPCPRPcjs.resolveCompetitorRoster.call(void 0, 
+function createRunManifest(config, tasks, resolvedCompetitors = _chunkSMG3PORScjs.resolveCompetitorRoster.call(void 0, 
   config.competitorIds
 )) {
-  const canonicalCompetitors = _chunk4TWPCPRPcjs.resolveCompetitorRoster.call(void 0, 
+  const canonicalCompetitors = _chunkSMG3PORScjs.resolveCompetitorRoster.call(void 0, 
     resolvedCompetitors.map((model) => model.id)
   );
   const competitors = canonicalCompetitors.map((model) => ({ ...manifestModel(model), role: "competitor" })).sort((left, right) => left.id.localeCompare(right.id));
-  const judges = _chunk4TWPCPRPcjs.listModels.call(void 0, "judge").map((model) => ({ ...manifestModel(model), role: "judge" })).sort((left, right) => left.id.localeCompare(right.id));
+  const judges = _chunkSMG3PORScjs.listModels.call(void 0, "judge").map((model) => ({ ...manifestModel(model), role: "judge" })).sort((left, right) => left.id.localeCompare(right.id));
   const manifest = {
     version: RUN_MANIFEST_VERSION,
-    methodologyVersion: _chunkCIXITJW6cjs.METHODOLOGY_VERSION,
+    methodologyVersion: _chunk5PCDG7VAcjs.METHODOLOGY_VERSION,
     engineVersion: ENGINE_VERSION,
     category: config.category,
     seed: config.seed,
@@ -162,8 +162,8 @@ function createRunManifest(config, tasks, resolvedCompetitors = _chunk4TWPCPRPcj
       privateHash: task.privateHash
     })).sort((left, right) => left.id.localeCompare(right.id)),
     promptPolicyHashes: {
-      competitor: _chunk4TWPCPRPcjs.competitorPromptPolicyHash.call(void 0, config.category),
-      judge: _chunk4TWPCPRPcjs.judgePromptPolicyHash.call(void 0, config.category)
+      competitor: _chunkSMG3PORScjs.competitorPromptPolicyHash.call(void 0, config.category),
+      judge: _chunkSMG3PORScjs.judgePromptPolicyHash.call(void 0, config.category)
     }
   };
   return RunManifestSchema.parse(manifest);
@@ -171,7 +171,7 @@ function createRunManifest(config, tasks, resolvedCompetitors = _chunk4TWPCPRPcj
 
