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

package/dist/{chunk-NJTYVNP4.cjs → chunk-VAS6KNJA.cjs}

@@ -13,7 +13,7 @@
 
 
 
-var _chunk4TWPCPRPcjs = require('./chunk-4TWPCPRP.cjs');
+var _chunkQMOPRKWDcjs = require('./chunk-QMOPRKWD.cjs');
 
 
 
@@ -23,7 +23,9 @@ var _chunk4TWPCPRPcjs = require('./chunk-4TWPCPRP.cjs');
 
 
 
-var _chunkCIXITJW6cjs = require('./chunk-CIXITJW6.cjs');
+
+
+var _chunkX3LPZGHScjs = require('./chunk-X3LPZGHS.cjs');
 
 // src/elo.ts
 var ELO_INITIAL = 1e3;
@@ -49,7 +51,7 @@ 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(_chunkQMOPRKWDcjs.packageRoot.call(void 0, _chunkX3LPZGHScjs.importMetaUrl), "package.json"), "utf8"))
 );
 var PACKAGE_NAME = metadata.name;
 var ENGINE_VERSION = metadata.version;
@@ -80,7 +82,7 @@ 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: _chunkX3LPZGHScjs.BenchmarkCategorySchema,
   seed: _zod.z.string(),
   matches: _zod.z.number().int().positive(),
   competitors: _zod.z.array(ManifestCompetitorSchema).min(2).superRefine((models, context) => {
@@ -104,7 +106,8 @@ var RunManifestSchema = _zod.z.object({
       id: _zod.z.string().min(1),
       version: _zod.z.string().min(1),
       publicHash: _zod.z.string().min(1),
-      privateHash: _zod.z.string().min(1)
+      // Null for public-only packs (e.g. the speed arena); judged packs bind a real hash.
+      privateHash: _zod.z.string().min(1).nullable()
     })
   ),
   promptPolicyHashes: _zod.z.object({
@@ -138,17 +141,17 @@ 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 = _chunkQMOPRKWDcjs.resolveCompetitorRoster.call(void 0, 
   config.competitorIds
 )) {
-  const canonicalCompetitors = _chunk4TWPCPRPcjs.resolveCompetitorRoster.call(void 0, 
+  const canonicalCompetitors = _chunkQMOPRKWDcjs.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 = _chunkQMOPRKWDcjs.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: _chunkX3LPZGHScjs.METHODOLOGY_VERSION,
     engineVersion: ENGINE_VERSION,
     category: config.category,
     seed: config.seed,
@@ -162,16 +165,58 @@ 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: _chunkQMOPRKWDcjs.competitorPromptPolicyHash.call(void 0, config.category),
+      judge: _chunkQMOPRKWDcjs.judgePromptPolicyHash.call(void 0, config.category)
     }
   };
   return RunManifestSchema.parse(manifest);
 }
 
