File size: 86,488 Bytes
57a963b 7a3b3da 57a963b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 | <!doctype html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>DarwinX: Evolving Agent Harnesses Through Natural Selection</title>
<meta name="description" content="DarwinX treats agent self-evolution as selection over a population of harnesses with the model frozen. 84.7% on Terminal-Bench 2.1, 68.3% on held-out TerminalWorld, 93.0% audit-clean on WebArena-Infinity, 84.2% zero-shot transfer to SWE-bench Verified." />
<meta property="og:type" content="article" />
<meta property="og:title" content="DarwinX: Evolving Agent Harnesses Through Natural Selection" />
<meta property="og:description" content="Population selection over agent harnesses with the model frozen. Terminal-Bench 2.1 frontier at 84.7%." />
<meta property="og:image" content="assets/teaser.png" />
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:title" content="DarwinX: Evolving Agent Harnesses Through Natural Selection" />
<meta name="twitter:description" content="A frozen model need not be a fixed agent: harness selection turns evaluation compute into durable capability." />
<meta name="twitter:image" content="assets/teaser.png" />
<link rel="icon" href="data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 32 32'><rect width='32' height='32' rx='7' fill='%23032d60'/><path d='M12.6 14.8 19.4 10.6M12.6 17.2 19.4 21.4' stroke='%23fff' stroke-width='1.7'/><circle cx='10' cy='16' r='3.3' fill='%230176d3'/><circle cx='22' cy='9' r='3.3' fill='%23fff'/><circle cx='22' cy='23' r='3.3' fill='%23fff'/></svg>" />
<style>
:root {
--navy: #032d60;
--blue: #0176d3;
--ink: #17202c;
--muted: #5a6b7d;
--line: #dfe6ee;
--tint: #f2f7fc;
--bg: #ffffff;
--maxw: 920px;
--navh: 48px;
}
* { box-sizing: border-box; }
html { -webkit-text-size-adjust: 100%; scroll-behavior: smooth; }
body {
margin: 0;
background: var(--bg);
color: var(--ink);
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "Helvetica Neue", Arial, sans-serif;
font-size: 17px;
line-height: 1.65;
}
.wrap { max-width: var(--maxw); margin: 0 auto; padding: 0 22px; }
a { color: var(--blue); text-decoration: none; }
a:hover { text-decoration: underline; }
/* ---------- sticky nav ---------- */
nav.top {
position: sticky; top: 0; z-index: 50;
background: rgba(255,255,255,0.94);
backdrop-filter: saturate(180%) blur(8px);
border-bottom: 1px solid var(--line);
height: var(--navh);
}
nav.top .wrap { display: flex; align-items: center; gap: 4px; height: 100%; overflow-x: auto; }
nav.top .brand { font-weight: 700; color: var(--navy); margin-right: 12px; white-space: nowrap; }
nav.top a {
color: var(--muted); font-size: 0.85rem; padding: 5px 9px; border-radius: 5px; white-space: nowrap;
}
nav.top a:hover { background: var(--tint); color: var(--navy); text-decoration: none; }
nav.top .spacer { flex: 1 1 auto; }
nav.top a.cta { color: #fff; background: var(--navy); font-weight: 500; }
nav.top a.cta:hover { background: #06407f; color: #fff; }
section[id] { scroll-margin-top: calc(var(--navh) + 14px); }
/* ---------- header ---------- */
header { padding: 52px 0 34px; border-bottom: 1px solid var(--line); }
h1 {
font-size: 2.05rem; line-height: 1.25; letter-spacing: -0.015em;
color: var(--navy); margin: 0 0 20px; font-weight: 700;
}
h1 .sub { display: block; font-size: 1.32rem; font-weight: 600; margin-top: 6px; color: var(--ink); }
.authors { font-size: 1.02rem; margin: 0 0 8px; }
.authors b { font-weight: 600; }
sup { font-size: 0.7em; color: var(--blue); }
.affil, .legend { color: var(--muted); font-size: 0.9rem; margin: 4px 0 0; }
.legend { margin-top: 12px; font-size: 0.84rem; }
.links { display: flex; flex-wrap: wrap; gap: 10px; margin: 26px 0 0; }
.btn {
display: inline-flex; align-items: center; padding: 9px 17px; border-radius: 7px;
background: var(--navy); color: #fff; font-size: 0.94rem; font-weight: 500;
border: 1px solid var(--navy); cursor: pointer;
}
.btn:hover { background: #06407f; text-decoration: none; }
.btn.ghost { background: #fff; color: var(--navy); }
.btn.ghost:hover { background: var(--tint); }
/* ---------- sections ---------- */
section { padding: 42px 0; border-bottom: 1px solid var(--line); }
section:last-of-type { border-bottom: 0; }
h2 {
font-size: 0.83rem; text-transform: uppercase; letter-spacing: 0.1em;
color: var(--muted); font-weight: 700; margin: 0 0 20px;
}
h3 { font-size: 1.16rem; color: var(--navy); margin: 0 0 4px; font-weight: 600; letter-spacing: -0.01em; }
h3 .tag { font-weight: 400; color: var(--muted); font-size: 0.8rem; letter-spacing: 0.02em; }
h4 { font-size: 1rem; color: var(--ink); margin: 0 0 5px; font-weight: 600; }
p { margin: 0 0 15px; }
p:last-child { margin-bottom: 0; }
figure { margin: 0; }
figure img { width: 100%; height: auto; display: block; border: 1px solid var(--line); border-radius: 8px; background: #fff; }
figure a.zoom { display: block; position: relative; }
figure a.zoom:hover img { border-color: var(--blue); }
figure a.zoom::after {
content: "\2317"; position: absolute; right: 9px; bottom: 9px;
font-size: 0.85rem; line-height: 1; padding: 5px 7px; border-radius: 5px;
background: rgba(3,45,96,0.82); color: #fff; opacity: 0; transition: opacity .15s;
}
figure a.zoom:hover::after { opacity: 1; }
figcaption { color: var(--muted); font-size: 0.87rem; margin-top: 10px; line-height: 1.55; }
.figrow { display: grid; grid-template-columns: 1fr 1fr; gap: 20px; align-items: start; }
.abstract { font-size: 1.02rem; }
/* ---------- headline cards ---------- */
.cards { display: grid; grid-template-columns: repeat(2, 1fr); gap: 14px; }
.card { border: 1px solid var(--line); border-radius: 8px; padding: 17px 18px; background: var(--tint); }
.card .label { font-size: 0.79rem; color: var(--muted); text-transform: uppercase; letter-spacing: 0.06em; font-weight: 600; }
.card .big { font-size: 1.75rem; font-weight: 700; color: var(--navy); margin: 7px 0 2px; letter-spacing: -0.02em; }
.card .big .delta { font-size: 0.9rem; font-weight: 600; color: var(--blue); margin-left: 7px; letter-spacing: 0; }
.card .note { font-size: 0.85rem; color: var(--muted); line-height: 1.5; }
/* ---------- per-benchmark blocks ---------- */
.bench { padding-top: 34px; margin-top: 34px; border-top: 1px solid var(--line); }