 // src/verification.ts
 var LEGACY_METHODOLOGY_VERSION = "reasoning-arena-v0.2.0";
-var SUPPORTED_METHODOLOGY_VERSIONS = /* @__PURE__ */ new Set([LEGACY_METHODOLOGY_VERSION, _chunkCIXITJW6cjs.METHODOLOGY_VERSION]);
+var SUPPORTED_METHODOLOGY_VERSIONS = /* @__PURE__ */ new Set([LEGACY_METHODOLOGY_VERSION, _chunk5PCDG7VAcjs.METHODOLOGY_VERSION]);
 var EPSILON = 1e-9;
 function fail(message) {
   throw new Error(message);
@@ -266,7 +266,7 @@ function verifyOutcome(match) {
   const judgeCost = _nullishCoalesce(_optionalChain([match, 'access', _6 => _6.panel, 'optionalAccess', _7 => _7.votes, 'access', _8 => _8.reduce, 'call', _9 => _9((sum, vote) => sum + (_nullishCoalesce(_optionalChain([vote, 'access', _10 => _10.completion, 'optionalAccess', _11 => _11.costUsd]), () => ( 0))), 0)]), () => ( 0));
   equalNumber(
     match.matchCostUsd,
-    _chunkCIXITJW6cjs.competitorCost.call(void 0, responseA) + _chunkCIXITJW6cjs.competitorCost.call(void 0, responseB) + judgeCost,
+    _chunk5PCDG7VAcjs.competitorCost.call(void 0, responseA) + _chunk5PCDG7VAcjs.competitorCost.call(void 0, responseB) + judgeCost,
     "matchCostUsd"
   );
 }
@@ -420,7 +420,7 @@ function ensureEntry(entries, modelId) {
   if (existing) return existing;
   let displayName = modelId;
   try {
-    displayName = _chunk4TWPCPRPcjs.getModel.call(void 0, modelId).displayName;
+    displayName = _chunkSMG3PORScjs.getModel.call(void 0, modelId).displayName;
   } catch (e) {
   }
   const created = emptyEntry(modelId, displayName);
@@ -430,7 +430,7 @@ function ensureEntry(entries, modelId) {
 function buildSnapshot(matches, category, options = {}) {
   const verified = verifyJournal(matches, category, options);
   const entries = /* @__PURE__ */ new Map();
-  const competitorIds = _nullishCoalesce(options.competitorIds, () => ( _chunk4TWPCPRPcjs.listModels.call(void 0, "competitor").map((model) => model.id)));
+  const competitorIds = _nullishCoalesce(options.competitorIds, () => ( _chunkSMG3PORScjs.listModels.call(void 0, "competitor").map((model) => model.id)));
   for (const modelId of competitorIds) {
     ensureEntry(entries, modelId);
   }
@@ -439,8 +439,8 @@ function buildSnapshot(matches, category, options = {}) {
     const modelB = ensureEntry(entries, match.competitors.modelB);
     modelA.elo = _nullishCoalesce(match.eloAfter[modelA.modelId], () => ( modelA.elo));
     modelB.elo = _nullishCoalesce(match.eloAfter[modelB.modelId], () => ( modelB.elo));
-    modelA.totalCostUsd += _chunkCIXITJW6cjs.competitorCost.call(void 0, match.competitors.responseA);
-    modelB.totalCostUsd += _chunkCIXITJW6cjs.competitorCost.call(void 0, match.competitors.responseB);
+    modelA.totalCostUsd += _chunk5PCDG7VAcjs.competitorCost.call(void 0, match.competitors.responseA);
+    modelB.totalCostUsd += _chunk5PCDG7VAcjs.competitorCost.call(void 0, match.competitors.responseB);
     if (match.outcome === "no-contest" || !match.winnerModelId) continue;
     modelA.matches += 1;
     modelB.matches += 1;
@@ -465,7 +465,7 @@ function buildSnapshot(matches, category, options = {}) {
   ).map((entry, index) => ({ rank: index + 1, ...entry }));
   return {
     version: "0.2.0",
-    methodologyVersion: _nullishCoalesce(verified.methodologyVersion, () => ( _chunkCIXITJW6cjs.METHODOLOGY_VERSION)),
+    methodologyVersion: _nullishCoalesce(verified.methodologyVersion, () => ( _chunk5PCDG7VAcjs.METHODOLOGY_VERSION)),
     category,
     generatedAt: (/* @__PURE__ */ new Date()).toISOString(),
     initialElo: ELO_INITIAL,
@@ -479,7 +479,7 @@ function incrementCluster(current, field) {
   return { ...next, [field]: next[field] + 1 };
 }
 function renderMarkdown(snapshot) {
-  const meta = _chunkCIXITJW6cjs.CATEGORY_META[snapshot.category];
+  const meta = _chunk5PCDG7VAcjs.CATEGORY_META[snapshot.category];
   const modelNames = new Map(
     snapshot.leaderboard.map((entry) => [entry.modelId, entry.displayName])
   );
@@ -567,7 +567,7 @@ function scheduleMatches(input) {
     throw new Error("Competitor model IDs must be unique");
   const random = mulberry32(hashSeed(`${input.category}|${input.seed}`));
   const runId = _nullishCoalesce(input.runId, () => ( `run-${stableId([
-    _chunkCIXITJW6cjs.METHODOLOGY_VERSION,
+    _chunk5PCDG7VAcjs.METHODOLOGY_VERSION,
     input.category,
     input.seed,
     String(input.count),
@@ -821,17 +821,17 @@ function failedResponse(modelId, error, latencyMs) {
   return {
     modelId,
     success: false,
-    error: _chunk4TWPCPRPcjs.sanitizeError.call(void 0, error),
+    error: _chunkSMG3PORScjs.sanitizeError.call(void 0, error),
     latencyMs
   };
 }
 var ArenaRunner = class {
-  constructor(gateway, store, onEvent, logger = _chunk4TWPCPRPcjs.noopLogger) {
+  constructor(gateway, store, onEvent, logger = _chunkSMG3PORScjs.noopLogger) {
     this.gateway = gateway;
     this.store = store;
     this.onEvent = onEvent;
     this.logger = logger;
-    this.judges = new (0, _chunk4TWPCPRPcjs.JudgePanel)(gateway, onEvent, logger);
+    this.judges = new (0, _chunkSMG3PORScjs.JudgePanel)(gateway, onEvent, logger);
   }
   
