Docs: Update benchmark results

2026-01-14 00:27:55 +00:00 · 2025-11-18 23:31:52 +00:00
parent 341252fec1
commit 5855cf8a6f
77 changed files with 972 additions and 1051 deletions
--- a/results.json
+++ b/results.json
@@ -13,69 +13,69 @@
    "9_stream_visualizer": 51.923362084000026
  },
  "anthropic/claude-sonnet-4.5 TEMP:0.7": {
-    "10_scrypt_hash": 2.994402775000024,
-    "11_geospatial": 3.173007857000048,
-    "1_dijkstra": 4.721320241999987,
-    "2_convex_hull": 4.782739678999991,
-    "3_lis": 2.566319416000042,
-    "4_determinant": 2.087922611000016,
-    "5_markdown_parser": 2.1563547050000054,
-    "6_csv_processor": 4.368516923999996,
-    "7_scheduler": 10.690530350000016,
-    "8_json_validator": 2.9041606949999696,
-    "9_stream_visualizer": 4.411488517000049
+    "10_scrypt_hash": 2.512005516999983,
+    "11_geospatial": 3.3480960069999566,
+    "1_dijkstra": 4.758364550999948,
+    "2_convex_hull": 5.67630803700001,
+    "3_lis": 2.4023967360000245,
+    "4_determinant": 2.215544780999946,
+    "5_markdown_parser": 1.9106662779999897,
+    "6_csv_processor": 4.774563675000041,
+    "7_scheduler": 10.645383548000012,
+    "8_json_validator": 3.43886771999998,
+    "9_stream_visualizer": 4.354729013999982
  },
  "openai/gpt-5.1-codex": {
-    "10_scrypt_hash": 19.945625629000016,
-    "11_geospatial": 7.752952860000019,
-    "1_dijkstra": 18.825615109000005,
-    "2_convex_hull": 24.519327066999978,
-    "3_lis": 13.449256432000023,
-    "4_determinant": 3.5313777560000306,
-    "5_markdown_parser": 4.871686004000018,
-    "6_csv_processor": 31.731031262999984,
-    "7_scheduler": 91.32724592999998,
-    "8_json_validator": 11.784968072999968,
-    "9_stream_visualizer": 17.32446443499997
+    "10_scrypt_hash": 7.498817803999991,
+    "11_geospatial": 10.405374638999987,
+    "1_dijkstra": 19.193891047999962,
+    "2_convex_hull": 14.263961956999964,
+    "3_lis": 3.3163609489999946,
+    "4_determinant": 5.17318697499996,
+    "5_markdown_parser": 6.449940595000051,
+    "6_csv_processor": 37.769467569000085,
+    "7_scheduler": 79.01968128499995,
+    "8_json_validator": 8.087602373000001,
+    "9_stream_visualizer": 12.885421481000026
  },
  "moonshotai/kimi-k2-thinking": {
-    "10_scrypt_hash": 54.48318050600006,
-    "11_geospatial": 123.94316572699998,
-    "1_dijkstra": 20.75571812199999,
-    "2_convex_hull": 281.26114928399994,
-    "3_lis": 78.25152572599984,
-    "4_determinant": 67.81623509099987,
-    "5_markdown_parser": 154.2897356459999,
-    "6_csv_processor": 85.1170599280002,
-    "7_scheduler": 389.4384002489999,
-    "8_json_validator": 17.501368902000134,
-    "9_stream_visualizer": 98.03424570999994
+    "10_scrypt_hash": 53.01161044100009,
+    "11_geospatial": 49.509398480999984,
+    "1_dijkstra": 75.16873453799997,
+    "2_convex_hull": 115.48629523699998,
+    "3_lis": 43.45383791499992,
+    "4_determinant": 99.70906514900003,
+    "5_markdown_parser": 22.150892672999994,
+    "6_csv_processor": 100.07209453999997,
+    "7_scheduler": 204.68869797400012,
+    "8_json_validator": 282.8534934129999,
+    "9_stream_visualizer": 445.8482685430001
  },
  "google/gemini-2.5-pro": {
-    "10_scrypt_hash": 33.5175050580001,
-    "11_geospatial": 31.009794073000318,
-    "1_dijkstra": 45.908529550999866,
-    "2_convex_hull": 85.32745476700039,
-    "3_lis": 32.75355649500014,
-    "4_determinant": 27.239392394999975,
-    "5_markdown_parser": 27.228569728999865,
-    "6_csv_processor": 69.48256337699992,
-    "7_scheduler": 54.015430929000026,
-    "8_json_validator": 29.125379850999916,
-    "9_stream_visualizer": 36.110917992000005
+    "10_scrypt_hash": 33.595784767000005,
+    "11_geospatial": 23.084225926999935,
+    "1_dijkstra": 39.2938456369997,
+    "2_convex_hull": 39.61738663299987,
+    "3_lis": 32.74601506800018,
+    "4_determinant": 12.779542751000262,
+    "5_markdown_parser": 17.71407721999986,
+    "6_csv_processor": 28.51295679199975,
+    "7_scheduler": 98.78428086399985,
+    "8_json_validator": 17.644947614999953,
+    "9_stream_visualizer": 42.817067843999716
  },
  "openrouter/sherlock-think-alpha": {
-    "10_scrypt_hash": 9.486372194000054,
-    "11_geospatial": 15.555460506999864,
-    "1_dijkstra": 21.514197755000087,
-    "2_convex_hull": 35.37556943300041,
-    "3_lis": 17.00031832400011,
-    "4_determinant": 8.939955472999719,
-    "5_markdown_parser": 15.81524256400019,
-    "6_csv_processor": 23.44287434099987,
-    "7_scheduler": 68.04158863600017,
-    "8_json_validator": 19.484740249000023,
-    "9_stream_visualizer": 28.426358369000255
+    "10_scrypt_hash": 14.262482631000225,
+    "11_geospatial": 21.945668632999993,
+    "1_dijkstra": 31.84612786399992,
+    "2_convex_hull": 51.9222613090002,
+    "3_lis": 27.120956264999695,
+    "4_determinant": 8.37700194299966,
+    "5_markdown_parser": 24.515614581000058,
+    "6_csv_processor": 25.41584125499986,
+    "7_scheduler": 93.37877358199982,
+    "8_json_validator": 43.54576423399988,
+    "9_stream_visualizer": 41.675588322999886
  },
  "gemini-3-pro-preview TEMP:1.1": {
    "10_scrypt_hash": 20.09197308799997,
@@ -349,5 +349,18 @@
    "7_scheduler": 43.091859620000236,
    "8_json_validator": 24.684869859999978,
    "9_stream_visualizer": 19.26046375499992
+  },
+  "google/gemini-3-pro-preview TEMP:0.35": {
+    "10_scrypt_hash": 17.112088825000004,
+    "11_geospatial": 34.95774618,
+    "1_dijkstra": 17.93846221700001,
+    "2_convex_hull": 51.98661096200001,
+    "3_lis": 33.88681510400001,
+    "4_determinant": 17.43103938499998,
+    "5_markdown_parser": 79.04342028300002,
+    "6_csv_processor": 74.62322437199997,
+    "7_scheduler": 79.87219575399999,
+    "8_json_validator": 28.535885181999998,
+    "9_stream_visualizer": 21.60573508199997
  }
 }
--- a/tests/10_scrypt_hash/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/10_scrypt_hash/outputs/anthropic_claude-sonnet-4.5
@@ -1,9 +1,9 @@
 async function hashPassword(password, salt) {
  const { scrypt } = await import('https://cdn.jsdelivr.net/npm/scrypt-js@3.0.1/+esm');
-  const enc = new TextEncoder();
+  const encoder = new TextEncoder();
  const hash = await scrypt(
-    enc.encode(password),
-    enc.encode(salt),
+    encoder.encode(password),
+    encoder.encode(salt),
    1024,
    8,
    1,
--- a/tests/10_scrypt_hash/outputs/google_gemini-2.5-pro.js
+++ b/tests/10_scrypt_hash/outputs/google_gemini-2.5-pro.js
@@ -1,5 +1,5 @@
 const hashPassword = async (password, salt) => {
-  const { default: scrypt } = await import('https://cdn.jsdelivr.net/npm/scrypt-js@3.0.1/scrypt.min.js');
+  const { scrypt } = await import('https://esm.sh/scrypt-js@3.0.1');

  const encoder = new TextEncoder();

@@ -13,7 +13,7 @@ const hashPassword = async (password, salt) => {
  );

  return [...hashBytes]
-    .map(byte => byte.toString(16).padStart(2, '0'))
+    .map(b => b.toString(16).padStart(2, '0'))
    .join('');
 };
 export default hashPassword;
--- a/tests/10_scrypt_hash/outputs/google_gemini-3-pro-preview
+++ b/tests/10_scrypt_hash/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,7 @@
+const hashPassword = async (p, s) => {
+  const { scrypt } = await import('https://esm.sh/scrypt-js')
+  const e = new TextEncoder()
+  const h = await scrypt(e.encode(p), e.encode(s), 1024, 8, 1, 32)
+  return [...h].map(b => b.toString(16).padStart(2, '0')).join('')
+}
+export default hashPassword;
--- a/tests/10_scrypt_hash/outputs/google_gemini-3-pro-preview.js
+++ b/tests/10_scrypt_hash/outputs/google_gemini-3-pro-preview.js
@@ -1,7 +0,0 @@
-const hashPassword = async (pwd, salt) => {
-  const { scrypt } = await import('https://unpkg.com/scrypt-js@3.0.1/scrypt.js')
-  const e = new TextEncoder()
-  const h = await scrypt(e.encode(pwd), e.encode(salt), 1024, 8, 1, 32)
-  return Array.from(h, b => b.toString(16).padStart(2, '0')).join('')
-}
-export default hashPassword;
--- a/tests/10_scrypt_hash/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/10_scrypt_hash/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,7 +1,9 @@
 async function hashPassword(password, salt) {
-  const { scrypt } = await import('https://cdn.jsdelivr.net/npm/scrypt-js@3.0.1/+esm');
+  const { default: scrypt } = await import('https://cdn.jsdelivr.net/npm/scrypt-js@3.0.1/index.js');
  const e = new TextEncoder();
-  const hash = await scrypt(e.encode(password), e.encode(salt), 1024, 8, 1, 32);
-  return Array.from(hash, b => b.toString(16).padStart(2, '0')).join('');
+  const p = e.encode(password);
+  const s = e.encode(salt);
+  const h = await scrypt(p, s, 1024, 8, 1, 32);
+  return [...h].map(b => b.toString(16).padStart(2, '0')).join('');
 }
 export default hashPassword;
--- a/tests/10_scrypt_hash/outputs/openai_gpt-5.1-codex.js
+++ b/tests/10_scrypt_hash/outputs/openai_gpt-5.1-codex.js
@@ -1,9 +1,7 @@
-const enc = new TextEncoder;
-let scryptLoader;
-const hashPassword = async (password, salt) => {
-  scryptLoader ||= import('https://cdn.jsdelivr.net/npm/scrypt-js@3.0.1/+esm').then(({ scrypt }) => scrypt);
-  const scrypt = await scryptLoader;
-  const hash = await scrypt(enc.encode(password), enc.encode(salt), 1024, 8, 1, 32);
-  return [...hash].map(v => v.toString(16).padStart(2, '0')).join('');
-};
+const e=new TextEncoder
+const hashPassword=async(p,s)=>{
+  const{ scrypt }=await import('https://cdn.skypack.dev/scrypt-js')
+  const h=await scrypt(e.encode(p),e.encode(s),1024,8,1,32)
+  return Array.from(h,x=>(x+256).toString(16).slice(-2)).join('')
+}
 export default hashPassword;
--- a/tests/10_scrypt_hash/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/10_scrypt_hash/outputs/openrouter_sherlock-think-alpha.js
@@ -1,2 +1,9 @@
-async function hashPassword(password,salt){const{scrypt}=await import('https://cdn.skypack.dev/scrypt-js');const e=new TextEncoder();const h=await scrypt(e.encode(password),e.encode(salt),1024,8,1,32);return Array.from(h).map(b=>b.toString(16).padStart(2,'0')).join('')}
+async function hashPassword(password, salt) {
+  const { scrypt } = await import('https://cdn.skypack.dev/scrypt-js');
+  const e = new TextEncoder();
+  const pw = e.encode(password);
+  const s = e.encode(salt);
+  const h = await scrypt(pw, s, 1024, 8, 1, 32);
+  return Array.from(h).map(b=>b.toString(16).padStart(2,'0')).join('');
+}
 export default hashPassword;
--- a/tests/11_geospatial/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/11_geospatial/outputs/anthropic_claude-sonnet-4.5
@@ -1,14 +1,14 @@
-const analyzeGPS = async (readings, boundary) => {
+async function analyzeGPS(readings, boundary) {
  const turf = await import('https://esm.sh/@turf/turf@7.1.0');
  
-  const validReadings = readings
-    .filter(r => turf.booleanPointInPolygon(turf.point(r.loc), boundary))
-    .sort((a, b) => a.ts - b.ts);
+  const valid = readings.filter(r => 
+    turf.booleanPointInPolygon(r.loc, boundary)
+  ).sort((a, b) => a.ts - b.ts);
  
-  if (validReadings.length < 2) return 0;
+  if (valid.length < 2) return 0;
  
-  const lineString = turf.lineString(validReadings.map(r => r.loc));
+  const line = turf.lineString(valid.map(r => r.loc));
  
-  return Math.round(turf.length(lineString) * 100) / 100;
-};
+  return Math.round(turf.length(line) * 100) / 100;
+}
 export default analyzeGPS;
--- a/tests/11_geospatial/outputs/google_gemini-2.5-pro.js
+++ b/tests/11_geospatial/outputs/google_gemini-2.5-pro.js
@@ -1,18 +1,18 @@
 const analyzeGPS = async (readings, boundary) => {
-  const t = await import('https://cdn.jsdelivr.net/npm/@turf/turf');
+  const { booleanPointInPolygon, lineString, length } = await import('https://cdn.jsdelivr.net/npm/@turf/turf/+esm');

-  const sortedLocations = readings
-    .filter(r => t.booleanPointInPolygon(r.loc, boundary, { ignoreBoundary: true }))
+  const sortedCoords = readings
+    .filter(({ loc }) => booleanPointInPolygon(loc, boundary, { ignoreBoundary: true }))
    .sort((a, b) => a.ts - b.ts)
-    .map(r => r.loc);
+    .map(({ loc }) => loc);