.bench:first-of-type { padding-top: 0; margin-top: 0; border-top: 0; }
.bench > .lede { color: var(--muted); font-size: 0.93rem; margin: 0 0 20px; }
.block { margin-top: 28px; }
table { width: 100%; border-collapse: collapse; font-size: 0.94rem; }
caption { text-align: left; color: var(--muted); font-size: 0.87rem; padding-bottom: 10px; line-height: 1.5; }
th, td { text-align: left; padding: 8px 9px; border-bottom: 1px solid var(--line); vertical-align: top; }
th { font-size: 0.78rem; text-transform: uppercase; letter-spacing: 0.04em; color: var(--muted); border-bottom: 1.5px solid #c9d4e0; font-weight: 700; }
td.num, th.num { text-align: right; font-variant-numeric: tabular-nums; }
tr.ours td { background: var(--tint); font-weight: 600; color: var(--navy); }
tr.base td { color: var(--muted); font-style: italic; }
tr.total td { border-top: 1.5px solid #c9d4e0; font-weight: 700; }
td.delta { color: var(--blue); font-weight: 600; }
.tnote { color: var(--muted); font-size: 0.85rem; margin-top: 11px; line-height: 1.55; }
.scroller { overflow-x: auto; }
.scroller table { min-width: 640px; }
code {
font-family: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
font-size: 0.85em; background: #eef3f8; padding: 1px 5px; border-radius: 4px; color: var(--navy);
white-space: nowrap;
}
ul.mech { list-style: none; padding: 0; margin: 0; }
ul.mech li { padding: 0 0 18px 0; }
ul.mech li:last-child { padding-bottom: 0; }
ul.lim { padding-left: 20px; margin: 0; }
ul.lim li { padding-bottom: 12px; }
ul.lim li:last-child { padding-bottom: 0; }
ul.lim b { color: var(--navy); }
/* ---------- interactive widgets ---------- */
.widget { border: 1px solid var(--line); border-radius: 9px; padding: 18px; background: #fcfdfe; }
.widget .wtop {
display: flex; flex-wrap: wrap; align-items: center; gap: 8px 14px;
padding-bottom: 15px; margin-bottom: 15px; border-bottom: 1px solid var(--line);
}
.widget .wtop .wlabel { font-size: 0.78rem; text-transform: uppercase; letter-spacing: 0.06em; color: var(--muted); font-weight: 700; }
.chips { display: flex; flex-wrap: wrap; gap: 7px; }
.chip {
font: inherit; font-size: 0.85rem; padding: 5px 12px; border-radius: 20px;
border: 1px solid var(--line); background: #fff; color: var(--muted); cursor: pointer;
transition: background .12s, border-color .12s, color .12s;
}
.chip:hover { border-color: #b9c7d6; }
.chip[aria-pressed="true"] { background: var(--navy); border-color: var(--navy); color: #fff; }
.chip .cnt { opacity: 0.62; font-variant-numeric: tabular-nums; }
.toggle { display: inline-flex; align-items: center; gap: 7px; font-size: 0.86rem; color: var(--muted); cursor: pointer; }
.toggle input { accent-color: var(--navy); width: 15px; height: 15px; }
.readout { display: flex; flex-wrap: wrap; gap: 10px; margin-bottom: 16px; }
.ro {
flex: 1 1 150px; border: 1px solid var(--line); border-radius: 7px;
padding: 11px 13px; background: #fff;
}
.ro .k { font-size: 0.75rem; text-transform: uppercase; letter-spacing: 0.05em; color: var(--muted); font-weight: 600; }
.ro .v { font-size: 1.35rem; font-weight: 700; color: var(--navy); font-variant-numeric: tabular-nums; letter-spacing: -0.01em; }
.ro.hi { background: var(--tint); border-color: #c9dcf0; }
.ro .v small { font-size: 0.62em; font-weight: 600; color: var(--muted); }
.tgrid { display: grid; grid-template-columns: repeat(auto-fill, minmax(26px, 1fr)); gap: 5px; }
.tcell {
position: relative; aspect-ratio: 1; border-radius: 4px; background: #e6ecf2;
border: 1px solid transparent; cursor: default; transition: background .12s;
}
.tcell.on { background: var(--blue); }
.tcell.gain { box-shadow: 0 0 0 2px #e8a33d; z-index: 1; }
.tcell.loss { box-shadow: 0 0 0 2px #d05a4e; z-index: 1; }
.tcell:hover { border-color: var(--navy); }
.wlegend { display: flex; flex-wrap: wrap; gap: 14px; margin-top: 14px; font-size: 0.8rem; color: var(--muted); }
.wlegend span { display: inline-flex; align-items: center; gap: 6px; }
.sw { width: 13px; height: 13px; border-radius: 3px; display: inline-block; }
#tip {
position: fixed; z-index: 100; pointer-events: none; opacity: 0;
background: var(--navy); color: #fff; font-size: 0.78rem; line-height: 1.45;
padding: 7px 10px; border-radius: 6px; max-width: 260px;
transition: opacity .1s; box-shadow: 0 4px 14px rgba(3,45,96,0.25);
}
#tip b { color: #9ecbf5; font-weight: 600; }
/* cluster bars */
.cbars { display: flex; flex-direction: column; gap: 9px; }
.cbar { display: grid; grid-template-columns: minmax(120px, 1.5fr) 3fr auto; gap: 12px; align-items: center; }
.cbar .cname { font-size: 0.86rem; color: var(--ink); }
.cbar .cname .cn { color: var(--muted); font-size: 0.78rem; }
.cbar .track { position: relative; height: 26px; background: #f0f4f8; border-radius: 4px; overflow: hidden; }
.cbar .fill { position: absolute; top: 0; height: 100%; border-radius: 4px; transition: width .35s ease, left .35s ease; }
.cbar .fill.b { background: #c6d3e0; }
.cbar .fill.e { background: var(--blue); opacity: 0.92; }
.cbar .cd { font-size: 0.84rem; font-weight: 700; font-variant-numeric: tabular-nums; min-width: 52px; text-align: right; }
.cbar .cd.up { color: var(--blue); }
.cbar .cd.down { color: #b3453a; }
/* line chart */
.chart { width: 100%; height: auto; display: block; overflow: visible; }
.chart .ax { stroke: #ccd7e2; stroke-width: 1; }
.chart .gl { stroke: #eef2f6; stroke-width: 1; }
.chart .lbl { font-size: 10.5px; fill: var(--muted); }
.chart .best { fill: none; stroke: var(--blue); stroke-width: 2.2; }
.chart .dot { fill: #b9c7d6; }
.chart .dot:hover, .chart .dot.act { fill: var(--navy); }
.chart .hit { fill: transparent; cursor: pointer; }
/* headline panels */
.panels { display: grid; grid-template-columns: repeat(4, 1fr); gap: 12px; }
.panel { text-align: center; }
.panel .pt { font-size: 0.95rem; font-weight: 700; color: var(--navy); }
.panel .ps { font-size: 0.74rem; color: var(--muted); line-height: 1.4; margin-bottom: 6px; }
.panel svg { width: 100%; height: auto; display: block; overflow: visible; }
.panel .bar { transition: y .35s ease, height .35s ease; }
.panel .bar.b { fill: #c6d3e0; }
.panel .bar.p { fill: #8fb4e8; }
.panel .bar.e { fill: var(--blue); }
.panel .dl { font-size: 11.5px; font-weight: 700; fill: var(--blue); }
.panel .tick { font-size: 9.5px; fill: var(--muted); }
.panel .axl { stroke: #ccd7e2; stroke-width: 1; }