   
@@ -892,7 +892,7 @@ var ArenaRunner = class {
     };
   }
   async prepareRun(config, tasks, signal) {
-    const competitors = _chunk4TWPCPRPcjs.resolveCompetitorRoster.call(void 0, config.competitorIds);
+    const competitors = _chunkSMG3PORScjs.resolveCompetitorRoster.call(void 0, config.competitorIds);
     const competitorIds = competitors.map((model) => model.id);
     const mismatched = tasks.filter((task) => task.public.category !== config.category);
     if (mismatched.length > 0) {
@@ -901,10 +901,10 @@ var ArenaRunner = class {
       );
     }
     const journal = this.store.readAll();
-    const incompatible = journal.find((match) => match.methodologyVersion !== _chunkCIXITJW6cjs.METHODOLOGY_VERSION);
+    const incompatible = journal.find((match) => match.methodologyVersion !== _chunk5PCDG7VAcjs.METHODOLOGY_VERSION);
     if (incompatible) {
       throw new Error(
-        `Cannot append ${_chunkCIXITJW6cjs.METHODOLOGY_VERSION} matches to a ${incompatible.methodologyVersion} journal; archive it and start a new journal`
+        `Cannot append ${_chunk5PCDG7VAcjs.METHODOLOGY_VERSION} matches to a ${incompatible.methodologyVersion} journal; archive it and start a new journal`
       );
     }
     this.store.rebuildEloState(competitorIds);
@@ -937,7 +937,7 @@ var ArenaRunner = class {
     if (_optionalChain([signal, 'optionalAccess', _25 => _25.aborted])) return prepared;
     try {
       await Promise.all(
-        [...competitors, ..._chunk4TWPCPRPcjs.listModels.call(void 0, "judge")].map(
+        [...competitors, ..._chunkSMG3PORScjs.listModels.call(void 0, "judge")].map(
           (model) => this.gateway.validateModel(model, signal)
         )
       );
@@ -1020,7 +1020,7 @@ var ArenaRunner = class {
         try {
           result = await this.runMatch(match, task, manifestHash, competitorIds, signal);
         } catch (error) {
-          if (_optionalChain([signal, 'optionalAccess', _30 => _30.aborted]) || _chunk4TWPCPRPcjs.isArenaCancellationError.call(void 0, error)) {
+          if (_optionalChain([signal, 'optionalAccess', _30 => _30.aborted]) || _chunkSMG3PORScjs.isArenaCancellationError.call(void 0, error)) {
             cancellation.requested();
             return this.cancelRun(runId, progress);
           }
@@ -1094,14 +1094,14 @@ var ArenaRunner = class {
     }
   }
   async runMatch(match, task, manifestHash, competitorIds, signal) {
-    _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+    _chunkSMG3PORScjs.throwIfCancelled.call(void 0, signal);
     const state = this.store.rebuildEloState(competitorIds);
-    const prompt = _chunk4TWPCPRPcjs.buildCompetitorPrompt.call(void 0, task);
+    const prompt = _chunkSMG3PORScjs.buildCompetitorPrompt.call(void 0, task);
     const [responseA, responseB] = await Promise.all([
       this.runCompetitor(match, "A", prompt, signal),
       this.runCompetitor(match, "B", prompt, signal)
     ]);
-    _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+    _chunkSMG3PORScjs.throwIfCancelled.call(void 0, signal);
     this.inspectResponse(match, responseA);
     this.inspectResponse(match, responseB);
     this.emit({
@@ -1112,20 +1112,20 @@ var ArenaRunner = class {
         modelA: {
           success: responseA.success,
           latencyMs: responseA.latencyMs,
-          costUsd: _chunkCIXITJW6cjs.competitorCost.call(void 0, responseA),
-          outputTokens: _chunkCIXITJW6cjs.competitorOutputTokens.call(void 0, responseA),
-          reasoningTokens: _nullishCoalesce(_chunkCIXITJW6cjs.competitorReasoningTokens.call(void 0, responseA), () => ( null))
+          costUsd: _chunk5PCDG7VAcjs.competitorCost.call(void 0, responseA),
+          outputTokens: _chunk5PCDG7VAcjs.competitorOutputTokens.call(void 0, responseA),
+          reasoningTokens: _nullishCoalesce(_chunk5PCDG7VAcjs.competitorReasoningTokens.call(void 0, responseA), () => ( null))
         },
         modelB: {
           success: responseB.success,
           latencyMs: responseB.latencyMs,
-          costUsd: _chunkCIXITJW6cjs.competitorCost.call(void 0, responseB),
-          outputTokens: _chunkCIXITJW6cjs.competitorOutputTokens.call(void 0, responseB),
-          reasoningTokens: _nullishCoalesce(_chunkCIXITJW6cjs.competitorReasoningTokens.call(void 0, responseB), () => ( null))
+          costUsd: _chunk5PCDG7VAcjs.competitorCost.call(void 0, responseB),
+          outputTokens: _chunk5PCDG7VAcjs.competitorOutputTokens.call(void 0, responseB),
+          reasoningTokens: _nullishCoalesce(_chunk5PCDG7VAcjs.competitorReasoningTokens.call(void 0, responseB), () => ( null))
         }
       }
     });
-    _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+    _chunkSMG3PORScjs.throwIfCancelled.call(void 0, signal);
     const eloBefore = {
       [match.modelA]: _nullishCoalesce(state.ratings[match.modelA], () => ( ELO_INITIAL)),
       [match.modelB]: _nullishCoalesce(state.ratings[match.modelB], () => ( ELO_INITIAL))
@@ -1141,10 +1141,10 @@ var ArenaRunner = class {
       this.emit({
         id: `${match.id}-judging-started`,
         type: "judging.started",
-        data: { matchId: match.id, judges: _chunk4TWPCPRPcjs.listModels.call(void 0, "judge").map((model) => model.id) }
+        data: { matchId: match.id, judges: _chunkSMG3PORScjs.listModels.call(void 0, "judge").map((model) => model.id) }
       });
       panel = await this.judges.judge({ match, task, responseA, responseB }, signal);
-      _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+      _chunkSMG3PORScjs.throwIfCancelled.call(void 0, signal);
       if (panel.winnerModelId) {
         outcome = "judged";
         winnerModelId = panel.winnerModelId;
@@ -1161,7 +1161,7 @@ var ArenaRunner = class {
     }
     const judgeCost = _nullishCoalesce(_optionalChain([panel, 'optionalAccess', _34 => _34.votes, 'access', _35 => _35.reduce, 'call', _36 => _36((sum, vote) => sum + (_nullishCoalesce(_optionalChain([vote, 'access', _37 => _37.completion, 'optionalAccess', _38 => _38.costUsd]), () => ( 0))), 0)]), () => ( 0));
     return {
-      methodologyVersion: _chunkCIXITJW6cjs.METHODOLOGY_VERSION,
+      methodologyVersion: _chunk5PCDG7VAcjs.METHODOLOGY_VERSION,
       runId: match.runId,
       runManifestHash: manifestHash,
       matchId: match.id,
@@ -1183,7 +1183,7 @@ var ArenaRunner = class {
       eloBefore,
       eloAfter,
       pointAwarded: winnerModelId !== null,
-      matchCostUsd: _chunkCIXITJW6cjs.competitorCost.call(void 0, responseA) + _chunkCIXITJW6cjs.competitorCost.call(void 0, responseB) + judgeCost
+      matchCostUsd: _chunk5PCDG7VAcjs.competitorCost.call(void 0, responseA) + _chunk5PCDG7VAcjs.competitorCost.call(void 0, responseB) + judgeCost
     };
   }
   /** Surface signals that make a journaled response untrustworthy or a task too easy. */
@@ -1226,19 +1226,19 @@ var ArenaRunner = class {
       });
     };
     try {
-      _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+      _chunkSMG3PORScjs.throwIfCancelled.call(void 0, signal);
       const completion = await this.gateway.complete({
-        model: _chunk4TWPCPRPcjs.getModel.call(void 0, modelId),
+        model: _chunkSMG3PORScjs.getModel.call(void 0, modelId),
         ...prompt,
         signal,
         onDelta: (text) => emitDelta(text, false)
       });
-      _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+      _chunkSMG3PORScjs.throwIfCancelled.call(void 0, signal);
       emitDelta(completion.content, true);
       return { modelId, success: true, ...completion };
     } catch (error) {
-      if (_optionalChain([signal, 'optionalAccess', _39 => _39.aborted]) || _chunk4TWPCPRPcjs.isArenaCancellationError.call(void 0, error)) {
-        throw new (0, _chunk4TWPCPRPcjs.ArenaCancellationError)();
+      if (_optionalChain([signal, 'optionalAccess', _39 => _39.aborted]) || _chunkSMG3PORScjs.isArenaCancellationError.call(void 0, error)) {
+        throw new (0, _chunkSMG3PORScjs.ArenaCancellationError)();
       }
       const failed = failedResponse(modelId, error, Date.now() - startedAt);
       emitDelta(failed.error, true, false);
@@ -1288,7 +1288,7 @@ var MockOpenRouterGateway = class {
         },
         violations: []
       };
-      void _chunk4TWPCPRPcjs.judgeVerdictJsonSchema.call(void 0, );
+      void _chunkSMG3PORScjs.judgeVerdictJsonSchema.call(void 0, );
       return baseCompletion(JSON.stringify(payload), `mock-judge-${request.model.id}`);
     }
     const variantA = [
@@ -1371,7 +1371,7 @@ var ArenaStore = class {
       } catch (e2) {
         throw new Error(`Malformed arena journal line ${index + 1}; refusing to continue`);
       }
-      const parsed = _chunkCIXITJW6cjs.MatchResultSchema.safeParse(decoded);
+      const parsed = _chunk5PCDG7VAcjs.MatchResultSchema.safeParse(decoded);
       if (!parsed.success) {
         const issue = parsed.error.issues[0];
         const location = _optionalChain([issue, 'optionalAccess', _46 => _46.path, 'access', _47 => _47.length]) ? ` at ${issue.path.join(".")}` : "";
@@ -1409,7 +1409,7 @@ var ArenaStore = class {
   completedMatchIds() {
     return new Set(this.readAll().map((result) => result.matchId));
   }
-  rebuildEloState(competitorIds = _chunk4TWPCPRPcjs.listModels.call(void 0, "competitor").map((model) => model.id)) {
+  rebuildEloState(competitorIds = _chunkSMG3PORScjs.listModels.call(void 0, "competitor").map((model) => model.id)) {
     const ratings = Object.fromEntries(competitorIds.map((modelId) => [modelId, ELO_INITIAL]));
     const points = Object.fromEntries(competitorIds.map((modelId) => [modelId, 0]));
     const verified = verifyJournal(this.readAll(), this.category, {
@@ -1464,4 +1464,3 @@ var ArenaStore = class {
 