-  if (sortedLocations.length < 2) {
+  if (sortedCoords.length < 2) {
    return 0;
  }

-  const path = t.lineString(sortedLocations);
-  const length = t.length(path, { units: 'kilometers' });
+  const path = lineString(sortedCoords);
+  const dist = length(path, { units: 'kilometers' });

-  return Math.round(length * 100) / 100;
+  return Math.round(dist * 1e2) / 1e2;
 };
 export default analyzeGPS;
--- a/tests/11_geospatial/outputs/google_gemini-3-pro-preview
+++ b/tests/11_geospatial/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,12 @@
+const analyzeGPS = async (readings, boundary) => {
+  const turf = await import('https://cdn.jsdelivr.net/npm/@turf/turf/+esm');
+  
+  const pts = readings
+    .filter(({ loc }) => turf.booleanPointInPolygon(turf.point(loc), boundary))
+    .sort((a, b) => a.ts - b.ts)
+    .map(r => r.loc);
+
+  return pts.length < 2 ? 0 : 
+    +turf.length(turf.lineString(pts), { units: 'kilometers' }).toFixed(2);
+};
+export default analyzeGPS;
--- a/tests/11_geospatial/outputs/google_gemini-3-pro-preview.js
+++ b/tests/11_geospatial/outputs/google_gemini-3-pro-preview.js
@@ -1,13 +0,0 @@
-const analyzeGPS = async (readings, boundary) => {
-  const { point, booleanPointInPolygon, lineString, length } = await import('https://esm.sh/@turf/turf');
-
-  const path = readings
-    .filter(({ loc }) => booleanPointInPolygon(point(loc), boundary, { ignoreBoundary: true }))
-    .sort((a, b) => a.ts - b.ts)
-    .map(({ loc }) => loc);
-
-  return path.length < 2 
-    ? 0 
-    : +length(lineString(path), { units: 'kilometers' }).toFixed(2);
-};
-export default analyzeGPS;
--- a/tests/11_geospatial/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/11_geospatial/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,12 +1,11 @@
-async function analyzeGPS(readings, boundary) {
-  try {
-    if (!Array.isArray(readings) || !boundary) return 0;
-    const turf = await import('https://cdn.skypack.dev/@turf/turf'),
-          v = readings.filter(r => r?.loc && turf.booleanPointInPolygon(r.loc, boundary, {ignoreBoundary: true}))
-                     .sort((a, b) => a.ts - b.ts);
-    return v.length < 2 ? 0 : Math.round(turf.length(turf.lineString(v.map(r => r.loc)), {units: 'kilometers'}) * 100) / 100;
-  } catch {
-    return 0;
-  }
+export async function analyzeGPS(readings, boundary) {
+  const { booleanPointInPolygon, lineString, length } = await import('https://cdn.skypack.dev/@turf/turf');
+  
+  const pts = readings
+    .filter(r => booleanPointInPolygon(r.loc, boundary))
+    .sort((a, b) => a.ts - b.ts)
+    .map(r => r.loc);
+  
+  return pts.length < 2 ? 0 : +length(lineString(pts)).toFixed(2);
 }
 export default analyzeGPS;
--- a/tests/11_geospatial/outputs/openai_gpt-5.1-codex.js
+++ b/tests/11_geospatial/outputs/openai_gpt-5.1-codex.js
@@ -1,8 +1,13 @@
-async function analyzeGPS(readings,boundary){
-const{booleanPointInPolygon,lineString,length}=await import('https://cdn.jsdelivr.net/npm/@turf/turf@6.5.0/dist/turf.esm.min.js');
-const pts=readings.filter(r=>booleanPointInPolygon({type:'Point',coordinates:r.loc},boundary,{ignoreBoundary:true})).sort((a,b)=>a.ts-b.ts);
-if(pts.length<2)return 0;
-const km=length(lineString(pts.map(r=>r.loc)),{units:'kilometers'});
-return Math.round(km*100)/100;
+let turf$;
+const useTurf = () => turf$ ??= import('https://cdn.skypack.dev/@turf/turf?min');
+
+async function analyzeGPS(readings, boundary) {
+  const t = await useTurf();
+  const pts = readings
+    .filter(r => t.booleanPointInPolygon(t.point(r.loc), boundary))
+    .sort((a, b) => a.ts - b.ts);
+  if (pts.length < 2) return 0;
+  const km = t.length(t.lineString(pts.map(r => r.loc)), { units: 'kilometers' });
+  return Math.round(km * 100) / 100;
 }
 export default analyzeGPS;
--- a/tests/11_geospatial/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/11_geospatial/outputs/openrouter_sherlock-think-alpha.js
@@ -1,8 +1,8 @@
 async function analyzeGPS(readings,boundary){
-  const{turf}=await import('https://cdn.skypack.dev/@turf/turf');
+  const{booleanPointInPolygon,lineString,length}=await import('https://cdn.skypack.dev/@turf/turf');
  const valid=readings
-    .filter(({loc})=>turf.booleanPointInPolygon([loc[0],loc[1]],boundary,{ignoreBoundary:true}))
+    .filter(r=>booleanPointInPolygon({type:'Point',coordinates:r.loc},boundary,{ignoreBoundary:true}))
    .sort((a,b)=>a.ts-b.ts);
-  return valid.length<2?0:+turf.length(turf.lineString(valid.map(({loc})=>loc)),{units:'kilometers'}).toFixed(2);
+  return valid.length<2?0:+length(lineString(valid.map(r=>r.loc)),{units:'kilometers'}).toFixed(2);
 }
 export default analyzeGPS;
--- a/tests/1_dijkstra/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/1_dijkstra/outputs/anthropic_claude-sonnet-4.5
@@ -1,27 +1,24 @@
 async function findShortestPath(graph, start, end) {
-  const { default: PriorityQueue } = await import('https://cdn.skypack.dev/js-priority-queue');
+  const { default: PriorityQueue } = await import('https://cdn.jsdelivr.net/npm/js-priority-queue@0.1.5/+esm');
  
-  const distances = {};
-  const visited = new Set();
-  const pq = new PriorityQueue({ comparator: (a, b) => a.dist - b.dist });
-  
-  for (const node in graph) distances[node] = Infinity;
+  const distances = Object.keys(graph).reduce((acc, node) => ({ ...acc, [node]: Infinity }), {});
  distances[start] = 0;
+  
+  const pq = new PriorityQueue({ comparator: (a, b) => a.dist - b.dist });
  pq.queue({ node: start, dist: 0 });
  
+  const visited = new Set();
+  
  while (pq.length) {
    const { node, dist } = pq.dequeue();
    
    if (visited.has(node)) continue;
-    if (node === end) return dist;
-    
    visited.add(node);
    
-    for (const neighbor in graph[node]) {
-      if (visited.has(neighbor)) continue;
-      
-      const newDist = dist + graph[node][neighbor];
-      
+    if (node === end) return dist;
+    
+    for (const [neighbor, weight] of Object.entries(graph[node] || {})) {
+      const newDist = dist + weight;
      if (newDist < distances[neighbor]) {
        distances[neighbor] = newDist;
        pq.queue({ node: neighbor, dist: newDist });
@@ -29,6 +26,6 @@ async function findShortestPath(graph, start, end) {
    }
  }
  
-  return Infinity;
+  return distances[end];
 }
 export default findShortestPath;
--- a/tests/1_dijkstra/outputs/google_gemini-2.5-pro.js
+++ b/tests/1_dijkstra/outputs/google_gemini-2.5-pro.js
@@ -1,34 +1,40 @@
-const findShortestPath = async (graph, start, end) => {
-  const { default: PriorityQueue } = await import('https://cdn.jsdelivr.net/npm/js-priority-queue@0.1.5/priority-queue.min.js');
-
-  const dists = Object.fromEntries(
-    Object.keys(graph).map(node => [node, Infinity])
+async function findShortestPath(graph, startNode, endNode) {
+  const { default: PriorityQueue } = await import(
+    'https://cdn.jsdelivr.net/npm/js-priority-queue@0.1.5/priority-queue.min.js'
  );

+  const distances = {};
  const pq = new PriorityQueue({ comparator: (a, b) => a.priority - b.priority });

-  dists[start] = 0;
-  pq.queue({ node: start, priority: 0 });
+  for (const node in graph) {
+    distances[node] = Infinity;
+  }
+  distances[startNode] = 0;
+
+  pq.queue({ node: startNode, priority: 0 });

  while (pq.length) {
-    const { node: u, priority: currentDist } = pq.dequeue();
+    const { node: currentNode, priority: currentPriority } = pq.dequeue();

-    if (currentDist > dists[u]) {
+    if (currentPriority > distances[currentNode]) {
      continue;
    }

-    if (u === end) break;
+    if (currentNode === endNode) {
+      return currentPriority;
+    }

-    for (const [v, weight] of Object.entries(graph[u])) {
-      const distThroughU = dists[u] + weight;
+    for (const neighbor in graph[currentNode]) {
+      const weight = graph[currentNode][neighbor];
+      const newDistance = currentPriority + weight;

-      if (distThroughU < dists[v]) {
-        dists[v] = distThroughU;
-        pq.queue({ node: v, priority: distThroughU });
+      if (newDistance < distances[neighbor]) {
+        distances[neighbor] = newDistance;
+        pq.queue({ node: neighbor, priority: newDistance });
      }
    }
  }

-  return dists[end];
-};
+  return Infinity;
+}
 export default findShortestPath;
--- a/tests/1_dijkstra/outputs/google_gemini-3-pro-preview
+++ b/tests/1_dijkstra/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,22 @@
+const findShortestPath = async (g, s, e) => {
+  const { default: Q } = await import('https://esm.sh/js-priority-queue')
+  const q = new Q({ comparator: (a, b) => a.w - b.w })
+  const d = { [s]: 0 }
+  q.queue({ n: s, w: 0 })
+
+  while (q.length) {
+    const { n, w } = q.dequeue()
+    if (n === e) return w
+    if (w > (d[n] ?? Infinity)) continue
+
+    for (const v in g[n]) {
+      const k = w + g[n][v]
+      if (k < (d[v] ?? Infinity)) {
+        d[v] = k
+        q.queue({ n: v, w: k })
+      }
+    }
+  }
+  return Infinity
+}
+export default findShortestPath;
--- a/tests/1_dijkstra/outputs/google_gemini-3-pro-preview.js
+++ b/tests/1_dijkstra/outputs/google_gemini-3-pro-preview.js
@@ -1,19 +0,0 @@
-const findShortestPath = async (g, s, e) => {
-  const { default: P } = await import('https://cdn.skypack.dev/js-priority-queue')
-  const d = { [s]: 0 }, q = new P({ comparator: (a, b) => a.w - b.w })
-  q.queue({ w: 0, n: s })
-  while (q.length) {
-    const { w, n } = q.dequeue()
-    if (n === e) return w
-    if (w > (d[n] ?? 1 / 0)) continue
-    for (const v in g[n]) {
-      const t = w + g[n][v]
-      if (t < (d[v] ?? 1 / 0)) {
-        d[v] = t
-        q.queue({ w: t, n: v })
-      }
-    }
-  }
-  return 1 / 0
-}
-export default findShortestPath;
--- a/tests/1_dijkstra/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/1_dijkstra/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,24 +1,28 @@
-async function findShortestPath(graph, start, end) {
-  const { default: Q } = await import('https://cdn.jsdelivr.net/npm/js-priority-queue@0.1.5/index.min.js');
-  const dist = { [start]: 0 }, visited = new Set();
-  const pq = new Q({ comparator: (a, b) => dist[a] - dist[b] });
-  pq.queue(start);
+export async function findShortestPath(graph, start, end) {
+  const { default: PriorityQueue } = await import('https://cdn.jsdelivr.net/npm/js-priority-queue@0.1.5/index.min.js');
+  
+  if (!graph[start] || !graph[end]) return Infinity;
+  if (start === end) return 0;
+  
+  const dist = { [start]: 0 };
+  const pq = new PriorityQueue({ comparator: (a, b) => a[0] - b[0] });
+  pq.queue([0, start]);
  