.panel .foot { font-size: 0.72rem; color: var(--muted); margin-top: 4px; }
/* horizontal bar rows (specialists, compute) */
.hrows { display: flex; flex-direction: column; gap: 8px; }
.hrow { display: grid; grid-template-columns: minmax(96px, 1.1fr) 3fr auto; gap: 12px; align-items: center; }
.hrow .hn { font-size: 0.86rem; }
.hrow .ht { position: relative; height: 24px; background: #f0f4f8; border-radius: 4px; }
.hrow .hf { position: absolute; left: 0; top: 0; height: 100%; border-radius: 4px; transition: width .35s ease; }
.hrow .hv { font-size: 0.84rem; font-weight: 700; font-variant-numeric: tabular-nums; min-width: 68px; text-align: right; color: var(--navy); }
.hrow.dim .hn, .hrow.dim .hv { color: var(--muted); }
.refline { position: absolute; top: -3px; bottom: -3px; width: 0; border-left: 2px dashed #e4572e; }
.reftag { position: absolute; top: -17px; transform: translateX(-50%); font-size: 0.7rem; color: #e4572e; white-space: nowrap; }
/* stacked composition */
.stack { display: flex; height: 30px; border-radius: 5px; overflow: hidden; background: #f0f4f8; }
.stack i { display: block; height: 100%; transition: width .35s ease; }
.srow { margin-bottom: 16px; }
.srow .sh { display: flex; justify-content: space-between; font-size: 0.82rem; color: var(--muted); margin-bottom: 5px; }
.srow .sh b { color: var(--navy); font-size: 0.95rem; }
.keys { display: flex; flex-wrap: wrap; gap: 6px 14px; margin-top: 12px; font-size: 0.78rem; color: var(--muted); }
.keys span { display: inline-flex; align-items: center; gap: 6px; }
/* dumbbell */
.dumb { display: flex; flex-direction: column; gap: 7px; }
.drow { display: grid; grid-template-columns: minmax(84px, 1fr) 4fr; gap: 12px; align-items: center; }
.drow.tot { padding-bottom: 7px; border-bottom: 1px solid var(--line); }
.drow.tot .dn { font-weight: 700; color: var(--navy); }
.drow .dn { font-size: 0.82rem; }
.drow .dt { position: relative; height: 30px;
background-image: repeating-linear-gradient(to right, #edf1f5 0 1px, transparent 1px 25%); }
.drow.dax { align-items: end; }
.drow.dax .dt { height: 16px; background: none; }
.drow.dax .dt span { position: absolute; bottom: 0; transform: translateX(-50%);
font-size: 0.7rem; color: var(--muted); }
.drow .seg { position: absolute; height: 3px; border-radius: 2px; }
.drow .pt2 { position: absolute; width: 9px; height: 9px; border-radius: 50%; transform: translate(-50%, -50%); }
.drow .hollow { background: #fff; }
.nojs { display: none; }
.nojs img { width: 100%; height: auto; border: 1px solid var(--line); border-radius: 8px; }
th.sortable { cursor: pointer; user-select: none; white-space: nowrap; }
th.sortable:hover { color: var(--navy); }
th.sortable::after { content: " \2195"; opacity: 0.35; font-size: 0.9em; }
th.sortable.asc::after { content: " \2191"; opacity: 1; color: var(--blue); }
th.sortable.desc::after { content: " \2193"; opacity: 1; color: var(--blue); }
.bibhead { display: flex; align-items: center; justify-content: space-between; gap: 12px; margin-bottom: 12px; }
.bibhead h2 { margin: 0; }
#copybtn {
font: inherit; font-size: 0.84rem; padding: 6px 13px; border-radius: 6px;
border: 1px solid var(--line); background: #fff; color: var(--navy); cursor: pointer;
}
#copybtn:hover { background: var(--tint); border-color: #c9d4e0; }
pre {
background: #f7f9fb; border: 1px solid var(--line); border-radius: 8px;
padding: 16px; overflow-x: auto; font-size: 0.85rem; line-height: 1.55;
font-family: ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
color: #223; margin: 0;
}
footer { padding: 30px 0 60px; color: var(--muted); font-size: 0.88rem; }
@media (max-width: 700px) {
body { font-size: 16px; }
h1 { font-size: 1.6rem; }
header { padding: 32px 0 26px; }
.cards, .figrow { grid-template-columns: 1fr; }
nav.top .hideable { display: none; }
.panels { grid-template-columns: repeat(2, 1fr); gap: 18px 10px; }
.hrow { grid-template-columns: minmax(72px, 1fr) 2fr auto; gap: 8px; }
.hrow .hn, .drow .dn { font-size: 0.78rem; }
.hrow .hv { min-width: 54px; font-size: 0.78rem; }
.drow { grid-template-columns: minmax(66px, 1fr) 3fr; gap: 8px; }
}
</style>
</head>
<body>
<noscript><style>.widget{display:none}.nojs{display:block}</style></noscript>
<nav class="top">
<div class="wrap">
<span class="brand">DarwinX</span>
<a href="#abstract">Abstract</a>
<a href="#method">Method</a>
<a href="#results">Results</a>
<a href="#benchmarks" class="hideable">Benchmarks</a>
<a href="#ablation" class="hideable">Ablation</a>
<a href="#limitations" class="hideable">Limitations</a>
<a href="#bibtex" class="hideable">BibTeX</a>
<span class="spacer"></span>
<a class="cta" href="https://arxiv.org/abs/2608.07545">arXiv</a>
</div>
</nav>
<header>
<div class="wrap">
<h1>DarwinX<span class="sub">Evolving Agent Harnesses Through Natural Selection</span></h1>
<p class="authors">
<b>Yifan Zhang</b><sup>◦1</sup>,
<b>Yutong Dai</b><sup>◦1</sup>,
Juntao Tan<sup>*1</sup>,
Luyu Yang<sup>*1</sup>,
Rishi Mullur<sup>1</sup>,
Thai Hoang<sup>1</sup>,
Zhiyuan Hu<sup>1</sup>,
James Zhu<sup>†2</sup>,
Phil Mui<sup>†2</sup>,
Silvio Savarese<sup>†1</sup>,
Ran Xu<sup>†1</sup>,
Zeyuan Chen<sup>†1</sup>
</p>
<p class="affil"><sup>1</sup>Salesforce AI Research <sup>2</sup>Salesforce Agentforce</p>
<p class="legend"><sup>◦</sup>First authors <sup>*</sup>Core authors <sup>†</sup>Senior authors</p>
<div class="links">
<a class="btn" href="https://arxiv.org/abs/2608.07545">arXiv:2608.07545</a>
<a class="btn ghost" href="https://arxiv.org/pdf/2608.07545">PDF</a>
<a class="btn ghost" href="https://huggingface.co/papers/2608.07545">Hugging Face</a>
<a class="btn ghost" href="mailto:yifan.zhang@salesforce.com?subject=DarwinX%20code%20request">Code available on request</a>
</div>
</div>
</header>
<section>
<div class="wrap">
<figure>
<a class="zoom" href="assets/full/teaser.png" target="_blank" rel="noopener">
<img src="assets/teaser_schematic.png" alt="Variants are evolved from a base agent, survivors are kept and merged, pruned lineages are dropped, all with the model frozen." />
</a>
<figcaption>
<b>With the base model frozen, evolving the harness alone matches or beats the strongest prior
agent on four benchmarks.</b> Variants survive on measured fitness (avg@<i>k</i>, no gold solutions)
and complementary survivors are merged. Click for the paper's full figure.