 
 exports.ELO_INITIAL = ELO_INITIAL; exports.ELO_K = ELO_K; exports.expectedScore = expectedScore; exports.applyEloWin = applyEloWin; exports.PACKAGE_NAME = PACKAGE_NAME; exports.ENGINE_VERSION = ENGINE_VERSION; exports.RUN_MANIFEST_VERSION = RUN_MANIFEST_VERSION; exports.RunManifestSchema = RunManifestSchema; exports.canonicalJson = canonicalJson; exports.runManifestHash = runManifestHash; exports.runIdFromManifest = runIdFromManifest; exports.createRunManifest = createRunManifest; exports.verifyJournal = verifyJournal; exports.verifiedEloState = verifiedEloState; exports.buildSnapshot = buildSnapshot; exports.renderMarkdown = renderMarkdown; exports.writeReports = writeReports; exports.scheduleMatches = scheduleMatches; exports.FAST_RESPONSE_MS = FAST_RESPONSE_MS; exports.LOW_OUTPUT_TOKENS = LOW_OUTPUT_TOKENS; exports.classifyError = classifyError; exports.detectResponseAnomalies = detectResponseAnomalies; exports.triageJournal = triageJournal; exports.formatTriage = formatTriage; exports.ArenaRunner = ArenaRunner; exports.MockOpenRouterGateway = MockOpenRouterGateway; exports.categoryStoreConfig = categoryStoreConfig; exports.ArenaStore = ArenaStore;
-//# sourceMappingURL=chunk-NJTYVNP4.cjs.map