+// src/speed.ts
+function isLiveResponse(response) {
+  return response.success && _chunkX3LPZGHScjs.competitorContent.call(void 0, response).trim().length > 0;
+}
+function outputTokensPerSecond(outputTokens, ttftMs, totalMs) {
+  const generationSeconds = Math.max(1e-3, (totalMs - ttftMs) / 1e3);
+  return outputTokens / generationSeconds;
+}
+function speedMetricFor(response) {
+  const ttftMs = _nullishCoalesce(response.ttftMs, () => ( 0));
+  const totalMs = _nullishCoalesce(response.totalMs, () => ( response.latencyMs));
+  const outputTokens = response.outputTokens;
+  return {
+    ttftMs,
+    totalMs,
+    outputTokens,
+    tps: outputTokensPerSecond(outputTokens, ttftMs, totalMs)
+  };
+}
+function speedWinner(metrics, modelA, modelB) {
+  return metrics.a.totalMs <= metrics.b.totalMs ? modelA : modelB;
+}
+function decideSpeedMatch(responseA, responseB, modelA, modelB) {
+  const aLive = isLiveResponse(responseA);
+  const bLive = isLiveResponse(responseB);
+  if (aLive && bLive) {
+    const speedMetrics = {
+      a: speedMetricFor(responseA),
+      b: speedMetricFor(responseB)
+    };
+    return {
+      outcome: "speed-decided",
+      winnerModelId: speedWinner(speedMetrics, modelA, modelB),
+      speedMetrics
+    };
+  }
+  if (aLive !== bLive) {
+    return { outcome: "forfeit", winnerModelId: aLive ? modelA : modelB, speedMetrics: null };
+  }
+  return { outcome: "no-contest", winnerModelId: null, speedMetrics: null };
+}
+
 // 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, _chunkX3LPZGHScjs.METHODOLOGY_VERSION]);
 var EPSILON = 1e-9;
 function fail(message) {
   throw new Error(message);
@@ -234,6 +279,39 @@ function verifyPanel(panel, modelA, modelB) {
   }
   return expected.winnerModelId;
 }
+function assertSpeedMetric(actual, expected, side) {
+  equalNumber(actual.ttftMs, expected.ttftMs, `speedMetrics.${side}.ttftMs`);
+  equalNumber(actual.totalMs, expected.totalMs, `speedMetrics.${side}.totalMs`);
+  equalNumber(actual.outputTokens, expected.outputTokens, `speedMetrics.${side}.outputTokens`);
+  equalNumber(actual.tps, expected.tps, `speedMetrics.${side}.tps`);
+}
+function expectedSpeedOutcome(match, modelA, modelB) {
+  const { responseA, responseB } = match.competitors;
+  if (isLiveResponse(responseA) && isLiveResponse(responseB)) {
+    if (!match.speedMetrics) {
+      fail("a speed-decided match must record speedMetrics for both competitors");
+    }
+    const derived = {
+      a: speedMetricFor(responseA),
+      b: speedMetricFor(responseB)
+    };
+    assertSpeedMetric(match.speedMetrics.a, derived.a, "a");
+    assertSpeedMetric(match.speedMetrics.b, derived.b, "b");
+    return {
+      expectedOutcome: "speed-decided",
+      expectedWinner: speedWinner(derived, modelA, modelB)
+    };
+  }
+  if (match.speedMetrics != null) {
+    fail("a speed forfeit or no-contest must not record speedMetrics");
+  }
+  const aLive = isLiveResponse(responseA);
+  const bLive = isLiveResponse(responseB);
+  if (aLive !== bLive) {
+    return { expectedOutcome: "forfeit", expectedWinner: aLive ? modelA : modelB };
+  }
+  return { expectedOutcome: "no-contest", expectedWinner: null };
+}
 function verifyOutcome(match) {
   const { modelA, modelB, responseA, responseB } = match.competitors;
   if (modelA === modelB) fail("a competitor cannot face itself");
@@ -242,7 +320,10 @@ function verifyOutcome(match) {
   }
   let expectedWinner = null;
   let expectedOutcome;
-  if (responseA.success !== responseB.success) {
+  if (match.task.category === "speed") {
+    if (match.panel !== null) fail("a speed match must not contain a judge panel");
+    ({ expectedOutcome, expectedWinner } = expectedSpeedOutcome(match, modelA, modelB));
+  } else if (responseA.success !== responseB.success) {
     expectedOutcome = "forfeit";
     expectedWinner = responseA.success ? modelA : modelB;
     if (match.panel !== null) fail("forfeit must not contain a judge panel");
@@ -266,7 +347,7 @@ 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,
+    _chunkX3LPZGHScjs.competitorCost.call(void 0, responseA) + _chunkX3LPZGHScjs.competitorCost.call(void 0, responseB) + judgeCost,
     "matchCostUsd"
   );
 }
@@ -396,6 +477,33 @@ function verifiedEloState(matches, category) {
 }
 