</figcaption>
</figure>
<div class="block">
<div class="widget" id="hl-widget">
<div class="wtop">
<span class="wlabel">y-axis</span>
<div class="chips" id="hl-chips"></div>
<span style="flex:1"></span>
<label class="toggle"><input type="checkbox" id="hl-prior" checked /> best prior agent</label>
</div>
<div class="panels" id="hl-panels"></div>
<div class="wlegend" id="hl-legend"></div>
<p class="tnote" id="hl-note"></p>
</div>
<noscript class="nojs"><img src="assets/teaser.png" alt="Four-benchmark summary: base vs. evolved vs. best prior agent." /></noscript>
</div>
</div>
</section>
<section id="abstract">
<div class="wrap">
<h2>Abstract</h2>
<div class="abstract">
<p>
An LLM agent's capability depends not only on model weights but on its <i>harness</i>: prompts,
tools, skills, and control flow. Self-improvement loops already edit harnesses, yet single-lineage
search is path-dependent and local wins often regress other tasks. We introduce <b>DarwinX</b>, which
treats self-evolution as <b>selection over a population of harnesses</b> with the model frozen: a
preserve-and-extend contract admits only variants that extend coverage without regressing, an archive
keeps alternative lineages for recombination, and failure-, teacher-, and self-derived evidence share
one edit interface. Fitness comes from each benchmark's own verifier: no gold solutions, no
hand-picked winners.
</p>
<p>
Across four benchmarks that progressively separate the evolution signal from the test, one loop adds
about 17 points on average: Terminal-Bench 2.1 rises +7.7 to 83.2% on a matched base and to the
verified frontier at <b>84.7%</b> on a stronger one; TerminalWorld's held-out split reaches
<b>68.3%</b>, ahead of every off-the-shelf agent; WebArena-Infinity real-task pass@1 rises from 43.5%
to <b>93.0%</b> audit-clean; and a Terminal-Bench 2.1 harness transfers unchanged to SWE-bench
Verified. What evolves is general agent competence, not benchmark-specific patches, so it survives
changes of task, verifier, and base model. A frozen model need not be a fixed agent: harness selection
turns evaluation compute into durable capability.
</p>
</div>
</div>
</section>
<section id="method">
<div class="wrap">
<h2>Method</h2>
<figure>
<a class="zoom" href="assets/full/overview.png" target="_blank" rel="noopener">
<img src="assets/overview.png" alt="Overview of the DarwinX loop." loading="lazy" />
</a>
</figure>
<ul class="mech" style="margin-top:26px">
<li>
<h4>Selection over a population, not one lineage</h4>
<p>A run maintains an archive shaped as a tree, where each node is a harness snapshot with its edit
delta, per-task scores, trial evidence, and distilled lessons. Nothing is discarded: a variant that
loses overall may still hold the one edit that, combined with another branch's, unlocks a task
neither solves alone. Parallel branches target different capability clusters, so the archive grows
specialists with different solved-task signatures.</p>
</li>
<li>
<h4>A preserve-and-extend contract</h4>
<p>A child is admitted only when the measured evidence shows it solved something new
<i>without breaking what it already solved</i>, which bounds what a win is allowed to cost
elsewhere. When variants solve complementary tasks, their additive edits are merged, and the merged
child is kept only if it covers the union of its parents' wins. Fitness is the benchmark's own
verifier at avg@<i>k</i> — no gold solutions and no hand-picked winners.</p>
</li>
<li>
<h4>One edit interface for three signals</h4>
<p>Failure-derived diagnosis, teacher-derived demonstrations, and self-derived rollout contrast are
all converted into harness edits rather than model-weight updates. The base model never changes, so
every gain reported here is a statement about the harness alone.</p>
</li>
</ul>
<figure style="margin-top:30px">
<a class="zoom" href="assets/full/population.png" target="_blank" rel="noopener">
<img src="assets/population.png" alt="DarwinX per-generation operators: mutation loop, variant classification, and the merge operator." loading="lazy" />
</a>
<figcaption>
<b>DarwinX's per-generation operators.</b> <b>Left:</b> the mutation loop and the three learning
signals that drive it. <b>Middle:</b> variants classified by how their solved set changes, where those
preserving inherited solves stay eligible for recombination while the rest contribute only distilled
lessons. <b>Right:</b> the merge operator and its acceptance criterion.
</figcaption>
</figure>
</div>
</section>
<section id="results">
<div class="wrap">
<h2>Results at a glance</h2>
<div class="cards">
<div class="card">
<div class="label">Terminal-Bench 2.1 · avg@5</div>
<div class="big">84.7%<span class="delta">+7.7 on matched base</span></div>
<div class="note">Verified-leaderboard frontier on a frozen GPT-5.6 Sol at <i>medium</i> effort.