  while (pq.length) {
-    const node = pq.dequeue();
-    if (node === end) return dist[end];
-    if (visited.has(node)) continue;
-    visited.add(node);
-    const nodeDist = dist[node];
-    for (const [adj, weight] of Object.entries(graph[node] || {})) {
-      if (visited.has(adj)) continue;
-      const newDist = nodeDist + weight;
-      if (newDist < (dist[adj] ?? Infinity)) {
-        dist[adj] = newDist;
-        pq.queue(adj);
+    const [d, node] = pq.dequeue();
+    
+    if (node === end) return d;
+    if (d > (dist[node] ?? Infinity)) continue;
+    
+    for (const [n, w] of Object.entries(graph[node] || {})) {
+      const nd = d + w;
+      if (nd < (dist[n] ?? Infinity)) {
+        dist[n] = nd;
+        pq.queue([nd, n]);
      }
    }
  }
+  
  return Infinity;
 }
 export default findShortestPath;
--- a/tests/1_dijkstra/outputs/openai_gpt-5.1-codex.js
+++ b/tests/1_dijkstra/outputs/openai_gpt-5.1-codex.js
@@ -1,27 +1,30 @@
-const src='https://esm.sh/js-priority-queue@1'
-let load
-const getPQ=()=>load||(load=import(src).then(m=>m.default))
-const findShortestPath=async(g,s,e)=>{
-  const PQ=await getPQ()
-  const pq=new PQ({comparator:(a,b)=>a[0]-b[0]})
-  const dist=new Map([[s,0]])
-  const done=new Set
-  pq.queue([0,s])
-  while(pq.length){
-    const [d,n]=pq.dequeue()
-    if(done.has(n))continue
-    done.add(n)
-    if(n===e)return d
-    const edges=g[n]
-    if(!edges)continue
-    for(const k in edges){
-      const w=d+edges[k]
-      if(w<(dist.get(k)??Infinity)){
-        dist.set(k,w)
-        pq.queue([w,k])
-      }
-    }
-  }
-  return Infinity
+let pqPromise
+const getPQ=()=>pqPromise??=import('https://cdn.jsdelivr.net/npm/js-priority-queue@latest/+esm')
+
+async function findShortestPath(g,s,t){
+if(!g||!(s in g)||!(t in g)) return Infinity
+if(s===t) return 0
+const {default:PriorityQueue}=await getPQ()
+const d={[s]:0}
+const seen=new Set
+const q=new PriorityQueue({comparator:(a,b)=>a[0]-b[0]})
+q.queue([0,s])
+while(q.length){
+const [w,n]=q.dequeue()
+if(seen.has(n)) continue
+seen.add(n)
+if(n===t) return w
+const edges=g[n]
+if(!edges) continue
+for(const k in edges){
+if(seen.has(k)) continue
+const nw=w+edges[k]
+if(nw<(d[k]??Infinity)){
+d[k]=nw
+q.queue([nw,k])
+}
+}
+}
+return Infinity
 }
 export default findShortestPath;
--- a/tests/1_dijkstra/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/1_dijkstra/outputs/openrouter_sherlock-think-alpha.js
@@ -1,23 +1,24 @@
 async function findShortestPath(graph,start,end){
-  const{default:PriorityQueue}=await import('https://cdn.skypack.dev/js-priority-queue');
+  const {PriorityQueue}=await import('https://cdn.skypack.dev/js-priority-queue');
+  if(start===end)return 0;
  const dist={};
-  const pq=new PriorityQueue({comparator:(x,y)=>x.d-y.d});
-  for(const n of Object.keys(graph))dist[n]=Infinity;
  dist[start]=0;
-  pq.queue({n:start,d:0});
+  const pq=new PriorityQueue({comparator:(a,b)=>a.dist-b.dist});
+  pq.queue({node:start,dist:0});
  while(!pq.isEmpty()){
-    const{n:n,d:d}=pq.dequeue();
-    if(d>dist[n])continue;
-    if(n===end)return d;
-    for(let nei in graph[n]){
-      const w=graph[n][nei];
-      const alt=d+w;
-      if(alt<dist[nei]){
-        dist[nei]=alt;
-        pq.queue({n:nei,d:alt});
+    const u=pq.dequeue();
+    if(u.dist>dist[u.node])continue;
+    const neighbors=graph[u.node];
+    if(!neighbors)continue;
+    for(const v in neighbors){
+      const w=neighbors[v];
+      const alt=dist[u.node]+w;
+      if(alt<(dist[v]??Infinity)){
+        dist[v]=alt;
+        pq.queue({node:v,dist:alt});
      }
    }
  }
-  return Infinity;
+  return dist[end]??Infinity;
 }
 export default findShortestPath;
--- a/tests/2_convex_hull/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/2_convex_hull/outputs/anthropic_claude-sonnet-4.5
@@ -1,29 +1,34 @@
 async function findConvexHull(points) {
-  const _ = await import('https://cdn.jsdelivr.net/npm/lodash-es@4.17.21/lodash.min.js').then(m => m.default);
+  const { sortBy, uniqWith } = await import('https://cdn.jsdelivr.net/npm/lodash-es@4.17.21/lodash.js');
  
  if (points.length < 3) return points;
  
-  const sorted = _.sortBy(_.uniqWith(points, _.isEqual), ['x', 'y']);
+  const cross = (o, a, b) => (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x);
+  
+  const sorted = sortBy(uniqWith(points, (a, b) => a.x === b.x && a.y === b.y), ['x', 'y']);
  
  if (sorted.length < 3) return sorted;
  
-  const cross = (o, a, b) => (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x);
-  
-  const buildHull = (pts) => {
-    const hull = [];
-    for (const p of pts) {
-      while (hull.length >= 2 && cross(hull[hull.length - 2], hull[hull.length - 1], p) <= 0) {
-        hull.pop();
-      }
-      hull.push(p);
+  const lower = [];
+  for (const p of sorted) {
+    while (lower.length >= 2 && cross(lower[lower.length - 2], lower[lower.length - 1], p) <= 0) {
+      lower.pop();
    }
-    hull.pop();
-    return hull;
-  };
+    lower.push(p);
+  }
  
-  const lower = buildHull(sorted);
-  const upper = buildHull([...sorted].reverse());
+  const upper = [];
+  for (let i = sorted.length - 1; i >= 0; i--) {
+    const p = sorted[i];
+    while (upper.length >= 2 && cross(upper[upper.length - 2], upper[upper.length - 1], p) <= 0) {
+      upper.pop();
+    }
+    upper.push(p);
+  }
  
-  return [...lower, ...upper];
+  lower.pop();
+  upper.pop();
+  
+  return lower.concat(upper);
 }
 export default findConvexHull;
--- a/tests/2_convex_hull/outputs/google_gemini-2.5-pro.js
+++ b/tests/2_convex_hull/outputs/google_gemini-2.5-pro.js
@@ -1,27 +1,28 @@
-const findConvexHull = async (points) => {
-  const _ = await import('https://cdn.jsdelivr.net/npm/lodash@4.17.21/lodash.min.js');
+async const findConvexHull = async (points) => {
+  const { sortBy, uniqWith, isEqual } = await import('https://cdn.jsdelivr.net/npm/lodash-es@4.17.21/lodash.min.js');

-  const uniquePts = _.uniqWith(points, _.isEqual);
-  if (uniquePts.length < 3) {
-    return uniquePts;
+  const cross = (p1, p2, p3) =>
+    (p2.x - p1.x) * (p3.y - p1.y) - (p2.y - p1.y) * (p3.x - p1.x);
+
+  const sortedPoints = sortBy(uniqWith(points, isEqual), ['x', 'y']);
+
+  if (sortedPoints.length <= 3) {
+    return sortedPoints;
  }

-  const crossProduct = (o, a, b) =>
-    (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x);
-
-  const sortedPts = _.sortBy(uniquePts, ['x', 'y']);
-
-  const buildHalfHull = (pts) =>
-    pts.reduce((hull, p) => {
-      while (hull.length >= 2 && crossProduct(hull.at(-2), hull.at(-1), p) <= 0) {
+  const buildHull = (pointSet) => {
+    const hull = [];
+    for (const p of pointSet) {
+      while (hull.length >= 2 && cross(hull[hull.length - 2], hull[hull.length - 1], p) <= 0) {
        hull.pop();
      }
      hull.push(p);
-      return hull;
-    }, []);
+    }
+    return hull;
+  };

-  const lower = buildHalfHull(sortedPts);
-  const upper = buildHalfHull([...sortedPts].reverse());
+  const lower = buildHull(sortedPoints);
+  const upper = buildHull([...sortedPoints].reverse());

  return [...lower.slice(0, -1), ...upper.slice(0, -1)];
 };
--- a/tests/2_convex_hull/outputs/google_gemini-3-pro-preview
+++ b/tests/2_convex_hull/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,20 @@
+const findConvexHull = async (pts) => {
+  const { sortBy: S, uniqWith: U, isEqual: E } = await import('https://cdn.jsdelivr.net/npm/lodash-es@4.17.21/+esm')
+  const p = S(U(pts, E), ['x', 'y'])
+  if (p.length < 3) return p
+  const X = (a, b, o) => (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x)
+  const B = (list) => {
+    const h = []
+    for (const pt of list) {
+      while (h.length >= 2 && X(h[h.length - 1], pt, h[h.length - 2]) <= 0) h.pop()
+      h.push(pt)
+    }
+    return h
+  }
+  const l = B(p)
+  const u = B(p.reverse())
+  l.pop()
+  u.pop()
+  return l.concat(u)
+}
+export default findConvexHull;
--- a/tests/2_convex_hull/outputs/google_gemini-3-pro-preview.js
+++ b/tests/2_convex_hull/outputs/google_gemini-3-pro-preview.js
@@ -1,22 +0,0 @@
-const findConvexHull = async (pts) => {
-  const { default: _ } = await import('https://cdn.skypack.dev/lodash');
-  
-  const P = _.sortBy(_.uniqWith(pts, _.isEqual), ['x', 'y']);
-  const n = P.length, k = [], C = (a, b, c) => (b.x - a.x) * (c.y - a.y) - (b.y - a.y) * (c.x - a.x);
-  
-  if (n < 3) return P;
-
-  for (let i = 0; i < n; i++) {
-    while (k.length >= 2 && C(k[k.length - 2], k[k.length - 1], P[i]) <= 0) k.pop();
-    k.push(P[i]);
-  }
-
-  for (let i = n - 2, t = k.length + 1; i >= 0; i--) {
-    while (k.length >= t && C(k[k.length - 2], k[k.length - 1], P[i]) <= 0) k.pop();
-    k.push(P[i]);
-  }
-
-  k.pop();
-  return k;
-};
-export default findConvexHull;
--- a/tests/2_convex_hull/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/2_convex_hull/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,19 +1,18 @@
-const findConvexHull = async (pts) => {
-  const { sortBy, uniqBy } = await import('https://esm.sh/lodash-es@4');
-  const cross = (a, b, c) => (b.x-a.x)*(c.y-a.y)-(b.y-a.y)*(c.x-a.x);
-  const sorted = sortBy(uniqBy(pts, p => `${p.x},${p.y}`), ['x', 'y']);
-  if (sorted.length < 3) return sorted;
+async function findConvexHull(pts) {
+  if (!Array.isArray(pts)) return [];
+  const { default: _ } = await import('https://cdn.jsdelivr.net/npm/lodash@4.17.21/lodash.min.js');
+  const points = _.sortBy(_.uniqWith(pts, (a, b) => a.x === b.x && a.y === b.y), ['x', 'y']);
+  if (points.length < 3) return points;
+  const cross = (o, a, b) => (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x);
+  const build = (h, p) => {
+    while (h.length >= 2 && cross(h[h.length - 2], h[h.length - 1], p) <= 0) h.pop();
+    h.push(p);
+  };
  const lower = [], upper = [];
-  for (const p of sorted) {
-    while (lower.length > 1 && cross(lower[lower.length-2], lower[lower.length-1], p) <= 0) lower.pop();
-    lower.push(p);
-  }
-  for (let i = sorted.length - 1; i >= 0; i--) {
-    const p = sorted[i];
-    while (upper.length > 1 && cross(upper[upper.length-2], upper[upper.length-1], p) <= 0) upper.pop();
-    upper.push(p);
-  }
-  lower.pop(); upper.pop();
-  return [...lower, ...upper];
-};
+  for (const p of points) build(lower, p);
+  for (let i = points.length; i--; ) build(upper, points[i]);
+  lower.pop();
+  upper.pop();
+  return lower.concat(upper);
+}
 export default findConvexHull;
--- a/tests/2_convex_hull/outputs/openai_gpt-5.1-codex.js
+++ b/tests/2_convex_hull/outputs/openai_gpt-5.1-codex.js
@@ -1,21 +1,19 @@
-async function findConvexHull(points){
-  const m=await import('https://cdn.skypack.dev/lodash?min')
-  const _=m.default||m
-  const pts=_.sortBy(_.uniqWith(points,(a,b)=>a.x===b.x&&a.y===b.y),['x','y'])
-  if(pts.length<3)return pts.slice()
-  const h=[]
-  const c=(o,a,b)=>(a.x-o.x)*(b.y-o.y)-(a.y-o.y)*(b.x-o.x)
-  for(const p of pts){
-    while(h.length>1&&c(h.at(-2),h.at(-1),p)<=0)h.pop()
-    h.push(p)
+async function findConvexHull(points) {
+  const { uniqWith, isEqual, sortBy } = await import('https://cdn.jsdelivr.net/npm/lodash-es@4.17.21/lodash.js')
+  const pts = sortBy(uniqWith(points, isEqual), ['x', 'y'])
+  if (pts.length < 3) return pts
+  const cross = (o, a, b) => (a.x - o.x) * (b.y - o.y) - (a.y - o.y) * (b.x - o.x)
+  const scan = (src, acc = []) => {
+    src.forEach(p => {
+      while (acc.length > 1 && cross(acc.at(-2), acc.at(-1), p) <= 0) acc.pop()
+      acc.push(p)
+    })
+    return acc
  }
-  const t=h.length+1
-  for(let i=pts.length-2;i>=0;i--){
-    const p=pts[i]
-    while(h.length>=t&&c(h.at(-2),h.at(-1),p)<=0)h.pop()
-    h.push(p)
-  }
-  h.pop()
-  return h
+  const lower = scan(pts)
+  const upper = scan([...pts].reverse())
+  lower.pop()
+  upper.pop()
+  return lower.concat(upper)
 }
 export default findConvexHull;
--- a/tests/2_convex_hull/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/2_convex_hull/outputs/openrouter_sherlock-think-alpha.js
@@ -1,22 +1,12 @@
-async function findConvexHull(points){
-  const{default:L}=await import('https://cdn.skypack.dev/lodash-es');
-  const{sortBy,uniqBy}=L;
-  let p=uniqBy(points,({x,y})=>`${x},${y}`);
+async function findConvexHull(pts){
+  const{ sortBy:s,uniqBy:u }=await import('https://cdn.skypack.dev/lodash');
+  if(!pts?.length)return[];
+  const p=u(s(pts,['x','y']),t=>`${t.x},${t.y}`);
  if(p.length<2)return p;
-  p=sortBy(p,['x','y']);
-  const cross=(o,a,b)=>(a.x-o.x)*(b.y-o.y)-(a.y-o.y)*(b.x-o.x);
-  let l=[];
-  for(let i=0;i<p.length;i++){
-    let pt=p[i];
-    while(l.length>=2&&cross(l[l.length-2],l[l.length-1],pt)<=0)l.pop();
-    l.push(pt);
-  }
-  let u=[];
-  for(let i=p.length-1;i>=0;i--){
-    let pt=p[i];
-    while(u.length>=2&&cross(u[u.length-2],u[u.length-1],pt)<=0)u.pop();
-    u.push(pt);
-  }
-  return[...l.slice(0,-1),...u.slice(0,-1)];
+  const c=(o,a,b)=>(a.x-o.x)*(b.y-o.y)-(a.y-o.y)*(b.x-o.x);
+  let l=[];for(let t of p){while(l.length>1&&c(l[l.length-2],l[l.length-1],t)<=0)l.pop();l.push(t)}
+  let r=[];for(let i=p.length-1;i>=0;i--){const t=p[i];while(r.length>1&&c(r[r.length-2],r[r.length-1],t)<=0)r.pop();r.push(t)}
+  l.pop();r.pop();
+  return[...l,...r];
 }
 export default findConvexHull;
--- a/tests/3_lis/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/3_lis/outputs/anthropic_claude-sonnet-4.5
@@ -1,5 +1,5 @@
-async function findLISLength(arr) {
-  const { bisectLeft } = await import('https://cdn.jsdelivr.net/npm/d3-array@3/+esm');
+const findLISLength = async (arr) => {
+  const { bisectLeft } = await import('https://cdn.skypack.dev/d3-array');
  
  if (!arr?.length) return 0;
  