 // src/report.ts
+function median(values) {
+  if (values.length === 0) return 0;
+  const sorted = [...values].sort((left, right) => left - right);
+  const mid = Math.floor(sorted.length / 2);
+  return sorted.length % 2 === 0 ? (sorted[mid - 1] + sorted[mid]) / 2 : sorted[mid];
+}
+function average(values) {
+  if (values.length === 0) return 0;
+  return values.reduce((sum, value) => sum + value, 0) / values.length;
+}
+function collectSpeedSample(samples, modelId, metric) {
+  const existing = _nullishCoalesce(samples.get(modelId), () => ( { ttftMs: [], tps: [] }));
+  existing.ttftMs.push(metric.ttftMs);
+  existing.tps.push(metric.tps);
+  samples.set(modelId, existing);
+}
+function speedStats(sample) {
+  const ttftMs = _nullishCoalesce(_optionalChain([sample, 'optionalAccess', _16 => _16.ttftMs]), () => ( []));
+  const tps = _nullishCoalesce(_optionalChain([sample, 'optionalAccess', _17 => _17.tps]), () => ( []));
+  return {
+    samples: ttftMs.length,
+    medianTtftMs: round(median(ttftMs), 1),
+    avgTtftMs: round(average(ttftMs), 1),
+    medianTps: round(median(tps), 2),
+    avgTps: round(average(tps), 2)
+  };
+}
 function round(value, digits = 2) {
   return Number(value.toFixed(digits));
 }
@@ -420,7 +528,7 @@ function ensureEntry(entries, modelId) {
   if (existing) return existing;
   let displayName = modelId;
   try {
-    displayName = _chunk4TWPCPRPcjs.getModel.call(void 0, modelId).displayName;
+    displayName = _chunkQMOPRKWDcjs.getModel.call(void 0, modelId).displayName;
   } catch (e) {
   }
   const created = emptyEntry(modelId, displayName);
@@ -430,7 +538,8 @@ 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 speedSamples = /* @__PURE__ */ new Map();
+  const competitorIds = _nullishCoalesce(options.competitorIds, () => ( _chunkQMOPRKWDcjs.listModels.call(void 0, "competitor").map((model) => model.id)));
   for (const modelId of competitorIds) {
     ensureEntry(entries, modelId);
   }
@@ -439,8 +548,12 @@ 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 += _chunkX3LPZGHScjs.competitorCost.call(void 0, match.competitors.responseA);
+    modelB.totalCostUsd += _chunkX3LPZGHScjs.competitorCost.call(void 0, match.competitors.responseB);
+    if (category === "speed" && match.speedMetrics) {
+      collectSpeedSample(speedSamples, modelA.modelId, match.speedMetrics.a);
+      collectSpeedSample(speedSamples, modelB.modelId, match.speedMetrics.b);
+    }
     if (match.outcome === "no-contest" || !match.winnerModelId) continue;
     modelA.matches += 1;
     modelB.matches += 1;
@@ -450,7 +563,7 @@ function buildSnapshot(matches, category, options = {}) {
     winner.wins += 1;
     loser.losses += 1;
     if (match.outcome === "forfeit") loser.forfeits += 1;
-    if (_optionalChain([match, 'access', _16 => _16.panel, 'optionalAccess', _17 => _17.agreement]) === "unanimous") winner.unanimousWins += 1;
+    if (_optionalChain([match, 'access', _18 => _18.panel, 'optionalAccess', _19 => _19.agreement]) === "unanimous") winner.unanimousWins += 1;
     const cluster = match.task.cluster;