On GPT-5.5, 75.5% → 83.2%.</div>
</div>
<div class="card">
<div class="label">TerminalWorld · held-out split</div>
<div class="big">68.3%<span class="delta">+7.3</span></div>
<div class="note">28/41 held-out tasks on a frozen Opus 4.8, above every off-the-shelf agent
we measured.</div>
</div>
<div class="card">
<div class="label">WebArena-Infinity · audit-clean pass@1</div>
<div class="big">93.0%<span class="delta">+49.5 pp</span></div>
<div class="note">1,260 real tasks, evolved on synthetic intents only. Base 43.5%; best same-model
baseline 86.1%.</div>
</div>
<div class="card">
<div class="label">SWE-bench Verified · zero-shot transfer</div>
<div class="big">84.2%<span class="delta">+3.4</span></div>
<div class="note">421/500 official pass@1 from the Terminal-Bench 2.1 harness run
<i>unchanged</i>, with no SWE-V feedback.</div>
</div>
</div>
<p class="tnote">
The four benchmarks progressively separate the evolution signal from the test: in-domain
(Terminal-Bench 2.1), held-out task split (TerminalWorld), synthetic → real intent
shift (WebArena-Infinity), and zero-shot cross-benchmark transfer (SWE-bench Verified).
</p>
</div>
</section>
<section id="benchmarks">
<div class="wrap">
<h2>Benchmark detail</h2>
<!-- ============ TB 2.1 ============ -->
<div class="bench">
<h3>Terminal-Bench 2.1 <span class="tag">— in-domain evolution</span></h3>
<p class="lede">89 tasks. Evolution and evaluation share the task distribution; the model is frozen and
only the harness changes.</p>
<table class="sortable">
<caption>Terminal-Bench 2.1, avg@5 under the strict leaderboard rule (binary avg@5,
<i>k</i>=5, errored trials score zero). Both DarwinX rows are leaderboard submissions, reported
before the leaderboard's uniform reward-hacking pass.</caption>
<thead>
<tr><th>Agent</th><th>Model / effort</th><th class="num">avg@5</th></tr>
</thead>
<tbody>
<tr class="ours"><td>Monet (DarwinX)</td><td>GPT-5.6 Sol / medium</td><td class="num">84.7 ± 1.2</td></tr>
<tr><td>Claude Code</td><td>Fable 5 / xhigh</td><td class="num">83.8 ± 1.2</td></tr>
<tr class="ours"><td>Monet (DarwinX)</td><td>GPT-5.5 / high</td><td class="num">83.2 ± 1.2</td></tr>
<tr><td>Codex</td><td>GPT-5.5 / xhigh</td><td class="num">83.1 ± 1.1</td></tr>
<tr><td>OpenAI reference</td><td>GPT-5.6 Sol / medium</td><td class="num">81.8</td></tr>
<tr><td>Terminus 2</td><td>GPT-5.5 / xhigh</td><td class="num">78.0 ± 1.2</td></tr>
<tr class="base"><td>Monet (base)</td><td>GPT-5.5 / default</td><td class="num">75.5 ± 3.5</td></tr>
</tbody>
</table>
<p class="tnote">
At 84.7% DarwinX matches or exceeds the verified leader (Claude Code + Fable 5, 83.8% at
<i>xhigh</i>) while running at a <i>lower</i> effort setting, and adds +2.9 points over OpenAI's own
native single-agent GPT-5.6 Sol at the same medium effort. Against a neutral harness on the
<i>same</i> GPT-5.5 base (Terminus 2, 78.0%), the pure harness gain is +5.2 points — so the
improvement is the harness, not the model or the effort setting.
</p>
<div class="block">
<h4>Where evolution helps</h4>
<p class="tnote" style="margin-top:0;margin-bottom:14px">
Per-cluster avg@5 on a frozen GPT-5.5. Sort by any column and hover a bar for exact rates.
</p>
<div class="widget" id="cl-widget">
<div class="wtop">
<span class="wlabel">Sort by</span>
<div class="chips" id="cl-chips"></div>
<span class="spacer" style="flex:1"></span>
<label class="toggle"><input type="checkbox" id="cl-base" checked /> show base</label>
</div>
<div class="cbars" id="cl-bars"></div>
<div class="wlegend">
<span><i class="sw" style="background:#c6d3e0"></i> Monet (base)</span>
<span><i class="sw" style="background:var(--blue)"></i> Monet (DarwinX)</span>
</div>
</div>
<noscript class="nojs"><img src="assets/tb21_category_gain.png" alt="Per-cluster Terminal-Bench 2.1 gain." /></noscript>
<p class="tnote">
Gains concentrate where a frozen base has the most headroom: <i>ML & scientific computing</i>
(+14.8, the largest cluster at 23 tasks) and <i>data & databases</i> (+13.8). Clusters where the
base is already strong barely move, and none regresses beyond the per-task noise band. Paired
protocol over 88 tasks, which excludes one task with no paired evolved trial; the headline 83.2%
uses the stricter 89-task leaderboard rule above.
</p>
</div>
<div class="block">
<h4>The gain is the harness, not compute</h4>
<p class="tnote" style="margin-top:0;margin-bottom:14px">
Per-task compute, median over clean attempts.
</p>
<div class="widget" id="cp-widget">
<div class="wtop">
<span class="wlabel">Metric</span>
<div class="chips" id="cp-chips"></div>
</div>
<div class="hrows" id="cp-rows"></div>
<div class="wlegend">
<span><i class="sw" style="background:#c6d3e0"></i> Monet (base)</span>
<span><i class="sw" style="background:var(--blue)"></i> Monet (DarwinX)</span>
</div>
<p class="tnote" id="cp-note"></p>
</div>
<noscript class="nojs"><img src="assets/tb21_compute.png" alt="Per-task compute, newly solved vs. already solved." /></noscript>
</div>
</div>
<!-- ============ TerminalWorld ============ -->
<div class="bench">
<h3>TerminalWorld <span class="tag">— held-out task split</span></h3>
<p class="lede">Evolve on 94 verifier-scored training tasks, then freeze the harness and evaluate on 41
disjoint held-out tasks. Single-attempt pass@1, no retries and no best-of-<i>k</i>, so no held-out
reward can feed back into the archive.</p>
<table class="sortable">
<caption>TerminalWorld held-out pass@1 (41 tasks, single attempt, frozen base models, identical
verifier and infrastructure).</caption>
<thead>
<tr><th>Agent</th><th>Model</th><th class="num">pass@1</th></tr>
</thead>
<tbody>
<tr class="ours"><td>Monet (DarwinX)</td><td>Opus 4.8</td><td class="num">68.3%</td></tr>
<tr><td>Claude Code</td><td>Opus 4.8</td><td class="num">65.9%</td></tr>
<tr class="base"><td>Monet (base)</td><td>Opus 4.8</td><td class="num">61.0%</td></tr>
<tr><td>Terminus-2</td><td>GPT-5.5</td><td class="num">61.0%</td></tr>
<tr><td>Terminus-2</td><td>Opus 4.8</td><td class="num">58.5%</td></tr>
<tr><td>Monet (DarwinX)</td><td>GPT-5.5</td><td class="num">56.1%</td></tr>
<tr><td>Codex</td><td>GPT-5.5</td><td class="num">51.2%</td></tr>
<tr class="base"><td>Monet (base)</td><td>GPT-5.5</td><td class="num">48.8%</td></tr>
</tbody>
</table>
<div class="figrow block">
<div>
<div class="widget" id="sp-widget">
<div class="wtop">
<span class="wlabel">x-axis</span>
<div class="chips" id="sp-chips"></div>
<span style="flex:1"></span>
<label class="toggle"><input type="checkbox" id="sp-counts" /> task counts</label>
</div>
<div class="hrows" id="sp-rows"></div>
<p class="tnote" id="sp-note"></p>
</div>
<noscript class="nojs"><img src="assets/tw_heldout_specialists.png" alt="Held-out pass@1 of four specialists and their merge." /></noscript>
<p class="tnote">
Held-out generalization reflects <b>archive diversity</b>: four evolved specialists each solve a
different subset of the 41 tasks, and the merged Monet (DarwinX) exceeds every specialist and the
Claude Code reference (dashed).