@@ -11,5 +11,5 @@ async function findLISLength(arr) {
  }
  
  return tails.length;
-}
+};
 export default findLISLength;
--- a/tests/3_lis/outputs/google_gemini-2.5-pro.js
+++ b/tests/3_lis/outputs/google_gemini-2.5-pro.js
@@ -1,16 +1,16 @@
-async function findLISLength(nums) {
+const findLISLength = async (nums) => {
  if (!nums?.length) {
    return 0;
  }

-  const { bisectLeft } = await import('https://cdn.jsdelivr.net/npm/d3-array@3/+esm');
-  
-  const sub = [];
+  const { bisectLeft } = await import('https://cdn.jsdelivr.net/npm/d3-array/+esm');

-  for (const num of nums) {
-    sub[bisectLeft(sub, num)] = num;
-  }
+  const tails = [];

-  return sub.length;
-}
+  nums.forEach((num) => {
+    tails[bisectLeft(tails, num)] = num;
+  });
+
+  return tails.length;
+};
 export default findLISLength;
--- a/tests/3_lis/outputs/google_gemini-3-pro-preview
+++ b/tests/3_lis/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,9 @@
+const findLISLength = async (nums) => {
+  const { bisectLeft } = await import('https://cdn.jsdelivr.net/npm/d3-array@3/+esm');
+  const tails = [];
+  for (const n of nums) {
+    tails[bisectLeft(tails, n)] = n;
+  }
+  return tails.length;
+};
+export default findLISLength;
--- a/tests/3_lis/outputs/google_gemini-3-pro-preview.js
+++ b/tests/3_lis/outputs/google_gemini-3-pro-preview.js
@@ -1,7 +0,0 @@
-const findLISLength = async arr => {
-  const { bisectLeft } = await import('https://esm.sh/d3-array')
-  const tails = []
-  for (const n of arr) tails[bisectLeft(tails, n)] = n
-  return tails.length
-}
-export default findLISLength;
--- a/tests/3_lis/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/3_lis/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,12 +1,14 @@
-export const findLISLength = async (arr) => {
-  if (!Array.isArray(arr) || !arr.length) return 0;
-  try {
-    const { bisectLeft } = await import('https://cdn.skypack.dev/d3-array@3');
-    const t = [];
-    for (const v of arr) t[bisectLeft(t, v)] = v;
-    return t.length;
-  } catch {
-    return 0;
+export async function findLISLength(nums) {
+  if (!Array.isArray(nums) || nums.length === 0) return 0;
+  
+  const { bisectLeft } = await import('https://cdn.jsdelivr.net/npm/d3-array@3/+esm');
+  const tails = [];
+  
+  for (const num of nums) {
+    const idx = bisectLeft(tails, num);
+    idx === tails.length ? tails.push(num) : tails[idx] = num;
  }
-};
+  
+  return tails.length;
+}
 export default findLISLength;
--- a/tests/3_lis/outputs/openai_gpt-5.1-codex.js
+++ b/tests/3_lis/outputs/openai_gpt-5.1-codex.js
@@ -1,11 +1,10 @@
-const findLISLength = async nums => {
-  if (!Array.isArray(nums)) throw new TypeError('Expected array')
-  const { bisectLeft } = await import('https://cdn.jsdelivr.net/npm/d3-array@3/+esm')
-  const tails = []
-  for (const v of nums) {
-    const i = bisectLeft(tails, v)
-    i === tails.length ? tails.push(v) : tails[i] = v
+async function findLISLength(a) {
+  const {bisectLeft} = await import('https://cdn.jsdelivr.net/npm/d3-array@3/+esm')
+  const t=[]
+  for (const x of a||[]) {
+    const i = bisectLeft(t,x)
+    t[i]=x
  }
-  return tails.length
+  return t.length
 }
 export default findLISLength;
--- a/tests/3_lis/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/3_lis/outputs/openrouter_sherlock-think-alpha.js
@@ -1,2 +1,2 @@
-async function findLISLength(arr){const{bisectLeft}=await import('https://cdn.skypack.dev/d3-array');let tails=[];for(const num of arr){const i=bisectLeft(tails,num);i===tails.length?tails.push(num):tails[i]=num;}return tails.length;}
+async function findLISLength(a){const{bisectLeft}=await import('https://cdn.skypack.dev/d3-array');let t=[];for(let n of a){let i=bisectLeft(t,n);i===t.length?t.push(n):t[i]=n}return t.length}
 export default findLISLength;
--- a/tests/4_determinant/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/4_determinant/outputs/anthropic_claude-sonnet-4.5
@@ -1,5 +1,5 @@
 const calculateDeterminant = async (matrix) => {
-  const { det } = await import('https://cdn.jsdelivr.net/npm/mathjs@11.11.0/+esm');
+  const { det } = await import('https://cdn.jsdelivr.net/npm/mathjs@11/+esm');
  return det(matrix);
 };
 export default calculateDeterminant;
--- a/tests/4_determinant/outputs/google_gemini-3-pro-preview
+++ b/tests/4_determinant/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,5 @@
+const calculateDeterminant = async m => {
+  const { det } = await import('https://cdn.jsdelivr.net/npm/mathjs/+esm')
+  return det(m)
+}
+export default calculateDeterminant;
--- a/tests/4_determinant/outputs/google_gemini-3-pro-preview.js
+++ b/tests/4_determinant/outputs/google_gemini-3-pro-preview.js
@@ -1,5 +0,0 @@
-const calculateDeterminant = async m => {
-  const { det } = await import('https://esm.sh/mathjs')
-  return det(m)
-}
-export default calculateDeterminant;
--- a/tests/4_determinant/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/4_determinant/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,13 +1,11 @@
-const URL = 'https://cdn.jsdelivr.net/npm/mathjs@11.11.0/+esm';
-let det;
+let mathLib;

 async function calculateDeterminant(m) {
-  const v = Array.isArray(m) && m.length && m.every(r => Array.isArray(r) && r.length === m.length);
-  if (!v) throw new TypeError('Invalid matrix');
+  if (!Array.isArray(m) || !m.length || m.some(r => !Array.isArray(r) || r.length !== m.length)) {
+    throw new TypeError('Invalid square matrix');
+  }
  
-  det ||= await import(URL).then(m => m.det).catch(() => { throw new Error('Load failed'); });
-  
-  try { return det(m); }
-  catch { throw new Error('Calc failed'); }
+  const {det} = await (mathLib ||= import('https://cdn.jsdelivr.net/npm/mathjs@11.11.0/lib/browser/math.js'));
+  return det(m);
 }
 export default calculateDeterminant;
--- a/tests/4_determinant/outputs/openai_gpt-5.1-codex.js
+++ b/tests/4_determinant/outputs/openai_gpt-5.1-codex.js
@@ -1,5 +1,3 @@
-async function calculateDeterminant(m){
-  const {det}=await import('https://cdn.jsdelivr.net/npm/mathjs@11.11.1/+esm')
-  return det(m)
-}
+let d
+const calculateDeterminant=async m=> (await (d??=import('https://cdn.jsdelivr.net/npm/mathjs@11.11.0/+esm'))).det(m)
 export default calculateDeterminant;
--- a/tests/4_determinant/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/4_determinant/outputs/openrouter_sherlock-think-alpha.js
@@ -1,2 +1,2 @@
-async function calculateDeterminant(matrix){const math=await import('https://esm.sh/mathjs');return math.det(matrix);}
+async function calculateDeterminant(m){const{det}=await import('https://esm.sh/mathjs');return det(m);}
 export default calculateDeterminant;
--- a/tests/5_markdown_parser/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/5_markdown_parser/outputs/anthropic_claude-sonnet-4.5
@@ -1,5 +1,5 @@
-const parseMarkdown = async (md) => {
-  const { marked } = await import('https://cdn.jsdelivr.net/npm/marked@11/+esm');
+async function parseMarkdown(md) {
+  const { marked } = await import('https://cdn.jsdelivr.net/npm/marked@11.1.1/+esm');
  return marked.parse(md);
-};
+}
 export default parseMarkdown;
--- a/tests/5_markdown_parser/outputs/google_gemini-2.5-pro.js
+++ b/tests/5_markdown_parser/outputs/google_gemini-2.5-pro.js
@@ -1,12 +1,24 @@
-const parseMarkdown = async (markdownText) => {
-  const cdnUrl = 'https://cdn.jsdelivr.net/npm/marked@13.0.2/lib/marked.esm.js';
-
+const parseMarkdown = async (md) => {
  try {
-    const { marked } = await import(cdnUrl);
-    return marked.parse(markdownText || '');
+    const [
+      { marked },
+      { default: DOMPurify }
+    ] = await Promise.all([
+      import('https://cdn.jsdelivr.net/npm/marked@13.0.0/lib/marked.esm.js'),
+      import('https://cdn.jsdelivr.net/npm/dompurify@3.1.5/dist/purify.es.mjs')
+    ]);
+
+    const rawHtml = marked.parse(md || '', {
+      gfm: true,
+      breaks: true,
+    });
+
+    return DOMPurify.sanitize(rawHtml);
  } catch (err) {
-    console.error(`Failed to parse markdown: ${err}`);
-    return markdownText || '';
+    console.error('Failed to parse markdown:', err);
+    const el = document.createElement('div');
+    el.textContent = md || '';
+    return el.innerHTML;
  }
 };
 export default parseMarkdown;
--- a/tests/5_markdown_parser/outputs/google_gemini-3-pro-preview
+++ b/tests/5_markdown_parser/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,5 @@
+const parseMarkdown = async text => {
+  const { parse } = await import('https://esm.sh/marked')
+  return parse(text)
+}
+export default parseMarkdown;
--- a/tests/5_markdown_parser/outputs/google_gemini-3-pro-preview.js
+++ b/tests/5_markdown_parser/outputs/google_gemini-3-pro-preview.js
@@ -1,5 +0,0 @@
-const parseMarkdown = async md => {
-  const { parse } = await import('https://cdn.jsdelivr.net/npm/marked@12.0.1/lib/marked.esm.js')
-  return parse(md)
-}
-export default parseMarkdown;
--- a/tests/5_markdown_parser/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/5_markdown_parser/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,13 +1,15 @@
 async function parseMarkdown(md) {
-  const u = 'https://cdn.jsdelivr.net/npm/marked@9/lib/marked.esm.js';
-  const o = {gfm:true,breaks:true,headerIds:false,mangle:false};
-  
+  if (!md?.trim()) return '';
  try {
-    const {marked} = await import(u);
-    marked.setOptions(o);
-    return marked.parse(md);
-  } catch(e) {
-    return `<p>${e.message}</p>`;
+    const [{ marked }, { default: DOMPurify }] = await Promise.all([
+      import('https://esm.sh/marked@9'),
+      import('https://esm.sh/dompurify@3')
+    ]);
+    marked.setOptions({ gfm: true, breaks: true });
+    return DOMPurify.sanitize(marked.parse(md));
+  } catch (e) {
+    console.error('Markdown parse error:', e);
+    return '';
  }
 }
 export default parseMarkdown;
--- a/tests/5_markdown_parser/outputs/openai_gpt-5.1-codex.js
+++ b/tests/5_markdown_parser/outputs/openai_gpt-5.1-codex.js
@@ -1,5 +1,12 @@
-const parseMarkdown=async m=>{
-  const {marked}=await import('https://cdn.jsdelivr.net/npm/marked@11.1.0/lib/marked.esm.js')
-  return marked.parse(m)
-}
+const parseMarkdown=(()=>{
+  let m,r
+  return async t=>{
+    if(!r){
+      ({marked:m}=await import('https://cdn.jsdelivr.net/npm/marked@11.1.1/+esm'))
+      m.setOptions({gfm:1,mangle:0,headerIds:0})
+      r=1
+    }
+    return m.parse(t)
+  }
+})()
 export default parseMarkdown;
--- a/tests/5_markdown_parser/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/5_markdown_parser/outputs/openrouter_sherlock-think-alpha.js
@@ -1,5 +1,6 @@
 async function parseMarkdown(md){
  const{marked}=await import('https://esm.sh/marked');
-  return marked.parse(md);
+  const{DOMPurify}=await import('https://esm.sh/dompurify');
+  return DOMPurify.sanitize(marked.parse(md));
 }
 export default parseMarkdown;
--- a/tests/6_csv_processor/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/6_csv_processor/outputs/anthropic_claude-sonnet-4.5
@@ -1,23 +1,26 @@
-const processCSV = async (csvString, { filterColumn, filterValue, groupBy, aggregateColumn, operation }) => {
-  const { default: Papa } = await import('https://cdn.jsdelivr.net/npm/[email protected]/+esm');
-  
-  const { data } = Papa.parse(csvString, { header: true, skipEmptyLines: true });
+const processCSV = async (csv, { filterColumn, filterValue, groupBy, aggregateColumn, operation }) => {
+  const [{ default: Papa }] = await Promise.all([
+    import('https://cdn.jsdelivr.net/npm/papaparse@5.4.1/+esm')
+  ]);
+
+  const { data } = Papa.parse(csv, { header: true, skipEmptyLines: true });
  
  const filtered = data.filter(row => row[filterColumn] == filterValue);
  
  const groups = filtered.reduce((acc, row) => {
    const key = row[groupBy];
-    (acc[key] ??= []).push(row);
+    if (!acc[key]) acc[key] = [];
+    acc[key].push(row);
    return acc;
  }, {});
-  
+
  return Object.entries(groups).map(([key, rows]) => {
-    const values = rows.map(r => +r[aggregateColumn] || 0);
+    const values = rows.map(r => Number(r[aggregateColumn]) || 0);
    const sum = values.reduce((a, b) => a + b, 0);
    
    return {
      [groupBy]: key,
-      result: operation === 'sum' ? sum : operation === 'avg' ? sum / values.length : rows.length
+      result: operation === 'sum' ? sum : operation === 'avg' ? sum / values.length : values.length
    };
  });
 };
--- a/tests/6_csv_processor/outputs/google_gemini-2.5-pro.js
+++ b/tests/6_csv_processor/outputs/google_gemini-2.5-pro.js
@@ -1,38 +1,34 @@
 async function processCSV(csvString, config) {
  const {
-    filterColumn,
-    filterValue,
-    groupBy,
-    aggregateColumn,
-    operation
+    filterColumn: fc,
+    filterValue: fv,
+    groupBy: gb,
+    aggregateColumn: ac,
+    operation: op
  } = config;