package/dist/{chunk-CIXITJW6.cjs → chunk-5PCDG7VA.cjs}

@@ -246,4 +246,3 @@ var CONTRACTS_VERSION = "1.0.0";
 
 
 exports.importMetaUrl = importMetaUrl; exports.METHODOLOGY_VERSION = METHODOLOGY_VERSION; exports.BenchmarkCategorySchema = BenchmarkCategorySchema; exports.CATEGORIES = CATEGORIES; exports.CATEGORY_CLUSTERS = CATEGORY_CLUSTERS; exports.CATEGORY_META = CATEGORY_META; exports.ModelRoleSchema = ModelRoleSchema; exports.ModelCompletionSchema = ModelCompletionSchema; exports.CompetitorResponseSchema = CompetitorResponseSchema; exports.JudgeVerdictSchema = JudgeVerdictSchema; exports.JUDGE_VERDICT_TRANSPORT_SCHEMA = JUDGE_VERDICT_TRANSPORT_SCHEMA; exports.JudgeVoteSchema = JudgeVoteSchema; exports.PanelDecisionSchema = PanelDecisionSchema; exports.MatchResultSchema = MatchResultSchema; exports.competitorCost = competitorCost; exports.competitorOutputTokens = competitorOutputTokens; exports.competitorReasoningTokens = competitorReasoningTokens; exports.competitorContent = competitorContent; exports.TaskArtifactSchema = TaskArtifactSchema; exports.TaskPublicSchema = TaskPublicSchema; exports.TaskPrivateSchema = TaskPrivateSchema; exports.CONTRACTS_VERSION = CONTRACTS_VERSION;
-//# sourceMappingURL=chunk-CIXITJW6.cjs.map

package/dist/{chunk-EQHRUV2I.js → chunk-KGZEQ6R5.js}

@@ -13,7 +13,7 @@ import {
   resolveCompetitorRoster,
   sanitizeError,
   throwIfCancelled
-} from "./chunk-JTVNKSMO.js";
+} from "./chunk-3CX4FIFZ.js";
 import {
   BenchmarkCategorySchema,
   CATEGORY_META,
@@ -22,7 +22,7 @@ import {
   competitorCost,
   competitorOutputTokens,
   competitorReasoningTokens
-} from "./chunk-UECBSKTD.js";
+} from "./chunk-UXFKTSQS.js";
 
 // src/elo.ts
 var ELO_INITIAL = 1e3;
@@ -1463,4 +1463,3 @@ export {
   categoryStoreConfig,
   ArenaStore
 };
-//# sourceMappingURL=chunk-EQHRUV2I.js.map

package/dist/{chunk-LFKEV2YL.js → chunk-MBC4C3VA.js}

@@ -6,17 +6,17 @@ import {
   createRunManifest,
   runIdFromManifest,
   runManifestHash
-} from "./chunk-EQHRUV2I.js";
+} from "./chunk-KGZEQ6R5.js";
 import {
   OpenRouterClient,
   TaskLoader,
   noopLogger,
   redactSecrets,
   resolveCompetitorRoster
-} from "./chunk-JTVNKSMO.js";
+} from "./chunk-3CX4FIFZ.js";
 import {
   METHODOLOGY_VERSION
-} from "./chunk-UECBSKTD.js";
+} from "./chunk-UXFKTSQS.js";
 
 // src/api-client.ts
 var ADMIN_KEY_HEADER = "x-bridgebench-admin-key";
@@ -395,4 +395,3 @@ export {
   createRemoteRun,
   runRemoteArena
 };
-//# sourceMappingURL=chunk-LFKEV2YL.js.map

package/dist/{chunk-4TWPCPRP.cjs → chunk-SMG3PORS.cjs}

@@ -5,7 +5,7 @@
 
 
 
-var _chunkCIXITJW6cjs = require('./chunk-CIXITJW6.cjs');
+var _chunk5PCDG7VAcjs = require('./chunk-5PCDG7VA.cjs');
 