</p>
</div>
<div>
<h4>The in-loop proxy overfits; the population absorbs it</h4>
<p style="font-size:0.95rem">
During evolution the training-subset score saturates from 0.505 to 1.000, yet held-out pass@1 is
68.3% — a <b>31.7-point gap</b> between the proxy the search maximizes and the held-out
truth it never sees. Crucially, the variant that best fits the proxy is <i>not</i> the best
generalizer: four high-scoring specialists solve 24, 25, 26, and 27 of the 41 held-out tasks on
overlapping but distinct subsets, and their merge reaches 28/41. Keeping a population, rather than
following the single best-scoring lineage, is what converts an overfit proxy into held-out gain.
</p>
</div>
</div>
<div class="block">
<h4>Explore the merge: which specialists cover which held-out tasks</h4>
<p class="tnote" style="margin-top:0;margin-bottom:14px">
Every cell is one of the 41 held-out tasks, filled when at least one selected specialist solves it.
Toggle specialists to watch coverage grow, and turn on the comparison to see how the harness DarwinX
actually produced differs from the plain union. Hover any cell for its task id.
Built from the per-task records of the five evaluation runs.
</p>
<div class="widget" id="tw-widget">
<div class="wtop">
<span class="wlabel">Specialists</span>
<div class="chips" id="tw-chips"></div>
<span class="spacer" style="flex:1"></span>
<label class="toggle"><input type="checkbox" id="tw-cmp" /> compare with realized merge</label>
</div>
<div class="readout" id="tw-readout"></div>
<div class="tgrid" id="tw-grid"></div>
<div class="wlegend" id="tw-legend">
<span><i class="sw" style="background:var(--blue)"></i> solved by selection</span>
<span><i class="sw" style="background:#e6ecf2"></i> unsolved</span>
</div>
<p class="tnote" id="tw-note"></p>
</div>
</div>
</div>
<!-- ============ WebArena-Infinity ============ -->
<div class="bench">
<h3>WebArena-Infinity <span class="tag">— synthetic → real intent shift</span></h3>
<p class="lede">Evolution sees synthetic intents only; evaluation is the official 1,260-task real suite
across 10 applications. We report <i>audit-clean</i> pass@1 throughout: trajectories flagged invalid
by the validity audit count as failures even when the verifier passes them.</p>
<div class="scroller">
<table class="sortable">
<caption>WebArena-Infinity audit-clean pass@1 by application. Click any column header to sort. External baselines are as-reported
without re-auditing, which makes this comparison conservative for DarwinX. BU = Browser Use.</caption>
<thead>
<tr>
<th>Application</th>
<th class="num">Kimi</th><th class="num">Qwen</th><th class="num">Gemini+BU</th>
<th class="num">GPT-5.5+BU</th><th class="num">Monet (base)</th>
<th class="num">Monet (DarwinX)</th><th class="num">Δ</th>
</tr>
</thead>
<tbody>
<tr><td>Elation clinical records</td><td class="num">50.0</td><td class="num">54.2</td><td class="num">81.7</td><td class="num">92.5</td><td class="num">95.8</td><td class="num"><b>96.7</b></td><td class="num delta">+0.9</td></tr>
<tr><td>Elation prescriptions</td><td class="num">23.3</td><td class="num">41.7</td><td class="num">80.8</td><td class="num">90.8</td><td class="num">20.0</td><td class="num"><b>95.0</b></td><td class="num delta">+75.0</td></tr>
<tr><td>GitLab plan and track</td><td class="num">39.3</td><td class="num">37.1</td><td class="num">63.6</td><td class="num">77.9</td><td class="num">63.6</td><td class="num"><b>97.9</b></td><td class="num delta">+34.3</td></tr>
<tr><td>Gmail</td><td class="num">70.0</td><td class="num">56.7</td><td class="num">75.0</td><td class="num">85.0</td><td class="num">25.0</td><td class="num"><b>98.3</b></td><td class="num delta">+73.3</td></tr>
<tr><td>Gmail accounts and contacts</td><td class="num">40.0</td><td class="num">33.3</td><td class="num">61.7</td><td class="num">87.5</td><td class="num">21.7</td><td class="num"><b>91.7</b></td><td class="num delta">+70.0</td></tr>
<tr><td>Handshake career exploration</td><td class="num">50.0</td><td class="num">50.5</td><td class="num">50.5</td><td class="num">83.5</td><td class="num">36.5</td><td class="num"><b>84.0</b></td><td class="num delta">+47.5</td></tr>
<tr><td>Linear account settings</td><td class="num">54.2</td><td class="num">65.8</td><td class="num">73.3</td><td class="num">81.7</td><td class="num">43.3</td><td class="num"><b>94.2</b></td><td class="num delta">+50.9</td></tr>
<tr><td>PayPal wallet</td><td class="num">70.7</td><td class="num">71.4</td><td class="num">88.6</td><td class="num">90.0</td><td class="num">49.3</td><td class="num"><b>95.7</b></td><td class="num delta">+46.4</td></tr>
<tr><td>Superhuman general</td><td class="num">15.0</td><td class="num">25.8</td><td class="num">50.0</td><td class="num">80.8</td><td class="num">31.7</td><td class="num"><b>87.5</b></td><td class="num delta">+55.8</td></tr>
<tr><td>Xero invoicing</td><td class="num">52.5</td><td class="num">55.8</td><td class="num">80.8</td><td class="num">93.3</td><td class="num">39.2</td><td class="num"><b>96.7</b></td><td class="num delta">+57.5</td></tr>
<tr class="total"><td>Overall</td><td class="num">43.3</td><td class="num">48.3</td><td class="num">69.3</td><td class="num">86.1</td><td class="num">43.5</td><td class="num">93.0</td><td class="num delta">+49.5</td></tr>
</tbody>
</table>
</div>
<p class="tnote">
93.0% audit-clean beats the strongest same-model baseline (GPT-5.5 + Browser Use, 86.1%) by
6.9 points and the top public agent (Gemini 3 Flash + Browser Use, 69.3%) by
23.7 points. Given differing action spaces across public agents, we treat the same-model gap as
the primary comparison. In counts, the base solves 548/1,260 and the evolved harness 1,171/1,260.