     winner.byCluster[cluster] = incrementCluster(winner.byCluster[cluster], "wins");
     loser.byCluster[cluster] = incrementCluster(loser.byCluster[cluster], "losses");
@@ -459,13 +572,16 @@ function buildSnapshot(matches, category, options = {}) {
     ...entry,
     elo: round(_nullishCoalesce(verified.ratings[entry.modelId], () => ( ELO_INITIAL))),
     winRate: entry.matches === 0 ? 0 : round(entry.wins / entry.matches * 100, 1),
-    totalCostUsd: round(entry.totalCostUsd, 6)
+    totalCostUsd: round(entry.totalCostUsd, 6),
+    // Speed aggregates are surfaced only for the speed category; judged
+    // categories keep their existing entry shape unchanged.
+    ...category === "speed" ? { speed: speedStats(speedSamples.get(entry.modelId)) } : {}
   })).sort(
     (a, b) => b.elo - a.elo || b.points - a.points || a.displayName.localeCompare(b.displayName)
   ).map((entry, index) => ({ rank: index + 1, ...entry }));
   return {
     version: "0.2.0",
-    methodologyVersion: _nullishCoalesce(verified.methodologyVersion, () => ( _chunkCIXITJW6cjs.METHODOLOGY_VERSION)),
+    methodologyVersion: _nullishCoalesce(verified.methodologyVersion, () => ( _chunkX3LPZGHScjs.METHODOLOGY_VERSION)),
     category,
     generatedAt: (/* @__PURE__ */ new Date()).toISOString(),
     initialElo: ELO_INITIAL,
@@ -479,13 +595,21 @@ function incrementCluster(current, field) {
   return { ...next, [field]: next[field] + 1 };
 }
 function renderMarkdown(snapshot) {
-  const meta = _chunkCIXITJW6cjs.CATEGORY_META[snapshot.category];
+  const meta = _chunkX3LPZGHScjs.CATEGORY_META[snapshot.category];
   const modelNames = new Map(
     snapshot.leaderboard.map((entry) => [entry.modelId, entry.displayName])
   );
-  const rows = snapshot.leaderboard.map(
-    (entry) => `| ${entry.rank} | ${entry.displayName} | ${entry.elo.toFixed(2)} | ${entry.points} | ${entry.wins}-${entry.losses} | ${entry.forfeits} | ${entry.winRate.toFixed(1)}% | $${entry.totalCostUsd.toFixed(4)} |`
-  );
+  const isSpeed = snapshot.category === "speed";
+  const speedHeader = isSpeed ? " Median TTFT (ms) | Median TPS |" : "";
+  const speedAlign = isSpeed ? "---:|---:|" : "";
+  const rows = snapshot.leaderboard.map((entry) => {
+    const base = `| ${entry.rank} | ${entry.displayName} | ${entry.elo.toFixed(2)} | ${entry.points} | ${entry.wins}-${entry.losses} | ${entry.forfeits} | ${entry.winRate.toFixed(1)}% | $${entry.totalCostUsd.toFixed(4)} |`;
+    if (!isSpeed) return base;
+    const speed = entry.speed;
+    const ttft = speed && speed.samples > 0 ? speed.medianTtftMs.toFixed(0) : "\u2014";
+    const tps = speed && speed.samples > 0 ? speed.medianTps.toFixed(1) : "\u2014";
+    return `${base} ${ttft} | ${tps} |`;
+  });
   const recent = snapshot.matches.slice(-20).reverse().map((match) => {
     const a = _nullishCoalesce(modelNames.get(match.competitors.modelA), () => ( match.competitors.modelA));
     const b = _nullishCoalesce(modelNames.get(match.competitors.modelB), () => ( match.competitors.modelB));
@@ -500,8 +624,8 @@ Generated ${snapshot.generatedAt}. Ratings start at ${snapshot.initialElo} with
 