 // src/tasks.ts
 var _crypto = require('crypto');
@@ -269,7 +269,7 @@ function assertPromptSize(request) {
   }
 }
 function judgeVerdictJsonSchema() {
-  return _chunkCIXITJW6cjs.JUDGE_VERDICT_TRANSPORT_SCHEMA;
+  return _chunk5PCDG7VAcjs.JUDGE_VERDICT_TRANSPORT_SCHEMA;
 }
 function judgeResponseFormat() {
   return {
@@ -558,7 +558,7 @@ var OpenRouterClient = class {
 };
 function parseJudgeVerdict(content) {
   const trimmed = content.trim().replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/, "");
-  return _chunkCIXITJW6cjs.JudgeVerdictSchema.parse(JSON.parse(trimmed));
+  return _chunk5PCDG7VAcjs.JudgeVerdictSchema.parse(JSON.parse(trimmed));
 }
 
 // src/judges.ts
@@ -837,7 +837,7 @@ function findProjectRoot(fromUrl) {
 }
 
 // src/tasks.ts
-var ROOT = packageRoot(_chunkCIXITJW6cjs.importMetaUrl);
+var ROOT = packageRoot(_chunk5PCDG7VAcjs.importMetaUrl);
 var TASKS_PER_CATEGORY = 12;
 var TASKS_PER_CLUSTER = 2;
 var APPROXIMATE_CHARS_PER_TOKEN = 4;
@@ -875,7 +875,7 @@ function validatePublicTask(file, task, category) {
   if (task.category !== category) {
     throw new Error(`${file} declares category ${task.category} inside the ${category} pack`);
   }
-  if (!_chunkCIXITJW6cjs.CATEGORY_CLUSTERS[category].includes(task.cluster)) {
+  if (!_chunk5PCDG7VAcjs.CATEGORY_CLUSTERS[category].includes(task.cluster)) {
     throw new Error(`${file} uses cluster ${task.cluster}, which is not a ${category} cluster`);
   }
   const artifactIds = task.artifacts.map((artifact) => artifact.id);
@@ -929,7 +929,7 @@ function validatePackComposition(category, tasks) {
   if (duplicates.length > 0) {
     throw new Error(`${category} task IDs must be unique: ${duplicates.join(", ")}`);
   }
-  for (const cluster of _chunkCIXITJW6cjs.CATEGORY_CLUSTERS[category]) {
+  for (const cluster of _chunk5PCDG7VAcjs.CATEGORY_CLUSTERS[category]) {
     const count = tasks.filter((task) => task.public.cluster === cluster).length;
     if (count !== TASKS_PER_CLUSTER) {
       throw new Error(`Expected exactly ${TASKS_PER_CLUSTER} ${cluster} tasks, found ${count}`);
@@ -941,7 +941,7 @@ function validatePackComposition(category, tasks) {
 }
 async function validatePublicTaskFile(filePath) {
   const resolved = _path2.default.resolve(filePath);
-  const { raw, value } = await parseYamlFile(resolved, _chunkCIXITJW6cjs.TaskPublicSchema);
+  const { raw, value } = await parseYamlFile(resolved, _chunk5PCDG7VAcjs.TaskPublicSchema);
   validatePublicTask(_path2.default.basename(resolved), value, value.category);
   const task = {
     public: value,
@@ -979,14 +979,14 @@ var TaskLoader = class {
       const publicFile = _path2.default.join(publicDir, file);
       const { raw: publicRaw, value: publicTask } = await parseYamlFile(
         publicFile,
-        _chunkCIXITJW6cjs.TaskPublicSchema
+        _chunk5PCDG7VAcjs.TaskPublicSchema
       );
       validatePublicTask(file, publicTask, this.category);
       let privateRaw = null;
       let privateTask = null;
       if (havePrivate) {
         try {
-          const loaded = await parseYamlFile(_path2.default.join(this.privateDir, file), _chunkCIXITJW6cjs.TaskPrivateSchema);
+          const loaded = await parseYamlFile(_path2.default.join(this.privateDir, file), _chunk5PCDG7VAcjs.TaskPrivateSchema);
           privateRaw = loaded.raw;
           privateTask = loaded.value;
         } catch (error) {
@@ -1020,7 +1020,7 @@ function mergePrivateHalves(tasks, privateHalves) {
     if (!/^[a-f0-9]{64}$/.test(half.hash)) {
       throw new Error(`Private hash for task ${task.public.id} is not a sha256 digest`);
     }
-    const privateTask = _chunkCIXITJW6cjs.TaskPrivateSchema.parse(half.value);
+    const privateTask = _chunk5PCDG7VAcjs.TaskPrivateSchema.parse(half.value);
     validatePrivatePair(task.public.id, task.public, privateTask);
     const complete = {
       public: task.public,
@@ -1094,4 +1094,3 @@ ${artifacts}`
 