-  const { default: Papa } = await import('https://cdn.jsdelivr.net/npm/papaparse@5.4.1/+esm');
+  const { default: Papa } = await import('https://cdn.jsdelivr.net/npm/papaparse@5.3.2/papaparse.min.js');

  const { data } = Papa.parse(csvString, {
    header: true,
    skipEmptyLines: true
  });

-  const groups = data
-    .filter(row => row[filterColumn] == filterValue)
-    .reduce((acc, row) => {
-      const key = row[groupBy];
-      const val = Number(row[aggregateColumn]) || 0;
+  const groups = data.reduce((acc, row) => {
+    if (row[fc] == fv) {
+      const key = row[gb];
+      const val = Number(row[ac]) || 0;

-      if (operation === 'avg') {
-        const group = (acc[key] ??= { sum: 0, count: 0 });
-        group.sum += val;
-        group.count++;
-      } else {
-        acc[key] = (acc[key] ?? 0) + (operation === 'sum' ? val : 1);
-      }
-      return acc;
-    }, {});
+      acc[key] = acc[key] || { sum: 0, count: 0 };
+      acc[key].sum += val;
+      acc[key].count++;
+    }
+    return acc;
+  }, {});

-  return Object.entries(groups).map(([groupValue, aggData]) => ({
-    [groupBy]: groupValue,
-    result: operation === 'avg' ? aggData.sum / aggData.count : aggData,
+  return Object.entries(groups).map(([key, { sum, count }]) => ({
+    [gb]: key,
+    result: op === 'avg' ? sum / count : op === 'sum' ? sum : count,
  }));
 }
 export default processCSV;
--- a/tests/6_csv_processor/outputs/google_gemini-3-pro-preview
+++ b/tests/6_csv_processor/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,19 @@
+export const processCSV = async (csvString, { filterColumn, filterValue, groupBy, aggregateColumn, operation }) => {
+  const [{ csvParse }, { rollup, sum, mean }] = await Promise.all([
+    import('https://esm.sh/d3-dsv'),
+    import('https://esm.sh/d3-array')
+  ]);
+
+  const getValue = d => +(d[aggregateColumn]) || 0;
+  
+  const grouped = rollup(
+    csvParse(csvString).filter(row => row[filterColumn] == filterValue),
+    group => operation === 'count' 
+      ? group.length 
+      : (operation === 'sum' ? sum : mean)(group, getValue),
+    row => row[groupBy]
+  );
+
+  return Array.from(grouped, ([key, result]) => ({ [groupBy]: key, result }));
+};
+export default processCSV;
--- a/tests/6_csv_processor/outputs/google_gemini-3-pro-preview.js
+++ b/tests/6_csv_processor/outputs/google_gemini-3-pro-preview.js
@@ -1,22 +0,0 @@
-const processCSV = async (csv, { filterColumn: fKey, filterValue: fVal, groupBy: gKey, aggregateColumn: aKey, operation: op }) => {
-  const { parse } = await import('https://cdn.jsdelivr.net/npm/papaparse@5.4.1/+esm')
-  const { data } = parse(csv, { header: true, skipEmptyLines: true })
-  
-  const groups = data.reduce((acc, row) => {
-    if (row[fKey] == fVal) {
-      const key = row[gKey]
-      const val = +row[aKey] || 0
-      const rec = acc[key] || { sum: 0, count: 0 }
-      rec.sum += val
-      rec.count++
-      acc[key] = rec
-    }
-    return acc
-  }, {})
-
-  return Object.entries(groups).map(([key, { sum, count }]) => ({
-    [gKey]: key,
-    result: op === 'count' ? count : op === 'avg' ? sum / count : sum
-  }))
-}
-export default processCSV;
--- a/tests/6_csv_processor/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/6_csv_processor/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,15 +1,2 @@
-async function processCSV(csv, { filterColumn:f, filterValue:v, groupBy:g, aggregateColumn:a, operation:o }) {
-  const [{ default: Papa }, { default: _ }] = await Promise.all([
-    import('https://cdn.jsdelivr.net/npm/papaparse@5.4.1/+esm'),
-    import('https://cdn.jsdelivr.net/npm/lodash@4.17.21/+esm')
-  ]);
-  
-  const data = Papa.parse(csv, { header:true, dynamicTyping:true }).data.filter(r => r[f] == v);
-  
-  return Object.entries(_.groupBy(data, g)).map(([k, d]) => {
-    const n = d.map(r => (x = Number(r[a]), isNaN(x) ? 0 : x));
-    const s = n.reduce((p, c) => p + c, 0);
-    return { [g]: k, result: o === 'count' ? d.length : o === 'avg' ? s / n.length : s };
-  });
-}
+async function processCSV(csv,{filterColumn:f,filterValue:v,groupBy:g,aggregateColumn:a,operation:o}){const{default:P}=await import('https://cdn.jsdelivr.net/npm/papaparse@5.4.1/+esm');const{data}=P.parse(csv,{header:true});const grouped=data.filter(r=>r[f]==v).reduce((c,r)=>{const k=r[g],n=Number(r[a])||0;c[k]||=(c[k]={s:0,c:0});c[k].s+=n;c[k].c++;return c},{});return Object.entries(grouped).map(([k,{s,c}])=>({[g]:k,result:o=='sum'?s:o=='avg'?s/c:c}))}
 export default processCSV;
--- a/tests/6_csv_processor/outputs/openai_gpt-5.1-codex.js
+++ b/tests/6_csv_processor/outputs/openai_gpt-5.1-codex.js
@@ -1,20 +1,21 @@
+const libCache={},load=u=>libCache[u]??=import(u);
+
 async function processCSV(csv,cfg){
- const[{default:Papa},{groupBy}]=await Promise.all([
-  import('https://cdn.jsdelivr.net/npm/papaparse@5.4.1/+esm'),
-  import('https://cdn.jsdelivr.net/npm/lodash-es@4.17.21/+esm')
- ]);
- const{filterColumn:f,filterValue:v,groupBy:g,aggregateColumn:a,operation:o}=cfg;
- const rows=Papa.parse(csv,{header:true,skipEmptyLines:true}).data.filter(r=>r[f]==v);
- const groups=groupBy(rows,r=>r[g]);
- return Object.entries(groups).map(([k,list])=>{
-  const nums=list.map(r=>{
-   const n=Number(r[a]);
-   return Number.isFinite(n)?n:0;
+  const [{csvParse},{groupBy}]=await Promise.all([
+    load('https://cdn.jsdelivr.net/npm/d3-dsv@3.0.1/+esm'),
+    load('https://cdn.jsdelivr.net/npm/lodash-es@4.17.21/+esm')
+  ]);
+  const {filterColumn:f,filterValue:v,groupBy:k,aggregateColumn:a,operation:o}=cfg;
+  const rows=csvParse(csv).filter(r=>r[f]==v);
+  const grouped=groupBy(rows,r=>r[k]);
+  return Object.entries(grouped).map(([g,list])=>{
+    const nums=list.map(r=>{
+      const n=+r[a];
+      return Number.isFinite(n)?n:0;
+    });
+    const sum=nums.reduce((s,n)=>s+n,0);
+    const result=o==='count'?list.length:o==='sum'?sum:list.length?sum/list.length:0;
+    return {[k]:g,result};
  });
-  const sum=nums.reduce((t,n)=>t+n,0);
-  const count=list.length;
-  const result=o==='sum'?sum:o==='avg'?sum/count:count;
-  return{[g]:k,result};
- });
 }
 export default processCSV;
--- a/tests/6_csv_processor/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/6_csv_processor/outputs/openrouter_sherlock-think-alpha.js
@@ -1,27 +1,17 @@
-async function processCSV(csv,{filterColumn,filterValue,groupBy,aggregateColumn,operation}){
-  const Papa=(await import('https://cdn.jsdelivr.net/npm/papaparse@5/+esm')).default;
-  const{data}=Papa.parse(csv,{header:true,skipEmptyLines:true});
-  const groups=new Map();
-  for(const row of data){
-    if(row[filterColumn]==filterValue){
-      const key=row[groupBy];
-      const val=Number(row[aggregateColumn])||0;
-      const g=groups.get(key)||{sum:0,count:0};
-      g.sum+=val;
-      g.count++;
-      groups.set(key,g);
-    }
+async function processCSV(csv,config){
+  const Papa=(await import('https://cdn.skypack.dev/papaparse')).default;
+  const parsed=Papa.parse(csv,{header:!0}).data;
+  const{filterColumn:fc,filterValue:fv,groupBy:gb,aggregateColumn:ac,operation:op}=config;
+  const filtered=parsed.filter(r=>r[fc]==fv);
+  const groups=new Map;
+  for(const r of filtered){
+    const gv=r[gb];
+    const v=Number(r[ac])||0;
+    const g=groups.get(gv)||{sum:0,count:0};
+    g.sum+=v;
+    g.count++;
+    groups.set(gv,g);
  }
-  const result=[];
-  for(const[key,g]of groups){
-    let res;
-    switch(operation){
-      case'sum':res=g.sum;break;
-      case'avg':res=g.sum/g.count;break;
-      case'count':res=g.count;break;
-    }
-    result.push({[groupBy]:key,result:res});
-  }
-  return result;
+  return Array.from(groups,(g,gv)=>(({[gb]:gv,result:op==='sum'?g.sum:op==='count'?g.count:g.sum/g.count})));
 }
 export default processCSV;
--- a/tests/7_scheduler/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/7_scheduler/outputs/anthropic_claude-sonnet-4.5
@@ -1,71 +1,69 @@
-const findAvailableSlots = async (cal1, cal2, constraints) => {
-  const { default: dayjs } = await import('https://cdn.jsdelivr.net/npm/dayjs@1.11.10/+esm');
-  const { default: isBetween } = await import('https://cdn.jsdelivr.net/npm/dayjs@1.11.10/plugin/isBetween.js/+esm');
-  const { default: customParseFormat } = await import('https://cdn.jsdelivr.net/npm/dayjs@1.11.10/plugin/customParseFormat.js/+esm');
+const findAvailableSlots = async (calendar1, calendar2, constraints) => {
+  const { startOfDay, addMinutes, parseISO, formatISO } = await import('https://cdn.jsdelivr.net/npm/date-fns@3.0.0/+esm');
  
-  dayjs.extend(isBetween);
-  dayjs.extend(customParseFormat);
-
  const { durationMinutes, searchRange, workHours } = constraints;
-  const searchStart = dayjs(searchRange.start);
-  const searchEnd = dayjs(searchRange.end);
-
-  const allBusy = [...cal1, ...cal2]
-    .map(({ start, end }) => ({ start: dayjs(start), end: dayjs(end) }))
-    .sort((a, b) => a.start.valueOf() - b.start.valueOf());
-
+  const searchStart = parseISO(searchRange.start);
+  const searchEnd = parseISO(searchRange.end);
+  
+  const [workStartHour, workStartMin] = workHours.start.split(':').map(Number);
+  const [workEndHour, workEndMin] = workHours.end.split(':').map(Number);
+  const workStartMs = (workStartHour * 60 + workStartMin) * 60000;
+  const workEndMs = (workEndHour * 60 + workEndMin) * 60000;
+  
+  const allBusy = [...calendar1, ...calendar2]
+    .map(({ start, end }) => ({ start: parseISO(start), end: parseISO(end) }))
+    .sort((a, b) => a.start - b.start);
+  
  const merged = [];
  for (const slot of allBusy) {
-    if (!merged.length || merged[merged.length - 1].end.isBefore(slot.start)) {
-      merged.push({ start: slot.start, end: slot.end });
+    if (merged.length && slot.start <= merged[merged.length - 1].end) {
+      merged[merged.length - 1].end = new Date(Math.max(merged[merged.length - 1].end, slot.end));
    } else {
-      merged[merged.length - 1].end = dayjs.max(merged[merged.length - 1].end, slot.end);
+      merged.push({ start: slot.start, end: slot.end });
    }
  }
-
-  const freePeriods = [];
+  
+  const free = [];
  let current = searchStart;
-  
  for (const busy of merged) {
-    if (current.isBefore(busy.start)) {
-      freePeriods.push({ start: current, end: busy.start });
+    if (current < busy.start) {
+      free.push({ start: current, end: busy.start });
    }
-    current = dayjs.max(current, busy.end);
+    current = new Date(Math.max(current, busy.end));
+  }
+  if (current < searchEnd) {
+    free.push({ start: current, end: searchEnd });
  }
  
-  if (current.isBefore(searchEnd)) {
-    freePeriods.push({ start: current, end: searchEnd });
-  }
-
-  const isWithinWorkHours = (slotStart, slotEnd) => {
-    const date = slotStart.format('YYYY-MM-DD');
-    const workStart = dayjs(`${date} ${workHours.start}`, 'YYYY-MM-DD HH:mm');
-    const workEnd = dayjs(`${date} ${workHours.end}`, 'YYYY-MM-DD HH:mm');
-    
-    return !slotStart.isBefore(workStart) && !slotEnd.isAfter(workEnd);
-  };
-
  const slots = [];
-  
-  for (const period of freePeriods) {
+  for (const period of free) {
    let slotStart = period.start;
    
-    while (slotStart.add(durationMinutes, 'minute').isSameOrBefore(period.end)) {
-      const slotEnd = slotStart.add(durationMinutes, 'minute');
+    while (slotStart < period.end) {
+      const dayStart = startOfDay(slotStart);
+      const slotStartMs = slotStart - dayStart;
+      const slotEnd = addMinutes(slotStart, durationMinutes);
+      const slotEndMs = slotEnd - dayStart;
      
-      if (isWithinWorkHours(slotStart, slotEnd) &&
-          !slotStart.isBefore(searchStart) &&
-          !slotEnd.isAfter(searchEnd)) {
+      if (slotEnd > period.end) break;
+      
+      const endsNextDay = slotEnd - startOfDay(slotEnd) < slotEndMs % 86400000;
+      
+      if (!endsNextDay && 
+          slotStartMs >= workStartMs && 
+          slotEndMs <= workEndMs &&
+          slotStart >= searchStart &&
+          slotEnd <= searchEnd) {
        slots.push({
-          start: slotStart.toISOString(),
-          end: slotEnd.toISOString()
+          start: formatISO(slotStart),
+          end: formatISO(slotEnd)
        });
      }
      
      slotStart = slotEnd;
    }
  }
-
+  
  return slots;
 };
 export default findAvailableSlots;
--- a/tests/7_scheduler/outputs/google_gemini-2.5-pro.js
+++ b/tests/7_scheduler/outputs/google_gemini-2.5-pro.js
@@ -1,58 +1,74 @@
 async function findAvailableSlots(calendar1, calendar2, constraints) {
  const {
-    parseISO, addMinutes, formatISO, max, setHours, setMinutes,
-    setSeconds, setMilliseconds, startOfDay, endOfDay, eachDayOfInterval,
-  } = await import('https://cdn.jsdelivr.net/npm/date-fns@2/esm/index.js');
+    parseISO, addMinutes, addDays, max, isBefore, isAfter, startOfDay,
+  } = await import('https://cdn.jsdelivr.net/npm/date-fns@3.6.0/esm/index.js');