 ## Leaderboard
 
-| Rank | Model | Elo | Points | W-L | Forfeits | Win rate | Competitor cost |
-|---:|---|---:|---:|---:|---:|---:|---:|
+| Rank | Model | Elo | Points | W-L | Forfeits | Win rate | Competitor cost |${speedHeader}
+|---:|---|---:|---:|---:|---:|---:|---:|${speedAlign}
 ${rows.join("\n")}
 
 ## Recent matches
@@ -567,7 +691,7 @@ 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,
+    _chunkX3LPZGHScjs.METHODOLOGY_VERSION,
     input.category,
     input.seed,
     String(input.count),
@@ -674,7 +798,7 @@ function triageJournal(results) {
       seed: runResults[0].seed,
       firstTimestamp: runResults[0].timestamp,
       matches: runResults.length,
-      outcomes: { judged: 0, forfeit: 0, "no-contest": 0 },
+      outcomes: { judged: 0, forfeit: 0, "no-contest": 0, "speed-decided": 0 },
       totalCostUsd: 0,
       errorClasses: {},
       models: {},
@@ -772,7 +896,7 @@ function formatTriage(reports) {
   for (const report of reports) {
     lines.push(`run ${report.runId} (seed "${report.seed}", ${report.firstTimestamp})`);
     lines.push(
-      `  matches ${report.matches} | judged ${report.outcomes.judged}, forfeit ${report.outcomes.forfeit}, no-contest ${report.outcomes["no-contest"]} | spend $${report.totalCostUsd.toFixed(4)}`
+      `  matches ${report.matches} | judged ${report.outcomes.judged}, speed-decided ${report.outcomes["speed-decided"]}, forfeit ${report.outcomes.forfeit}, no-contest ${report.outcomes["no-contest"]} | spend $${report.totalCostUsd.toFixed(4)}`
     );
     if (Object.keys(report.errorClasses).length > 0) {
       lines.push(
@@ -821,17 +945,17 @@ function failedResponse(modelId, error, latencyMs) {
   return {
     modelId,
     success: false,
-    error: _chunk4TWPCPRPcjs.sanitizeError.call(void 0, error),
+    error: _chunkQMOPRKWDcjs.sanitizeError.call(void 0, error),
     latencyMs
   };
 }
 var ArenaRunner = class {
-  constructor(gateway, store, onEvent, logger = _chunk4TWPCPRPcjs.noopLogger) {
+  constructor(gateway, store, onEvent, logger = _chunkQMOPRKWDcjs.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, _chunkQMOPRKWDcjs.JudgePanel)(gateway, onEvent, logger);
   }
   