 
 exports.ArenaCancellationError = ArenaCancellationError; exports.isArenaCancellationError = isArenaCancellationError; exports.throwIfCancelled = throwIfCancelled; exports.redactSecrets = redactSecrets; exports.noopLogger = noopLogger; exports.FileArenaLogger = FileArenaLogger; exports.MODEL_REGISTRY = MODEL_REGISTRY; exports.SOL_FABLE_PILOT_COMPETITOR_IDS = SOL_FABLE_PILOT_COMPETITOR_IDS; exports.listModels = listModels; exports.resolveCompetitorRoster = resolveCompetitorRoster; exports.getModel = getModel; exports.MAX_PROMPT_CHARS = MAX_PROMPT_CHARS; exports.assertPromptSize = assertPromptSize; exports.judgeVerdictJsonSchema = judgeVerdictJsonSchema; exports.runOpenRouterAttempt = runOpenRouterAttempt; exports.isRetryableError = isRetryableError; exports.sanitizeError = sanitizeError; exports.OpenRouterClient = OpenRouterClient; exports.parseJudgeVerdict = parseJudgeVerdict; exports.anonymizeCompetitorOutput = anonymizeCompetitorOutput; exports.buildJudgePayload = buildJudgePayload; exports.judgePromptPolicyHash = judgePromptPolicyHash; exports.judgeSystemPrompt = judgeSystemPrompt; exports.JudgePanel = JudgePanel; exports.packageRoot = packageRoot; exports.findProjectRoot = findProjectRoot; exports.TASKS_PER_CATEGORY = TASKS_PER_CATEGORY; exports.TASKS_PER_CLUSTER = TASKS_PER_CLUSTER; exports.defaultTaskRoot = defaultTaskRoot; exports.validatePublicTaskFile = validatePublicTaskFile; exports.TaskLoader = TaskLoader; exports.mergePrivateHalves = mergePrivateHalves; exports.competitorPromptPolicyHash = competitorPromptPolicyHash; exports.buildCompetitorPrompt = buildCompetitorPrompt;
-//# sourceMappingURL=chunk-4TWPCPRP.cjs.map

package/dist/{chunk-UECBSKTD.js → chunk-UXFKTSQS.js}

@@ -241,4 +241,3 @@ export {
   TaskPrivateSchema,
   CONTRACTS_VERSION
 };
-//# sourceMappingURL=chunk-UECBSKTD.js.map

package/dist/{chunk-7YCJSOK7.cjs → chunk-XPFZBHV2.cjs}

@@ -6,17 +6,17 @@
 
 
 
-var _chunkNJTYVNP4cjs = require('./chunk-NJTYVNP4.cjs');
+var _chunk5OZGHLDHcjs = require('./chunk-5OZGHLDH.cjs');
 
 
 
 
 
 
-var _chunk4TWPCPRPcjs = require('./chunk-4TWPCPRP.cjs');
+var _chunkSMG3PORScjs = require('./chunk-SMG3PORS.cjs');
 
 
-var _chunkCIXITJW6cjs = require('./chunk-CIXITJW6.cjs');
+var _chunk5PCDG7VAcjs = require('./chunk-5PCDG7VA.cjs');
 
 // src/api-client.ts
 var ADMIN_KEY_HEADER = "x-bridgebench-admin-key";
@@ -74,7 +74,7 @@ async function getJson(config, path2, schema) {
   });
   const text = await response.text();
   if (!response.ok) {
-    throw new Error(`GET ${path2} \u2192 ${response.status}: ${_chunk4TWPCPRPcjs.redactSecrets.call(void 0, text).slice(0, 500)}`);
+    throw new Error(`GET ${path2} \u2192 ${response.status}: ${_chunkSMG3PORScjs.redactSecrets.call(void 0, text).slice(0, 500)}`);
   }
   let decoded;
   try {
@@ -110,7 +110,7 @@ async function postJson(config, path2, body, schema) {
     break;
   }
   if (!response.ok) {
-    throw new Error(`POST ${path2} \u2192 ${response.status}: ${_chunk4TWPCPRPcjs.redactSecrets.call(void 0, text).slice(0, 500)}`);
+    throw new Error(`POST ${path2} \u2192 ${response.status}: ${_chunkSMG3PORScjs.redactSecrets.call(void 0, text).slice(0, 500)}`);
   }
   let decoded;
   try {
@@ -138,7 +138,7 @@ async function postChunks(config, path2, batches, wrap, schema) {
 // src/publish.ts
 var _zod = require('zod');
 async function publishTasks(category, config = resolveApiConfig()) {
-  const loaded = await new (0, _chunk4TWPCPRPcjs.TaskLoader)(category).loadAll({
+  const loaded = await new (0, _chunkSMG3PORScjs.TaskLoader)(category).loadAll({
     requirePrivate: true
   });
   const results = await postChunks(
@@ -173,7 +173,7 @@ async function publishJournalFromStore(store, config = resolveApiConfig()) {
   return { imported, skipped, matches: journal.length };
 }
 async function publishJournal(category, config = resolveApiConfig()) {
-  return publishJournalFromStore(new (0, _chunkNJTYVNP4cjs.ArenaStore)(_chunkNJTYVNP4cjs.categoryStoreConfig.call(void 0, category)), config);
+  return publishJournalFromStore(new (0, _chunk5OZGHLDHcjs.ArenaStore)(_chunk5OZGHLDHcjs.categoryStoreConfig.call(void 0, category)), config);
 }
 