</p>
<div class="block">
<h4>Evolution as optimization</h4>
<p class="tnote" style="margin-top:0;margin-bottom:14px">
Screening score of every evaluated variant in the WebArena-Infinity run, in archive order. Hover a
point for its score; the line is the best-so-far envelope.
</p>
<div class="widget" id="cv-widget">
<div class="wtop">
<span class="wlabel">Show</span>
<label class="toggle"><input type="checkbox" id="cv-all" checked /> every evaluated variant</label>
<label class="toggle"><input type="checkbox" id="cv-best" checked /> best so far</label>
</div>
<svg class="chart" id="cv-chart" viewBox="0 0 860 320" role="img"
aria-label="Screening score per evaluated variant, with the best-so-far envelope."></svg>
<p class="tnote" id="cv-note"></p>
</div>
<noscript class="nojs"><img src="assets/wai_evolution_curve.png" alt="Best-so-far screening score across accepted generations." /></noscript>
</div>
<figure class="block">
<a class="zoom" href="assets/wai_tree.png" target="_blank" rel="noopener">
<img src="assets/wai_tree.png" alt="Archive lineage tree with accepted and reverted variants, the primary lineage, and recombination edges." loading="lazy" />
</a>
<figcaption>
The same run as an archive lineage tree (node size ∝ screening score): accepted (blue) and
reverted (grey) variants, the primary lineage (gold, base → evolved), and
recombination edges (dashed). The reverted nodes are the preserve-and-extend contract doing its job:
a variant that broke an inherited solve contributes its distilled lesson but not its edit.
</figcaption>
</figure>
<div class="block">
<h4>Anti-cheating audit</h4>
<table style="margin-top:12px">
<caption>Validity audit on the 1,260 real tasks. Trajectories that reach the goal through the
evaluation plane, privileged knowledge, or exploits are counted as failures.</caption>
<thead>
<tr><th>Measure</th><th class="num">Monet (base)</th><th class="num">Monet (DarwinX)</th><th class="num">Δ</th></tr>
</thead>
<tbody>
<tr class="ours"><td>Audit-clean pass@1</td><td class="num">43.5</td><td class="num">93.0</td><td class="num delta">+49.5 pp</td></tr>
<tr><td>Confirmed invalid</td><td class="num">23.5%</td><td class="num">1.4%</td><td class="num">−22.1 pp</td></tr>
<tr><td>Human review</td><td class="num">5.1%</td><td class="num">0.1%</td><td class="num">−5.0 pp</td></tr>
<tr><td>Blocked attempt</td><td class="num">14.3%</td><td class="num">0.1%</td><td class="num">−14.2 pp</td></tr>
</tbody>
</table>
<p class="tnote">
Evolution does not buy the gain with shortcuts: it cuts invalid trajectories from 293 to 17, the
evaluation-plane, privileged-knowledge, and exploit mechanisms disappear entirely, and the residual
17 are all raw-state mutations concentrated in a single application. Dropping unaudited successes
too still leaves 92.9%.
</p>
<div class="block">
<div class="widget" id="iv-widget">
<div class="wtop">
<span class="wlabel">Break down by</span>
<div class="chips" id="iv-chips"></div>
<span style="flex:1"></span>
<label class="toggle"><input type="checkbox" id="iv-norm" /> scale each row to its own total</label>
</div>
<div id="iv-rows"></div>
<div class="keys" id="iv-keys"></div>
<p class="tnote" id="iv-note"></p>
</div>
<noscript class="nojs"><img src="assets/wai_invalid_composition.png" alt="Invalid trajectories before vs. after evolution, 293 to 17." /></noscript>
<p class="tnote">
Invalid trajectories before vs. after evolution. On a shared scale the second row is almost
invisible, which is the result: 293 invalid trajectories become 17. Scale each row to its own
total to inspect what the remainder is made of.
</p>
</div>
<div class="block">
<h4>What the audit costs each harness</h4>
<p class="tnote" style="margin-top:0;margin-bottom:14px">
Each line runs from raw pass@1 to pass@1 after the validity audit. A long line means the harness
was leaning on trajectories the audit rejects.
</p>
<div class="widget" id="ad-widget">
<div class="wtop">
<span class="wlabel">Show</span>
<label class="toggle"><input type="checkbox" id="ad-base" checked /> Monet (base)</label>
<label class="toggle"><input type="checkbox" id="ad-dx" checked /> Monet (DarwinX)</label>
<span style="flex:1"></span>
<label class="toggle"><input type="checkbox" id="ad-sort" /> sort by audit loss</label>
</div>
<div class="dumb" id="ad-rows"></div>
<div class="wlegend">
<span><i class="sw" style="background:#fff;box-shadow:inset 0 0 0 2px #9aa8b8;border-radius:50%"></i> raw pass@1</span>
<span><i class="sw" style="background:#8794a6;border-radius:50%"></i> base, audited</span>
<span><i class="sw" style="background:var(--blue);border-radius:50%"></i> DarwinX, audited</span>
</div>
<p class="tnote" id="ad-note"></p>
</div>
<noscript class="nojs"><img src="assets/wai_audit_by_app.png" alt="Raw vs. audited pass@1 per application." /></noscript>
</div>
</div>
</div>
<!-- ============ SWE-bench Verified ============ -->
<div class="bench">
<h3>SWE-bench Verified <span class="tag">— zero-shot cross-benchmark transfer</span></h3>
<p class="lede">The best Terminal-Bench 2.1 harness is run <i>unchanged</i> on all 500 SWE-bench
Verified instances. No SWE-V feedback ever enters the loop; we report it as a transfer target only and
run no in-domain SWE-V evolution.</p>
<table>
<caption>SWE-bench Verified official pass@1, graded by SWE-V's own test harness.</caption>
<thead>
<tr><th>Harness</th><th class="num">Solved</th><th class="num">pass@1</th></tr>
</thead>
<tbody>
<tr class="ours"><td>Monet (DarwinX), TB2.1-specialized, transferred unchanged</td><td class="num">421/500</td><td class="num">84.2%</td></tr>
<tr><td>Fix-skill reference</td><td class="num">—</td><td class="num">80.8%</td></tr>
</tbody>
</table>
<p class="tnote">
+3.4 points over the reference without receiving any SWE-V feedback. The transferred agent preserves
strong repository-level competence, so what evolved on terminal tasks is general agent capability
rather than benchmark-specific patching.