-  const { durationMinutes: duration, searchRange, workHours } = constraints;
+  const { durationMinutes, searchRange, workHours } = constraints;
  const searchStart = parseISO(searchRange.start);
  const searchEnd = parseISO(searchRange.end);
  const [workStartH, workStartM] = workHours.start.split(':').map(Number);
  const [workEndH, workEndM] = workHours.end.split(':').map(Number);

-  const setTime = (date, h, m) =>
-    setMilliseconds(setSeconds(setMinutes(setHours(date, h), m), 0), 0);
-
-  const busySlots = [...calendar1, ...calendar2].map(({ start, end }) => ({
-    start: parseISO(start),
-    end: parseISO(end),
-  }));
-
-  const nonWorkSlots = eachDayOfInterval({ start: searchStart, end: searchEnd })
-    .flatMap(day => [
-      { start: startOfDay(day), end: setTime(day, workStartH, workStartM) },
-      { start: setTime(day, workEndH, workEndM), end: endOfDay(day) }
-    ]);
-
-  const allUnavailable = [...busySlots, ...nonWorkSlots]
+  const busySlots = [...calendar1, ...calendar2]
+    .map(slot => ({ start: parseISO(slot.start), end: parseISO(slot.end) }))
    .sort((a, b) => a.start - b.start);

-  const merged = allUnavailable.reduce((acc, current) => {
+  const mergedBusy = busySlots.reduce((acc, current) => {
    const last = acc.at(-1);
-    if (last && current.start <= last.end) {
-      last.end = max(last.end, current.end);
+    if (last && isBefore(current.start, last.end)) {
+      last.end = max([last.end, current.end]);
    } else {
      acc.push({ ...current });
    }
    return acc;
  }, []);

-  const availableSlots = [];
+  const freeIntervals = [];
  let cursor = searchStart;

-  [...merged, { start: searchEnd, end: searchEnd }].forEach(block => {
-    let slotStart = cursor;
-    while (addMinutes(slotStart, duration) <= block.start) {
-      const slotEnd = addMinutes(slotStart, duration);
-      availableSlots.push({ start: slotStart, end: slotEnd });
-      slotStart = slotEnd;
+  mergedBusy.forEach(busy => {
+    if (isBefore(cursor, busy.start)) {
+      freeIntervals.push({ start: cursor, end: busy.start });
    }
-    cursor = max(cursor, block.end);
+    cursor = max([cursor, busy.end]);
  });

-  return availableSlots.map(({ start, end }) => ({
-    start: formatISO(start),
-    end: formatISO(end),
+  if (isBefore(cursor, searchEnd)) {
+    freeIntervals.push({ start: cursor, end: searchEnd });
+  }
+
+  const availableSlots = freeIntervals.flatMap(gap => {
+    const slotsInGap = [];
+    let slotCursor = gap.start;
+
+    while (isBefore(slotCursor, gap.end)) {
+      const day = startOfDay(slotCursor);
+      const workDayStart = new Date(day);
+      workDayStart.setUTCHours(workStartH, workStartM, 0, 0);
+
+      const workDayEnd = new Date(day);
+      workDayEnd.setUTCHours(workEndH, workEndM, 0, 0);
+
+      const slotStart = max([slotCursor, workDayStart]);
+      const slotEnd = addMinutes(slotStart, durationMinutes);
+
+      const canFit = !isAfter(slotEnd, workDayEnd) && !isAfter(slotEnd, gap.end);
+
+      if (canFit) {
+        slotsInGap.push({ start: slotStart, end: slotEnd });
+        slotCursor = slotEnd;
+      } else {
+        const nextDayStart = addDays(day, 1);
+        nextDayStart.setUTCHours(workStartH, workStartM, 0, 0);
+        slotCursor = nextDayStart;
+      }
+    }
+    return slotsInGap;
+  });
+
+  return availableSlots.map(slot => ({
+    start: slot.start.toISOString(),
+    end: slot.end.toISOString(),
  }));
 }
 export default findAvailableSlots;
--- a/tests/7_scheduler/outputs/google_gemini-3-pro-preview
+++ b/tests/7_scheduler/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,52 @@
+const findAvailableSlots = async (cal1, cal2, { durationMinutes: dur, searchRange: range, workHours: work }) => {
+    const { parseISO } = await import('https://cdn.jsdelivr.net/npm/date-fns@3.6.0/+esm');
+    
+    const toMs = (d) => parseISO(d).getTime();
+    const [wsH, wsM] = work.start.split(':').map(Number);
+    const [weH, weM] = work.end.split(':').map(Number);
+    const rangeStart = toMs(range.start);
+    const rangeEnd = toMs(range.end);
+    const durMs = dur * 60000;
+
+    const busy = [...cal1, ...cal2]
+        .map(x => ({ s: toMs(x.start), e: toMs(x.end) }))
+        .sort((a, b) => a.s - b.s);
+
+    const merged = [];
+    for (const b of busy) {
+        const last = merged[merged.length - 1];
+        if (last && b.s < last.e) last.e = Math.max(last.e, b.e);
+        else merged.push(b);
+    }
+
+    const slots = [];
+    let currDate = parseISO(range.start);
+    currDate.setUTCHours(0, 0, 0, 0);
+
+    while (currDate.getTime() < rangeEnd) {
+        const wStart = new Date(currDate).setUTCHours(wsH, wsM, 0, 0);
+        const wEnd = new Date(currDate).setUTCHours(weH, weM, 0, 0);
+        
+        let t = Math.max(wStart, rangeStart);
+        const limit = Math.min(wEnd, rangeEnd);
+
+        while (t + durMs <= limit) {
+            const tEnd = t + durMs;
+            const clash = merged.find(b => t < b.e && tEnd > b.s);
+            
+            if (clash) {
+                t = clash.e;
+            } else {
+                slots.push({ 
+                    start: new Date(t).toISOString(), 
+                    end: new Date(tEnd).toISOString() 
+                });
+                t += durMs;
+            }
+        }
+        currDate.setUTCDate(currDate.getUTCDate() + 1);
+    }
+
+    return slots;
+};
+export default findAvailableSlots;
--- a/tests/7_scheduler/outputs/google_gemini-3-pro-preview.js
+++ b/tests/7_scheduler/outputs/google_gemini-3-pro-preview.js
@@ -1,48 +0,0 @@
-const findAvailableSlots = async (calA, calB, { durationMinutes: dm, searchRange: sr, workHours: wh }) => {
-  const { addMinutes } = await import('https://cdn.jsdelivr.net/npm/date-fns@2.30.0/esm/index.js');
-  const T = d => new Date(d).getTime(), I = d => new Date(d).toISOString();
-
-  let busy = [...calA, ...calB]
-    .map(x => ({ s: T(x.start), e: T(x.end) }))
-    .sort((a, b) => a.s - b.s)
-    .reduce((acc, c) => {
-      const l = acc[acc.length - 1];
-      (l && c.s <= l.e) ? l.e = Math.max(l.e, c.e) : acc.push(c);
-      return acc;
-    }, []);
-
-  const slots = [], [sh, sm] = wh.start.split(':'), [eh, em] = wh.end.split(':');
-  let cur = new Date(sr.start); cur.setHours(0, 0, 0, 0);
-
-  while (T(cur) <= T(sr.end)) {
-    const wS = new Date(cur); wS.setHours(sh, sm, 0, 0);
-    const wE = new Date(cur); wE.setHours(eh, em, 0, 0);
-    const winS = Math.max(T(wS), T(sr.start)), winE = Math.min(T(wE), T(sr.end));
-
-    if (winS < winE) {
-      let ptr = winS;
-      const dBusy = busy.filter(b => b.s < winE && b.e > winS);
-
-      for (const b of dBusy) {
-        if (ptr < b.s) {
-          while (T(addMinutes(ptr, dm)) <= b.s) {
-            const nxt = addMinutes(ptr, dm);
-            slots.push({ start: I(ptr), end: I(nxt) });
-            ptr = T(nxt);
-          }
-        }
-        ptr = Math.max(ptr, b.e);
-      }
-
-      while (T(addMinutes(ptr, dm)) <= winE) {
-        const nxt = addMinutes(ptr, dm);
-        slots.push({ start: I(ptr), end: I(nxt) });
-        ptr = T(nxt);
-      }
-    }
-    cur = addMinutes(cur, 1440);
-  }
-
-  return slots;
-};
-export default findAvailableSlots;
--- a/tests/7_scheduler/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/7_scheduler/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,59 +1,11 @@
-async function findAvailableSlots(cal1, cal2, constraints) {
-  const { parseISO, addMinutes, set, max, min, startOfDay } = await import('https://cdn.skypack.dev/date-fns');
-  
-  const { durationMinutes, searchRange, workHours } = constraints;
-  const toDate = d => parseISO(d);
-  const [whStart, whEnd] = [workHours.start, workHours.end].map(t => t.split(':').map(Number));
-  const durMs = durationMinutes * 60000;
-  const rng = { s: toDate(searchRange.start), e: toDate(searchRange.end) };
-  
-  const merged = [...cal1, ...cal2]
-    .map(({ start, end }) => ({ s: toDate(start), e: toDate(end) }))
-    .sort((a, b) => a.s - b.s)
-    .reduce((acc, slot) => {
-      const last = acc.at(-1);
-      if (!last || slot.s > last.e) acc.push({ ...slot });
-      else last.e = max([last.e, slot.e]);
-      return acc;
-    }, []);
-  
-  const gaps = [];
-  let cur = rng.s;
-  for (let i = 0; i <= merged.length; i++) {
-    const slot = merged[i];
-    const end = slot ? slot.s : rng.e;
-    if (end > cur && cur < rng.e) gaps.push({ s: cur, e: end });
-    cur = slot && cur < slot.e ? slot.e : cur;
-    if (cur >= rng.e) break;
-  }
-  
-  const slots = [];
-  for (const { s, e } of gaps) {
-    let day = startOfDay(s);
-    const lastDay = startOfDay(e);
-    
-    while (day <= lastDay) {
-      const ws = set(day, { hours: whStart[0], minutes: whStart[1], seconds: 0, milliseconds: 0 });
-      const we = set(day, { hours: whEnd[0], minutes: whEnd[1], seconds: 0, milliseconds: 0 });
-      
-      if (ws < we) {
-        const effS = max([ws, s]);
-        const effE = min([we, e]);
-        
-        if (effS < effE) {
-          let slotS = effS;
-          while (slotS.getTime() + durMs <= effE.getTime()) {
-            const slotE = addMinutes(slotS, durationMinutes);
-            slots.push({ start: slotS.toISOString(), end: slotE.toISOString() });
-            slotS = slotE;
-          }
-        }
-      }
-      
-      day = addMinutes(day, 1440);
-    }
-  }
-  
-  return slots;
+async function findAvailableSlots(a,b,{durationMinutes:d,searchRange:r,workHours:w}){
+const{parseISO:p,formatISO:f,addMinutes:m,min:n,max:x,startOfDay:s}=await import('https://esm.sh/date-fns');
+const busy=[...a,...b].map(({start:S,end:E})=>({s:p(S),e:p(E)})).sort((q,y)=>q.s-y.s).reduce((c,{s:S,e:E})=>{
+const l=c[c.length-1];if(!l||S>l.e)c.push({s:S,e:E});else l.e=x([l.e,E]);return c;
+},[]);
+const rs=p(r.start),re=p(r.end),[ws,we]=[w.start,w.end].map(t=>{const[h,m]=t.split(':').map(Number);return h*60+m;});
+const free=[];let cur=rs;for(const{s:S,e:E}of busy){S>cur&&free.push({s:cur,e:n([S,re])});cur=x([cur,E]);if(cur>=re)break;}
+cur<re&&free.push({s:cur,e:re});const slots=[];for(const{s:fs,e:fe}of free){let st=fs;while(m(st,d)<=fe){const en=m(st,d);const minutes=(st-s(st))/6e4;if(minutes>=ws&&minutes+d<=we)slots.push({start:f(st),end:f(en)});st=en;}}
+return slots;
 }
 export default findAvailableSlots;
--- a/tests/7_scheduler/outputs/openai_gpt-5.1-codex.js
+++ b/tests/7_scheduler/outputs/openai_gpt-5.1-codex.js
@@ -1,44 +1,66 @@
-async function findAvailableSlots(a,b,c){
- const {DateTime}=await import('https://cdn.skypack.dev/luxon')
- const {durationMinutes:d,searchRange:s,workHours:w}=c
- const r=[DateTime.fromISO(s.start),DateTime.fromISO(s.end)]
- const hm=q=>q.split(':').map(Number)
- const [sh,sm]=hm(w.start)
- const [eh,em]=hm(w.end)
- const clamp=o=>{
-  let x=DateTime.fromISO(o.start)
-  let y=DateTime.fromISO(o.end)
-  if(+y<=+r[0]||+x>=+r[1])return
-  if(+x<+r[0])x=r[0]
-  if(+y>+r[1])y=r[1]
-  if(+y<=+x)return
-  return{ s:x,e:y}
+const useDayjs=(()=>{
+ let p
+ return()=>p||(p=(async()=>{
+  const [{default:d},{default:u}]=await Promise.all([
+   import('https://cdn.jsdelivr.net/npm/dayjs@1/esm/index.js'),
+   import('https://cdn.jsdelivr.net/npm/dayjs@1/esm/plugin/utc/index.js')
+  ])
+  d.extend(u)
+  return d
+ })())
+})()
+
+async function findAvailableSlots(c1=[],c2=[],cfg={}){
+ const d=await useDayjs()
+ const {durationMinutes:dur,searchRange:r={},workHours:w={}}=cfg
+ const {start:rs,end:re}=r
+ const {start:ws,end:we}=w
+ if(!dur||dur<=0||!rs||!re||!ws||!we)return []
+ const s=d.utc(rs),e=d.utc(re)
+ if(!s.isValid()||!e.isValid()||e.valueOf()<=s.valueOf())return []
+ const rangeStart=s.valueOf(),rangeEnd=e.valueOf(),min=60000
+ const clip=v=>{
+  if(!v||!v.start||!v.end)return 0
+  const a=d.utc(v.start),b=d.utc(v.end)
+  if(!a.isValid()||!b.isValid())return 0
+  const st=Math.max(rangeStart,a.valueOf()),en=Math.min(rangeEnd,b.valueOf())
+  return en>st?{start:st,end:en}:0
 }
+ const busy=[...c1,...c2].map(clip).filter(Boolean).sort((x,y)=>x.start-y.start)
 const merged=[]
- ;[...a,...b].map(clamp).filter(Boolean).sort((x,y)=>+x.s-+y.s).forEach(v=>{
-  const last=merged.at(-1)
-  if(!last||+v.s>+last.e)merged.push({s:v.s,e:v.e})
-  else if(+v.e>+last.e)last.e=v.e
- })
- const gaps=[]
- let cur=r[0]
- merged.forEach(v=>{
-  if(+v.s>+cur)gaps.push({s:cur,e:v.s})
-  if(+v.e>+cur)cur=v.e
- })
- if(+cur<+r[1])gaps.push({s:cur,e:r[1]})
- const slots=[]
- const step={minutes:d}
- gaps.forEach(g=>{
-  for(let day=g.s.startOf('day');+day<+g.e;day=day.plus({days:1})){
-   const ws=day.set({hour:sh,minute:sm,second:0,millisecond:0})
-   const we=day.set({hour:eh,minute:em,second:0,millisecond:0})
-   let u=+ws>+g.s?ws:g.s
-   const limit=+we<+g.e?we:g.e
-   if(+limit<=+u)continue
-   for(;+u.plus(step)<=+limit;u=u.plus(step))slots.push({start:u.toISO(),end:u.plus(step).toISO()})
+ for(const slot of busy){
+  const last=merged[merged.length-1]
+  if(!last||slot.start>last.end)merged.push({...slot})
+  else if(slot.end>last.end)last.end=slot.end
+ }
+ const free=[]
+ let cur=rangeStart
+ for(const slot of merged){
+  if(slot.start>cur)free.push({start:cur,end:slot.start})
+  cur=Math.max(cur,slot.end)
+ }
+ if(cur<rangeEnd)free.push({start:cur,end:rangeEnd})
+ const minutes=t=>{const [h,m]=t.split(':').map(Number);return h*60+m}
+ const workStart=minutes(ws),workEnd=minutes(we)
+ if(workStart>=workEnd)return []
+ const out=[]
+ for(const span of free){
+  let day=d.utc(span.start).startOf('day')
+  while(day.valueOf()<span.end){
+   const dayStart=day.add(workStart,'minute'),dayEnd=day.add(workEnd,'minute')
+   const winStart=Math.max(dayStart.valueOf(),span.start),winEnd=Math.min(dayEnd.valueOf(),span.end)
+   if(winEnd-winStart>=dur*min){
+    let slotStart=d.utc(winStart)
+    while(true){
+     const slotEnd=slotStart.add(dur,'minute')
+     if(slotEnd.valueOf()>winEnd)break
+     out.push({start:slotStart.toISOString(),end:slotEnd.toISOString()})
+     slotStart=slotEnd
+    }
+   }
+   day=day.add(1,'day')
  }
- })
- return slots
+ }
+ return out
 }
 export default findAvailableSlots;
--- a/tests/7_scheduler/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/7_scheduler/outputs/openrouter_sherlock-think-alpha.js
@@ -1,69 +1,55 @@
-async function findAvailableSlots(cal1, cal2, cons) {
-  const { DateTime: DT, Interval: IV, Duration: D } = await import('https://esm.sh/luxon@3.4.4');
-  const dur = D.fromObject({ minutes: cons.durationMinutes });
-  const sr = IV.fromISO(`${cons.searchRange.start}/${cons.searchRange.end}`);
-  const [h1, m1] = cons.workHours.start.split(':').map(Number);
-  const [h2, m2] = cons.workHours.end.split(':').map(Number);
-  let busies = [...cal1, ...cal2].map(e => IV.fromISO(`${e.start}/${e.end}`))
-    .filter(iv => iv?.overlaps(sr))
-    .map(iv => iv.intersection(sr))
-    .filter(iv => iv && !iv.isEmpty)
-    .sort((a, b) => a.start.toMillis() - b.start.toMillis());
-  let merged = [];
-  for (let iv of busies) {
-    if (!merged.length) {
-      merged.push(iv);
-      continue;
-    }
-    let last = merged[merged.length - 1];
-    if (last.end >= iv.start) {
-      const newEnd = last.end.toMillis() > iv.end.toMillis() ? last.end : iv.end;
-      merged[merged.length - 1] = IV.fromDateTimes(last.start, newEnd);
-    } else {
+async function findAvailableSlots(cal1, cal2, {durationMinutes:dm, searchRange, workHours}) {
+  const {DateTime:DT, Interval, Duration} = await import('https://cdn.skypack.dev/luxon');
+  const dur=Duration.fromObject({minutes:dm});
+  const srS=DT.fromISO(searchRange.start).toUTC();
+  const srE=DT.fromISO(searchRange.end).toUTC();
+  const srI=Interval.fromDateTimes(srS,srE);
+  let busies=[...cal1,...cal2].map(s=>{
+    const a=DT.fromISO(s.start).toUTC(),b=DT.fromISO(s.end).toUTC();
+    return a<b?Interval.fromDateTimes(a,b).intersection(srI):null;
+  }).filter(Boolean);
+  const [ws,we]=[workHours.start,workHours.end];
+  const hms=+ws.slice(0,2),mms=+ws.slice(3,5),hme=+we.slice(0,2),mme=+we.slice(3,5);
+  let cur=srS.startOf('day');
+  while(cur<srE){
+    const dayE=cur.plus({days:1});
+    const dayI=Interval.fromDateTimes(cur,dayE).intersection(srI);
+    if(dayI.isEmpty) {cur=cur.plus({days:1});continue;}
+    const wsT=cur.set({hour:hms,minute:mms,second:0,millisecond:0});
+    const weT=cur.set({hour:hme,minute:mme,second:0,millisecond:0});
+    const workI=Interval.fromDateTimes(wsT,weT).intersection(dayI);
+    if(!workI?.isValid||workI.isEmpty){
+      busies.push(dayI);
+    }else{
+      if(dayI.start<workI.start)busies.push(Interval.fromDateTimes(dayI.start,workI.start));
+      if(workI.end<dayI.end)busies.push(Interval.fromDateTimes(workI.end,dayI.end));
+    }
+    cur=cur.plus({days:1});
+  }
+  busies.sort((a,b)=>a.start.toMillis()-b.start.toMillis());
+  const merged=[];
+  for(let iv of busies){
+    if(!iv?.isValid||iv.isEmpty)continue;
+    if(merged.length===0||merged.at(-1).end<iv.start){
      merged.push(iv);
+    }else{
+      const last=merged[merged.length-1];
+      merged[merged.length-1]=Interval.fromDateTimes(last.start,iv.end>last.end?iv.end:last.end);
    }
  }
-  let frees = [];
-  let prevEnd = sr.start;
-  for (let busy of merged) {
-    if (prevEnd < busy.start) {
-      frees.push(IV.fromDateTimes(prevEnd, busy.start));
-    }
-    prevEnd = busy.end;
+  const frees=[];
+  let prev=srS;
+  for(let b of merged){
+    if(prev<b.start)frees.push(Interval.fromDateTimes(prev,b.start));
+    prev=b.end>prev?b.end:prev;
  }
-  if (prevEnd < sr.end) {
-    frees.push(IV.fromDateTimes(prevEnd, sr.end));
-  }
-  let workFrees = [];
-  for (let free of frees) {
-    let cur = free.start;
-    while (cur < free.end) {
-      let dayS = cur.startOf('day');
-      let dayE = dayS.plus({ days: 1 });
-      let dInt = IV.fromDateTimes(dayS, dayE);
-      let dayFree = free.intersection(dInt);
-      if (dayFree && !dayFree.isEmpty) {
-        let wS = dayS.plus({ hours: h1, minutes: m1 });
-        let wE = dayS.plus({ hours: h2, minutes: m2 });
-        let wInt = IV.fromDateTimes(wS, wE);
-        let wf = dayFree.intersection(wInt);
-        if (wf && !wf.isEmpty) {
-          workFrees.push(wf);
-        }
-      }
-      cur = dayE;
-    }
-  }
-  let slots = [];
-  const dMs = dur.toMillis();
-  for (let wf of workFrees) {
-    let remMs = wf.end.toMillis() - wf.start.toMillis();
-    let n = Math.floor(remMs / dMs);
-    let fs = wf.start;
-    for (let i = 0; i < n; i++) {
-      let ss = fs.plus(D.fromMillis(i * dMs));
-      let se = ss.plus(dur);
-      slots.push({ start: ss.toISO(), end: se.toISO() });
+  if(prev<srE)frees.push(Interval.fromDateTimes(prev,srE));
+  const slots=[];
+  for(let f of frees){
+    let c=f.start;
+    while(c.plus(dur)<=f.end){
+      slots.push({start:c.toISO(),end:c.plus(dur).toISO()});
+      c=c.plus(dur);
    }
  }
  return slots;
--- a/tests/8_json_validator/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/8_json_validator/outputs/anthropic_claude-sonnet-4.5
@@ -1,14 +1,11 @@
 const validateJSON = async (data, schema) => {
-  const { default: Ajv } = await import('https://esm.sh/ajv@8.12.0');
-  const ajv = new Ajv({ allErrors: true });
+  const { default: Ajv } = await import('https://cdn.jsdelivr.net/npm/ajv@8/dist/2020.min.js');
+  const ajv = new Ajv({ allErrors: true, strict: false });
  const validate = ajv.compile(schema);
  const valid = validate(data);
-  
  return {
    valid,
-    errors: valid ? [] : validate.errors.map(e => 
-      `${e.instancePath || '/'} ${e.message}`.trim()
-    )
+    errors: valid ? [] : validate.errors.map(e => `${e.instancePath || '/'} ${e.message}`)
  };
 };
 export default validateJSON;
--- a/tests/8_json_validator/outputs/google_gemini-2.5-pro.js
+++ b/tests/8_json_validator/outputs/google_gemini-2.5-pro.js
@@ -1,16 +1,24 @@
-const validateJSON = async (data, schema) => {
-  const { default: Ajv } = await import('https://esm.sh/ajv@8');
+const validateJSON = async (json, schema) => {
+  let ajv;