 // src/remote-events.ts
@@ -297,23 +297,23 @@ async function fetchExecutionPack(config, category) {
     `/arena/admin/${category}/execution-pack`,
     ExecutionPackSchema
   );
-  if (pack.methodologyVersion !== _chunkCIXITJW6cjs.METHODOLOGY_VERSION) {
+  if (pack.methodologyVersion !== _chunk5PCDG7VAcjs.METHODOLOGY_VERSION) {
     throw new Error(
-      `Execution pack methodology ${pack.methodologyVersion} does not match engine ${_chunkCIXITJW6cjs.METHODOLOGY_VERSION}`
+      `Execution pack methodology ${pack.methodologyVersion} does not match engine ${_chunk5PCDG7VAcjs.METHODOLOGY_VERSION}`
     );
   }
   return { tasks: pack.tasks };
 }
 async function createRemoteRun(apiConfig, runConfig, tasks) {
-  const manifest = _chunkNJTYVNP4cjs.createRunManifest.call(void 0, runConfig, [...tasks]);
-  const manifestHash = _chunkNJTYVNP4cjs.runManifestHash.call(void 0, manifest);
-  const runKey = _chunkNJTYVNP4cjs.runIdFromManifest.call(void 0, manifest);
-  const competitorIds = _chunk4TWPCPRPcjs.resolveCompetitorRoster.call(void 0, runConfig.competitorIds).map((model) => model.id);
+  const manifest = _chunk5OZGHLDHcjs.createRunManifest.call(void 0, runConfig, [...tasks]);
+  const manifestHash = _chunk5OZGHLDHcjs.runManifestHash.call(void 0, manifest);
+  const runKey = _chunk5OZGHLDHcjs.runIdFromManifest.call(void 0, manifest);
+  const competitorIds = _chunkSMG3PORScjs.resolveCompetitorRoster.call(void 0, runConfig.competitorIds).map((model) => model.id);
   const body = {
     runKey,
     category: runConfig.category,
     seed: runConfig.seed,
-    methodologyVersion: _chunkCIXITJW6cjs.METHODOLOGY_VERSION,
+    methodologyVersion: _chunk5PCDG7VAcjs.METHODOLOGY_VERSION,
     competitorIds,
     scheduledCount: runConfig.matches,
     maxCostUsd: runConfig.maxCostUsd,
@@ -324,7 +324,7 @@ async function createRemoteRun(apiConfig, runConfig, tasks) {
   return { runKey: response.run.runKey, created: response.created };
 }
 async function runRemoteArena(apiConfig, options) {
-  const { config, mock = false, publishMatches = true, logger = _chunk4TWPCPRPcjs.noopLogger } = options;
+  const { config, mock = false, publishMatches = true, logger = _chunkSMG3PORScjs.noopLogger } = options;
   const { tasks } = await fetchExecutionPack(apiConfig, config.category);
   const { runKey } = await createRemoteRun(apiConfig, config, tasks);
   const resultsRoot = _path2.default.join(
@@ -332,10 +332,10 @@ async function runRemoteArena(apiConfig, options) {
     "remote",
     config.category
   );
-  const store = new (0, _chunkNJTYVNP4cjs.ArenaStore)(remoteStoreConfig(config.category, resultsRoot));
+  const store = new (0, _chunk5OZGHLDHcjs.ArenaStore)(remoteStoreConfig(config.category, resultsRoot));
   const eventSink = new RemoteArenaEventSink(apiConfig, runKey);
-  const gateway = mock ? new (0, _chunkNJTYVNP4cjs.MockOpenRouterGateway)({ judgeWinner: "MODEL_A" }) : new (0, _chunk4TWPCPRPcjs.OpenRouterClient)(_nullishCoalesce(process.env.OPENROUTER_API_KEY, () => ( "")), logger);
-  const runner = new (0, _chunkNJTYVNP4cjs.ArenaRunner)(gateway, store, eventSink.sink, logger);
+  const gateway = mock ? new (0, _chunk5OZGHLDHcjs.MockOpenRouterGateway)({ judgeWinner: "MODEL_A" }) : new (0, _chunkSMG3PORScjs.OpenRouterClient)(_nullishCoalesce(process.env.OPENROUTER_API_KEY, () => ( "")), logger);
+  const runner = new (0, _chunk5OZGHLDHcjs.ArenaRunner)(gateway, store, eventSink.sink, logger);
   const cancellation = new AbortController();
   const requestCancellation = () => {
     if (cancellation.signal.aborted) return;
@@ -395,4 +395,3 @@ async function runRemoteArena(apiConfig, options) {
 
 
 exports.ADMIN_KEY_HEADER = ADMIN_KEY_HEADER; exports.DEFAULT_API_TIMEOUT_MS = DEFAULT_API_TIMEOUT_MS; exports.REQUEST_SPACING_MS = REQUEST_SPACING_MS; exports.resolveApiConfig = resolveApiConfig; exports.publishTarget = publishTarget; exports.delay = delay; exports.chunk = chunk; exports.getJson = getJson; exports.postJson = postJson; exports.postChunks = postChunks; exports.publishTasks = publishTasks; exports.publishJournalFromStore = publishJournalFromStore; exports.publishJournal = publishJournal; exports.RemoteArenaEventSink = RemoteArenaEventSink; exports.fetchExecutionPack = fetchExecutionPack; exports.createRemoteRun = createRemoteRun; exports.runRemoteArena = runRemoteArena;
-//# sourceMappingURL=chunk-7YCJSOK7.cjs.map