   
@@ -839,7 +963,7 @@ var ArenaRunner = class {
   
   
   emit(event) {
-    _optionalChain([this, 'access', _18 => _18.onEvent, 'optionalCall', _19 => _19({ ...event, timestamp: (/* @__PURE__ */ new Date()).toISOString() })]);
+    _optionalChain([this, 'access', _20 => _20.onEvent, 'optionalCall', _21 => _21({ ...event, timestamp: (/* @__PURE__ */ new Date()).toISOString() })]);
   }
   observeCancellation(runId, progress, signal) {
     let emitted = false;
@@ -861,11 +985,11 @@ var ArenaRunner = class {
         }
       });
     };
-    if (_optionalChain([signal, 'optionalAccess', _20 => _20.aborted])) requested();
-    else _optionalChain([signal, 'optionalAccess', _21 => _21.addEventListener, 'call', _22 => _22("abort", requested, { once: true })]);
+    if (_optionalChain([signal, 'optionalAccess', _22 => _22.aborted])) requested();
+    else _optionalChain([signal, 'optionalAccess', _23 => _23.addEventListener, 'call', _24 => _24("abort", requested, { once: true })]);
     return {
       requested,
-      dispose: () => _optionalChain([signal, 'optionalAccess', _23 => _23.removeEventListener, 'call', _24 => _24("abort", requested)])
+      dispose: () => _optionalChain([signal, 'optionalAccess', _25 => _25.removeEventListener, 'call', _26 => _26("abort", requested)])
     };
   }
   cancelRun(runId, progress) {
@@ -892,7 +1016,7 @@ var ArenaRunner = class {
     };
   }
   async prepareRun(config, tasks, signal) {
-    const competitors = _chunk4TWPCPRPcjs.resolveCompetitorRoster.call(void 0, config.competitorIds);
+    const competitors = _chunkQMOPRKWDcjs.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 +1025,10 @@ 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 !== _chunkX3LPZGHScjs.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 ${_chunkX3LPZGHScjs.METHODOLOGY_VERSION} matches to a ${incompatible.methodologyVersion} journal; archive it and start a new journal`
       );
     }
     this.store.rebuildEloState(competitorIds);
@@ -934,18 +1058,19 @@ var ArenaRunner = class {
         stoppedForBudget: false
       }
     };
-    if (_optionalChain([signal, 'optionalAccess', _25 => _25.aborted])) return prepared;
+    if (_optionalChain([signal, 'optionalAccess', _27 => _27.aborted])) return prepared;
+    const judgesToValidate = config.category === "speed" ? [] : _chunkQMOPRKWDcjs.listModels.call(void 0, "judge");
     try {
       await Promise.all(
-        [...competitors, ..._chunk4TWPCPRPcjs.listModels.call(void 0, "judge")].map(
+        [...competitors, ...judgesToValidate].map(
           (model) => this.gateway.validateModel(model, signal)
         )
       );
     } catch (error) {
-      if (_optionalChain([signal, 'optionalAccess', _26 => _26.aborted])) return prepared;
+      if (_optionalChain([signal, 'optionalAccess', _28 => _28.aborted])) return prepared;
       throw error;
     }
-    if (_optionalChain([signal, 'optionalAccess', _27 => _27.aborted])) return prepared;
+    if (_optionalChain([signal, 'optionalAccess', _29 => _29.aborted])) return prepared;
     this.store.writeRunManifest(runId, manifest);
     return prepared;
   }
@@ -955,7 +1080,7 @@ var ArenaRunner = class {
     const competitorIds = competitors.map((model) => model.id);
     const cancellation = this.observeCancellation(runId, progress, signal);
     try {
-      if (_optionalChain([signal, 'optionalAccess', _28 => _28.aborted])) return this.cancelRun(runId, progress);
+      if (_optionalChain([signal, 'optionalAccess', _30 => _30.aborted])) return this.cancelRun(runId, progress);
       this.logger.info("run.started", {
         runId,
         category: config.category,
@@ -979,7 +1104,7 @@ var ArenaRunner = class {
         }
       });
       for (const match of schedule) {
-        if (_optionalChain([signal, 'optionalAccess', _29 => _29.aborted])) return this.cancelRun(runId, progress);
+        if (_optionalChain([signal, 'optionalAccess', _31 => _31.aborted])) return this.cancelRun(runId, progress);
         if (completedIds.has(match.id)) {
           if (config.resume) continue;
           throw new Error(
@@ -1020,13 +1145,13 @@ 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', _32 => _32.aborted]) || _chunkQMOPRKWDcjs.isArenaCancellationError.call(void 0, error)) {
             cancellation.requested();
             return this.cancelRun(runId, progress);
           }
           throw error;
         }