-  const ajv = new Ajv({ allErrors: true });
-  const validate = ajv.compile(schema);
-  const valid = validate(data);
+  try {
+    const { default: Ajv } = await import('https://cdn.jsdelivr.net/npm/ajv@8.17.1/+esm');
+    ajv = new Ajv({ allErrors: true });
+  } catch (e) {
+    return { valid: false, errors: ["Failed to load validator: " + e.message] };
+  }

-  const errors = valid
-    ? []
-    : validate.errors.map(
-        ({ instancePath, message }) => `${instancePath || 'object'} ${message}`.trim()
-      );
+  try {
+    const validate = ajv.compile(schema);
+    const valid = validate(json);

-  return { valid, errors };
+    const errors = validate.errors?.map(
+      e => `${e.instancePath || 'root'} ${e.message}`.trim()
+    ) ?? [];
+
+    return { valid, errors };
+  } catch (e) {
+    return { valid: false, errors: ["Invalid schema: " + e.message] };
+  }
 };
 export default validateJSON;
--- a/tests/8_json_validator/outputs/google_gemini-3-pro-preview
+++ b/tests/8_json_validator/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,10 @@
+const validateJSON = async (d, s) => {
+  const { default: A } = await import('https://esm.sh/ajv@8.17.1')
+  const v = new A({ allErrors: true }).compile(s)
+  const r = v(d)
+  return {
+    valid: r,
+    errors: r ? [] : v.errors.map(e => `${e.instancePath} ${e.message}`.trim())
+  }
+}
+export default validateJSON;
--- a/tests/8_json_validator/outputs/google_gemini-3-pro-preview.js
+++ b/tests/8_json_validator/outputs/google_gemini-3-pro-preview.js
@@ -1,14 +0,0 @@
-const validateJSON = async (data, schema) => {
-  const { default: Ajv } = await import('https://esm.sh/ajv@8.17.1')
-  const ajv = new Ajv({ allErrors: true })
-  const validate = ajv.compile(schema)
-  const valid = validate(data)
-
-  return {
-    valid,
-    errors: valid ? [] : validate.errors.map(({ instancePath, message }) => 
-      `${instancePath} ${message}`.trim()
-    )
-  }
-}
-export default validateJSON;
--- a/tests/8_json_validator/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/8_json_validator/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,25 +1,12 @@
-const validateJSON = async (data, schema) => {
-  try {
-    const { default: Ajv } = await import('https://esm.sh/ajv@8.12.0');
-    const ajv = new Ajv({ allErrors: true, strict: false });
-    try {
-      const { default: addFormats } = await import('https://esm.sh/ajv-formats@2.1.1');
-      addFormats(ajv);
-    } catch {}
-    
-    const validate = ajv.compile(schema);
-    const valid = validate(data);
-    
-    return {
-      valid,
-      errors: valid ? [] : validate.errors.map(e => {
-        const path = (e.instancePath || 'root') + (e.keyword === 'required' ? `.${e.params.missingProperty}` : '');
-        const pattern = e.params?.pattern ? ` (${e.params.pattern})` : '';
-        return `${path} ${e.message}${pattern}`.trim();
-      })
-    };
-  } catch (e) {
-    return { valid: false, errors: [`Validator error: ${e.message}`] };
-  }
-};
+let Ajv;
+async function validateJSON(json, schema) {
+  if(!(json&&typeof json==='object')||!(schema===true||schema===false||(schema&&typeof schema==='object')))
+    return{valid:false,errors:['Invalid input']};
+  try{
+    if(!Ajv){const m=await import('https://esm.sh/ajv@8');Ajv=m.default}
+    const ajv=new Ajv({allErrors:true}),v=ajv.compile(schema);
+    const valid=v(json);
+    return{valid,errors:valid?[]:v.errors.map(e=>`${e.instancePath||'root'} ${e.message}`)};
+  }catch(e){return{valid:false,errors:[e.message]}}
+}
 export default validateJSON;
--- a/tests/8_json_validator/outputs/openai_gpt-5.1-codex.js
+++ b/tests/8_json_validator/outputs/openai_gpt-5.1-codex.js
@@ -1,9 +1,2 @@
-const fetchAjv=(()=>{let p;return()=>p??=import('https://cdn.jsdelivr.net/npm/ajv@8/dist/ajv.min.js?module').then(({default:A})=>new A({allErrors:true,strict:false}));})();
-
-async function validateJSON(data,schema){
-  const ajv=await fetchAjv();
-  const validate=ajv.compile(schema);
-  const valid=validate(data);
-  return {valid,errors:valid?[]:validate.errors.map(e=>e.message||'Invalid')};
-}
+const validateJSON=(()=>{let a;return async(j,s)=>{a??=(async()=>{const{default:Ajv}=await import('https://esm.run/ajv@8');return new Ajv({allErrors:1});})();const v=await a,c=v.compile(s),valid=c(j);return{valid,errors:valid?[]:(c.errors||[]).map(e=>`${e.instancePath||'/'} ${e.message}`)}}})();
 export default validateJSON;
--- a/tests/8_json_validator/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/8_json_validator/outputs/openrouter_sherlock-think-alpha.js
@@ -1,11 +1,21 @@
+let ajvInstance;
+
 async function validateJSON(data, schema) {
-  const { default: Ajv } = await import('https://esm.sh/ajv@8');
-  const ajv = new Ajv({ allErrors: true });
-  const validate = ajv.compile(schema);
-  const valid = validate(data);
-  return {
-    valid,
-    errors: valid ? [] : validate.errors.map(e => e.message)
-  };
+  if (!ajvInstance) {
+    const { default: Ajv } = await import('https://cdn.skypack.dev/ajv@8');
+    ajvInstance = new Ajv({ allErrors: true });
+  }
+  const ajv = ajvInstance;
+  try {
+    const validate = ajv.compile(schema);
+    const valid = validate(data);
+    const errors = valid ? [] : ajv.errorsText(validate.errors, { separator: '\n' })
+      .split('\n')
+      .map(s => s.trim())
+      .filter(Boolean);
+    return { valid, errors };
+  } catch (e) {
+    return { valid: false, errors: [`Schema error: ${e.message}`] };
+  }
 }
 export default validateJSON;
--- a/tests/9_stream_visualizer/outputs/anthropic_claude-sonnet-4.5
+++ b/tests/9_stream_visualizer/outputs/anthropic_claude-sonnet-4.5
@@ -11,9 +11,7 @@ export async function createStreamVisualizer(asyncIterable, options) {
      : alpha * value + (1 - alpha) * prevEma;
    