-        if (_optionalChain([signal, 'optionalAccess', _31 => _31.aborted])) return this.cancelRun(runId, progress);
+        if (_optionalChain([signal, 'optionalAccess', _33 => _33.aborted])) return this.cancelRun(runId, progress);
         this.store.append(result);
         writeReports(this.store, { competitorIds });
         progress.costUsd += result.matchCostUsd;
@@ -1049,7 +1174,7 @@ var ArenaRunner = class {
           }
         });
         if (onMatchResult) await onMatchResult(result);
-        if (_optionalChain([signal, 'optionalAccess', _32 => _32.aborted])) return this.cancelRun(runId, progress);
+        if (_optionalChain([signal, 'optionalAccess', _34 => _34.aborted])) return this.cancelRun(runId, progress);
         if (config.healthStop !== false && progress.completed >= HEALTH_STOP_MIN_MATCHES && progress.matchesWithFailures / progress.completed >= HEALTH_STOP_FAILURE_RATE) {
           this.logger.error("run.health-stopped", {
             runId,
@@ -1062,7 +1187,7 @@ var ArenaRunner = class {
           );
         }
       }
-      if (_optionalChain([signal, 'optionalAccess', _33 => _33.aborted])) return this.cancelRun(runId, progress);
+      if (_optionalChain([signal, 'optionalAccess', _35 => _35.aborted])) return this.cancelRun(runId, progress);
       writeReports(this.store, { competitorIds });
       this.logger.info("run.completed", {
         runId,
@@ -1094,14 +1219,14 @@ var ArenaRunner = class {
     }
   }
   async runMatch(match, task, manifestHash, competitorIds, signal) {
-    _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+    _chunkQMOPRKWDcjs.throwIfCancelled.call(void 0, signal);
     const state = this.store.rebuildEloState(competitorIds);
-    const prompt = _chunk4TWPCPRPcjs.buildCompetitorPrompt.call(void 0, task);
+    const prompt = _chunkQMOPRKWDcjs.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);
+    _chunkQMOPRKWDcjs.throwIfCancelled.call(void 0, signal);
     this.inspectResponse(match, responseA);
     this.inspectResponse(match, responseB);
     this.emit({
@@ -1112,20 +1237,20 @@ 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: _chunkX3LPZGHScjs.competitorCost.call(void 0, responseA),
+          outputTokens: _chunkX3LPZGHScjs.competitorOutputTokens.call(void 0, responseA),
+          reasoningTokens: _nullishCoalesce(_chunkX3LPZGHScjs.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: _chunkX3LPZGHScjs.competitorCost.call(void 0, responseB),
+          outputTokens: _chunkX3LPZGHScjs.competitorOutputTokens.call(void 0, responseB),
+          reasoningTokens: _nullishCoalesce(_chunkX3LPZGHScjs.competitorReasoningTokens.call(void 0, responseB), () => ( null))
         }
       }
     });
-    _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+    _chunkQMOPRKWDcjs.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))
@@ -1134,17 +1259,28 @@ var ArenaRunner = class {
     let outcome = "no-contest";
     let winnerModelId = null;
     let panel = null;
-    if (responseA.success !== responseB.success) {
+    let speedMetrics = null;
+    if (task.public.category === "speed") {
+      const decision = decideSpeedMatch(responseA, responseB, match.modelA, match.modelB);
+      outcome = decision.outcome;
+      winnerModelId = decision.winnerModelId;
+      speedMetrics = decision.speedMetrics;
+    } else if (responseA.success !== responseB.success) {
       outcome = "forfeit";
       winnerModelId = responseA.success ? match.modelA : match.modelB;
     } else if (responseA.success && responseB.success) {
+      if (!_chunkX3LPZGHScjs.isCompleteArenaTask.call(void 0, task)) {
+        throw new Error(
+          `A judged ${task.public.category} match requires a task with its hidden reference`
+        );
+      }
       this.emit({
         id: `${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: _chunkQMOPRKWDcjs.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);
+      _chunkQMOPRKWDcjs.throwIfCancelled.call(void 0, signal);
       if (panel.winnerModelId) {
         outcome = "judged";
         winnerModelId = panel.winnerModelId;
@@ -1159,9 +1295,9 @@ var ArenaRunner = class {
       eloAfter[match.modelA] = update.ratingA;
       eloAfter[match.modelB] = update.ratingB;
     }
-    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));
+    const judgeCost = _nullishCoalesce(_optionalChain([panel, 'optionalAccess', _36 => _36.votes, 'access', _37 => _37.reduce, 'call', _38 => _38((sum, vote) => sum + (_nullishCoalesce(_optionalChain([vote, 'access', _39 => _39.completion, 'optionalAccess', _40 => _40.costUsd]), () => ( 0))), 0)]), () => ( 0));
     return {
-      methodologyVersion: _chunkCIXITJW6cjs.METHODOLOGY_VERSION,
+      methodologyVersion: _chunkX3LPZGHScjs.METHODOLOGY_VERSION,
       runId: match.runId,
       runManifestHash: manifestHash,
       matchId: match.id,
@@ -1180,10 +1316,11 @@ var ArenaRunner = class {
       outcome,
       winnerModelId,
       panel,
+      speedMetrics,
       eloBefore,
       eloAfter,
       pointAwarded: winnerModelId !== null,
-      matchCostUsd: _chunkCIXITJW6cjs.competitorCost.call(void 0, responseA) + _chunkCIXITJW6cjs.competitorCost.call(void 0, responseB) + judgeCost
+      matchCostUsd: _chunkX3LPZGHScjs.competitorCost.call(void 0, responseA) + _chunkX3LPZGHScjs.competitorCost.call(void 0, responseB) + judgeCost
     };
   }
   /** Surface signals that make a journaled response untrustworthy or a task too easy. */
@@ -1226,19 +1363,19 @@ var ArenaRunner = class {
       });
     };
     try {
-      _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+      _chunkQMOPRKWDcjs.throwIfCancelled.call(void 0, signal);
       const completion = await this.gateway.complete({
-        model: _chunk4TWPCPRPcjs.getModel.call(void 0, modelId),
+        model: _chunkQMOPRKWDcjs.getModel.call(void 0, modelId),
         ...prompt,
         signal,
         onDelta: (text) => emitDelta(text, false)
       });
-      _chunk4TWPCPRPcjs.throwIfCancelled.call(void 0, signal);
+      _chunkQMOPRKWDcjs.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', _41 => _41.aborted]) || _chunkQMOPRKWDcjs.isArenaCancellationError.call(void 0, error)) {
+        throw new (0, _chunkQMOPRKWDcjs.ArenaCancellationError)();
       }
       const failed = failedResponse(modelId, error, Date.now() - startedAt);
       emitDelta(failed.error, true, false);
@@ -1271,7 +1408,7 @@ var MockOpenRouterGateway = class {
   async validateModel(_model) {
   }
   async complete(request) {
-    if (_optionalChain([request, 'access', _40 => _40.signal, 'optionalAccess', _41 => _41.aborted])) {
+    if (_optionalChain([request, 'access', _42 => _42.signal, 'optionalAccess', _43 => _43.aborted])) {
       throw new Error("Mock completion aborted");
     }
     if (request.structured) {
@@ -1288,7 +1425,7 @@ var MockOpenRouterGateway = class {
         },
         violations: []
       };
-      void _chunk4TWPCPRPcjs.judgeVerdictJsonSchema.call(void 0, );
+      void _chunkQMOPRKWDcjs.judgeVerdictJsonSchema.call(void 0, );
       return baseCompletion(JSON.stringify(payload), `mock-judge-${request.model.id}`);
     }
     const variantA = [
@@ -1307,11 +1444,11 @@ var MockOpenRouterGateway = class {
     const parts = _nullishCoalesce(text.match(/.{1,24}/g), () => ( [text]));
     let cumulative = "";
     for (const part of parts) {
-      if (_optionalChain([request, 'access', _42 => _42.signal, 'optionalAccess', _43 => _43.aborted])) {
+      if (_optionalChain([request, 'access', _44 => _44.signal, 'optionalAccess', _45 => _45.aborted])) {
         throw new Error("Mock completion aborted");
       }
       cumulative += part;
-      _optionalChain([request, 'access', _44 => _44.onDelta, 'optionalCall', _45 => _45(cumulative)]);
+      _optionalChain([request, 'access', _46 => _46.onDelta, 'optionalCall', _47 => _47(cumulative)]);
       if (delayMs > 0) await sleep(delayMs);
     }
     return baseCompletion(cumulative, `mock-competitor-${request.model.id}`);
@@ -1371,12 +1508,12 @@ 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 = _chunkX3LPZGHScjs.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(".")}` : "";
+        const location = _optionalChain([issue, 'optionalAccess', _48 => _48.path, 'access', _49 => _49.length]) ? ` at ${issue.path.join(".")}` : "";
         throw new Error(
-          `Invalid arena journal line ${index + 1}${location}: ${_nullishCoalesce(_optionalChain([issue, 'optionalAccess', _48 => _48.message]), () => ( "schema mismatch"))}`
+          `Invalid arena journal line ${index + 1}${location}: ${_nullishCoalesce(_optionalChain([issue, 'optionalAccess', _50 => _50.message]), () => ( "schema mismatch"))}`
         );
       }
       return parsed.data;
@@ -1409,7 +1546,7 @@ 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 = _chunkQMOPRKWDcjs.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 +1601,3 @@ 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