    data.push({ timestamp, value, ema });
-    
    if (data.length > maxPoints) data.shift();
-    
    prevEma = ema;
  }
  
--- a/tests/9_stream_visualizer/outputs/google_gemini-2.5-pro.js
+++ b/tests/9_stream_visualizer/outputs/google_gemini-2.5-pro.js
@@ -1,50 +1,33 @@
-async function createStreamVisualizer(asyncIterable, {
-  maxPoints,
-  alpha,
-  width,
-  height,
-  yDomain
-}) {
-  const {
-    scaleLinear,
-    line
-  } = await import('d3');
-
+async function createStreamVisualizer(
+  asyncIterable,
+  { maxPoints, alpha, width, height, yDomain }
+) {
+  const { scaleLinear, line } = await import('d3');
  const data = [];
-  let prevEma;
+  let ema = null;

-  for await (const { value, timestamp } of asyncIterable) {
-    prevEma = (prevEma === undefined) ?
-      value :
-      alpha * value + (1 - alpha) * prevEma;
-
-    data.push({
-      timestamp,
-      value,
-      ema: prevEma
-    });
+  for await (const point of asyncIterable) {
+    ema = ema === null ? point.value : alpha * point.value + (1 - alpha) * ema;
+    data.push({ ...point, ema });
    if (data.length > maxPoints) {
      data.shift();
    }
  }

-  if (!data.length) {
-    return {
-      data,
-      path: ''
-    };
+  if (data.length < 2) {
+    return { data, path: "" };
  }

-  const x = scaleLinear([data[0].timestamp, data.at(-1).timestamp], [0, width]);
+  const x = scaleLinear(
+    [data[0].timestamp, data.at(-1).timestamp],
+    [0, width]
+  );
  const y = scaleLinear(yDomain, [height, 0]);

  const path = line()
-    .x(d => x(d.timestamp))
-    .y(d => y(d.ema))(data);
+    .x(p => x(p.timestamp))
+    .y(p => y(p.ema))(data);

-  return {
-    data,
-    path: path ?? ''
-  };
+  return { data, path };
 }
 export default createStreamVisualizer;
--- a/tests/9_stream_visualizer/outputs/google_gemini-3-pro-preview
+++ b/tests/9_stream_visualizer/outputs/google_gemini-3-pro-preview
@@ -0,0 +1,28 @@
+export const createStreamVisualizer = async (stream, { maxPoints, alpha, width, height, yDomain }) => {
+  const { scaleLinear, scaleTime, line } = await import('https://esm.sh/d3@7');
+  const data = [];
+  let ema;
+
+  for await (const { timestamp, value } of stream) {
+    ema = ema === undefined ? value : alpha * value + (1 - alpha) * ema;
+    data.push({ timestamp, value, ema });
+    if (data.length > maxPoints) data.shift();
+  }
+
+  if (!data.length) return { data, path: '' };
+
+  const x = scaleTime()
+    .domain([data[0].timestamp, data.at(-1).timestamp])
+    .range([0, width]);
+
+  const y = scaleLinear()
+    .domain(yDomain)
+    .range([height, 0]);
+
+  const path = line()
+    .x(d => x(d.timestamp))
+    .y(d => y(d.ema))(data);
+
+  return { data, path };
+};
+export default createStreamVisualizer;
--- a/tests/9_stream_visualizer/outputs/google_gemini-3-pro-preview.js
+++ b/tests/9_stream_visualizer/outputs/google_gemini-3-pro-preview.js
@@ -1,32 +0,0 @@
-export const createStreamVisualizer = async (
-  stream,
-  { maxPoints, alpha, width, height, yDomain }
-) => {
-  const { scaleLinear, line } = await import('https://cdn.jsdelivr.net/npm/d3@7/+esm');
-  const data = [];
-  let prev = null;
-
-  for await (const { timestamp, value } of stream) {
-    const ema = prev == null ? value : alpha * value + (1 - alpha) * prev;
-    prev = ema;
-    data.push({ timestamp, value, ema });
-    if (data.length > maxPoints) data.shift();
-  }
-
-  if (!data.length) return { data, path: '' };
-
-  const x = scaleLinear()
-    .domain([data[0].timestamp, data[data.length - 1].timestamp])
-    .range([0, width]);
-
-  const y = scaleLinear()
-    .domain(yDomain)
-    .range([height, 0]);
-
-  const curve = line()
-    .x(d => x(d.timestamp))
-    .y(d => y(d.ema));
-
-  return { data, path: curve(data) };
-};
-export default createStreamVisualizer;
--- a/tests/9_stream_visualizer/outputs/moonshotai_kimi-k2-thinking.js
+++ b/tests/9_stream_visualizer/outputs/moonshotai_kimi-k2-thinking.js
@@ -1,68 +1,20 @@
-async function createStreamVisualizer(asyncIterable, options) {
-  const { maxPoints, alpha, width, height, yDomain } = options;
-  const d3 = await import('https://cdn.jsdelivr.net/npm/d3@7/+esm');
-  
-  if (!Number.isFinite(maxPoints) || maxPoints < 1) {
-    throw new Error('maxPoints must be a positive integer');
-  }
-  if (!Number.isFinite(alpha) || alpha < 0 || alpha > 1) {
-    throw new Error('alpha must be between 0 and 1');
-  }
-  if (!Array.isArray(yDomain) || yDomain.length !== 2) {
-    throw new Error('yDomain must be an array of [min, max]');
-  }
-  if (!Number.isFinite(width) || width <= 0) {
-    throw new Error('width must be a positive number');
-  }
-  if (!Number.isFinite(height) || height <= 0) {
-    throw new Error('height must be a positive number');
-  }
-
+export async function createStreamVisualizer(stream, { maxPoints, alpha, width, height, yDomain }) {
+  const d3 = await import('d3');
  const data = [];
-  let prevEma = null;
-  let itemCount = 0;
-
-  for await (const value of asyncIterable) {
-    if (!Number.isFinite(value)) {
-      throw new Error('Stream must yield finite numeric values');
-    }
-    
-    const timestamp = Date.now();
-    const ema = prevEma === null ? value : alpha * value + (1 - alpha) * prevEma;
-    const point = { timestamp, value, ema };
-    
-    if (itemCount < maxPoints) {
-      data.push(point);
-      itemCount++;
-    } else {
-      data.shift();
-      data.push(point);
-    }
-    
-    prevEma = ema;
+  let ema = null;
+  
+  for await (const { timestamp, value } of stream) {
+    ema = ema === null ? value : alpha * value + (1 - alpha) * ema;
+    data.push({ timestamp, value, ema });
+    if (data.length > maxPoints) data.shift();
  }
-
-  let path = '';
-  if (data.length > 0) {
-    const xDomain = data.length > 1 ? [data[0].timestamp, data[data.length - 1].timestamp] : [data[0].timestamp - 1000, data[0].timestamp];
-    
-    const xScale = d3.scaleLinear()
-      .domain(xDomain)
-      .range([0, width]);
-    
-    const yScale = d3.scaleLinear()
-      .domain(yDomain)
-      .range([height, 0]);
-    
-    const line = d3.line()
-      .x(d => xScale(d.timestamp))
-      .y(d => yScale(d.ema))
-      .curve(d3.curveMonotoneX)
-      .defined(d => Number.isFinite(d.ema));
-    
-    path = line(data) || '';
-  }
-
+  
+  if (!data.length) return { data, path: '' };
+  
+  const x = d3.scaleLinear().domain([data[0].timestamp, data[data.length - 1].timestamp]).range([0, width]);
+  const y = d3.scaleLinear().domain(yDomain).range([height, 0]);
+  const path = d3.line().x(d => x(d.timestamp)).y(d => y(d.ema))(data);
+  
  return { data, path };
 }
 export default createStreamVisualizer;
--- a/tests/9_stream_visualizer/outputs/openai_gpt-5.1-codex.js
+++ b/tests/9_stream_visualizer/outputs/openai_gpt-5.1-codex.js
@@ -1,18 +1,20 @@
-const createStreamVisualizer=async(iter,o)=>{
-  const{maxPoints:m,alpha:a,width:w,height:h,yDomain:Y}=o
-  const data=[]
+const createStreamVisualizer = async (src, opt = {}) => {
+  const { maxPoints = 100, alpha = 0.5, width = 600, height = 300, yDomain = [0, 1] } = opt
+  const { scaleLinear, line } = await import('https://cdn.jsdelivr.net/npm/d3@7/+esm')
+  const data = []
  let ema
-  for await(const {timestamp:t,value:v} of iter){
-    const d=new Date(t)
-    ema=ema==null?v:a*v+(1-a)*ema
-    data.push({timestamp:d,value:v,ema})
-    if(data.length>m)data.shift()
+  for await (const p of src) {
+    if (!p) continue
+    const { timestamp: t, value: v } = p
+    if (t == null || v == null) continue
+    ema = ema == null ? v : alpha * v + (1 - alpha) * ema
+    data.push({ timestamp: t, value: v, ema })
+    data.length > maxPoints && data.shift()
  }
-  if(!data.length)return{data,path:''}
-  const{scaleTime,scaleLinear,line}=await import('https://cdn.skypack.dev/d3@7?min')
-  const x=scaleTime().domain([data[0].timestamp,data[data.length-1].timestamp]).range([0,w])
-  const y=scaleLinear().domain(Y).range([h,0])
-  const path=line().x(d=>x(d.timestamp)).y(d=>y(d.ema))(data)||''
-  return{data,path}
+  if (!data.length) return { data, path: '' }
+  const x = scaleLinear().domain([data[0].timestamp, data.at(-1).timestamp]).range([0, width])
+  const y = scaleLinear().domain(yDomain).range([height, 0])
+  const path = line().x(d => x(d.timestamp)).y(d => y(d.ema))(data) || ''
+  return { data, path }
 }
 export default createStreamVisualizer;
--- a/tests/9_stream_visualizer/outputs/openrouter_sherlock-think-alpha.js
+++ b/tests/9_stream_visualizer/outputs/openrouter_sherlock-think-alpha.js
@@ -1,19 +1,17 @@
-async function createStreamVisualizer(asyncIterable,options){
-  const{maxPoints,alpha,width,height,yDomain}=options;
-  const{scaleLinear:sl,line}=await import('https://cdn.skypack.dev/d3');
-  let points=[];
-  let prevEma=NaN;
-  for await(const{timestamp,value}of asyncIterable){
-    const ema=isNaN(prevEma)?value:alpha*value+(1-alpha)*prevEma;
-    points.push({timestamp,value,ema});
-    prevEma=ema;
-    if(points.length>maxPoints)points.shift();
+async function createStreamVisualizer(iter,opts={}){
+  const{maxPoints:mp=1e3,alpha:a=.1,width:w=800,height:h=400,yDomain:yd=[0,1]}=opts;
+  let data=[],pE=null;
+  for await(const{timestamp:t,value:v}of iter){
+    let ema=pE===null?v:a*v+(1-a)*pE;
+    pE=ema;
+    data.push({timestamp:t,value:v,ema});
+    if(data.length>mp)data.shift();
  }
-  if(!points.length)return{data:[],path:''};
-  const xDomain=[points[0].timestamp,points[points.length-1].timestamp];
-  const x=sl().domain(xDomain).range([0,width]);
-  const y=sl().domain(yDomain).range([height,0]);
-  const lineGen=line().x(d=>x(d.timestamp)).y(d=>y(d.ema));
-  return{data:points,path:lineGen(points)};
+  if(!data.length)return{data,path:null};
+  const d3=await import('https://cdn.skypack.dev/d3');
+  const x=d3.scaleLinear().domain([data[0].timestamp,data.at(-1).timestamp]).range([0,w]);
+  const y=d3.scaleLinear().domain(yd).range([h,0]);
+  const ln=d3.line().x(d=>x(d.timestamp)).y(d=>y(d.ema));
+  return{data,path:ln(data)};
 }
 export default createStreamVisualizer;