</p>
</div>
</div>
</section>
<section id="ablation">
<div class="wrap">
<h2>Ablation: what evolution changes</h2>
<p>
Anchored on Terminal-Bench 2.1, we compare base Monet (75.5%) with the evolved best (83.2%) along
two axes: the <b>skill-bundle diff</b> between the two harnesses, and <b>where</b> on the benchmark the
gains land. This is an <i>exploratory attribution, not a per-skill causal ablation</i> — the skills
were co-selected, not independently randomized.
</p>
<table style="margin-top:20px">
<caption>The seven skills the evolved lineage adds over base Monet. Every one belongs to a single
family: verification / artifact-contract. None adds domain knowledge.</caption>
<thead>
<tr><th style="width:42%">Evolved skills</th><th>Role</th></tr>
</thead>
<tbody>
<tr>
<td><code>verifier-contract</code><br /><code>contract-candidate</code></td>
<td>Derive the task's acceptance contract and check the solution against it before finalizing.</td>
</tr>
<tr>
<td><code>graded-artifact-final-check</code><br /><code>artifact-verification-loop</code></td>
<td>Verify the graded artifact (output file, format, and values) and iterate a fix-and-recheck loop.</td>
</tr>
<tr>
<td><code>real-tool-artifact</code><br /><code>tool-grounded-artifact</code></td>
<td>Ground outputs in real tool execution rather than asserted or simulated results.</td>
</tr>
<tr>
<td><code>security-contract-repair</code></td>
<td>Repair the solution against security and contract checks.</td>
</tr>
</tbody>
</table>
<div class="block" style="margin-top:30px">
<h4>Where the gains land</h4>
<p>
Gains concentrate where a frozen base has the most headroom and where verifying against a contract
unblocks fragile multi-step work: <i>ML & scientific-computing</i> (+14.8, 60→75%) and
<i>data/database</i> (+13.8, 84→98%). Clusters where the base is already strong barely move
(system administration 92→98%, security 85→84%, within noise), and no cluster regresses
beyond the per-task noise band. <b>That asymmetry is the empirical footprint of the
preserve-and-extend rule</b>: extend the fragile clusters, hold the solved ones fixed. The unblocked
difficulty is procedural — long dependency installs, environment setup, output verification,
multi-step tool use — rather than knowledge-bound, which matches a verification/artifact-contract
bundle rather than a stronger model.
</p>
</div>
<div class="block">
<h4>Cross-benchmark cross-checks</h4>
<p>
<b>WebArena-Infinity, the same family in a different modality.</b> The evolved browser harness adds
the same kind of skills (state and action contracts), and its confirmed-invalid rate falls from 23.5%
to 1.4% while audit-clean pass@1 rises 49.5 points, with the largest gains on state-changing
applications. Verification-before-finalization is thus visible under a different interaction modality
and reward source.
</p>
<p>
<b>TerminalWorld, diversity rather than a single skill.</b> Individual specialists solve 24–27
held-out tasks and the merged harness solves 28, so the archive is valuable as a source of
complementary behaviors even when the training-subset ranking is noisy.
</p>
<p class="tnote">
Together these make verification-before-finalization and contract-aware tool use a plausible shared
mechanism across benchmarks, offered as an explanation rather than a per-skill causal estimate.
</p>
</div>
</div>
</section>
<section id="limitations">
<div class="wrap">
<h2>Limitations</h2>
<ul class="lim">
<li><b>Transfer is measured in one direction, and is much smaller.</b> A TB2.1-evolved harness reaches
84.2% on SWE-bench Verified, ahead of the fix-skill reference but inside a narrow band around it
(official scores across the harnesses we compare span just 80.8–84.2%). The transferred gain is
far smaller than the in-domain ones, and we make no in-domain SWE-V claim.</li>
<li><b>Recombination's own contribution is not isolated.</b> The archive and merge machinery can
preserve and combine variants, but population search needs diverse wins before inheritance is useful,
and the contribution of recombination relative to single-lineage mutation still requires controlled
ablation.</li>
<li><b>The claim is system-level, not per-operator.</b> The experiments evaluate the complete system:
archive, parent selector, recombination operator, and inference effort are not independently
randomized, and public leaderboard rows use different models and effort settings. Matched-model deltas
support a harness claim; any individual operator's contribution, and the attribution to a
verification/contract mechanism, remain plausible rather than causal.</li>
<li><b>TerminalWorld is small.</b> With only 41 held-out tasks, one solve moves pass@1 by 2.4 points,
and the matched Opus comparison (25/41 vs. 28/41, McNemar <i>p</i>=0.45) is suggestive rather than
decisive.</li>
<li><b>The validity audit is not a formal sandbox.</b> Our WAI policy permits client-visible observation
and semantic application operations but rejects privileged knowledge, evaluation-plane access,
raw-state fabrication, database manipulation, and exploits. The static-plus-LLM audit is far stronger
than a keyword heuristic, yet deeply dynamic construction can require human review and a small number
of trajectories were unavailable, so we report both raw and conservative audit-clean scores.</li>
<li><b>Freezing the base model is a choice, not a ceiling.</b> It is what makes these deltas
attributable to the harness. Coupling harness selection with weight updates, and treating the harness
as an asset that outlives a base-model generation, are left to future work.</li>
</ul>
</div>
</section>
<section id="bibtex">
<div class="wrap">
<div class="bibhead">
<h2>BibTeX</h2>
<button id="copybtn" type="button">Copy</button>
</div>
<pre id="bib">@article{zhang2026darwinx,
title = {DarwinX: Evolving Agent Harnesses Through Natural Selection},
author = {Zhang, Yifan and Dai, Yutong and Tan, Juntao and Yang, Luyu and
Mullur, Rishi and Hoang, Thai and Hu, Zhiyuan and Zhu, James and
Mui, Phil and Savarese, Silvio and Xu, Ran and Chen, Zeyuan},
journal = {arXiv preprint arXiv:2608.07545},
year = {2026}
}</pre>
</div>
</section>
<footer>
<div class="wrap">
Salesforce AI Research · 2026. All figures and numbers are from
<a href="https://arxiv.org/abs/2608.07545">arXiv:2608.07545</a>.
Click any figure to open it at full resolution.
</div>
</footer>
<div id="tip" role="tooltip"></div>
<script src="assets/data.js"></script>
<script>
(function () {
"use strict";
var D = window.DARWINX;
var $ = function (id) { return document.getElementById(id); };
/* ---------------- shared tooltip ---------------- */
var tip = $("tip");
function showTip(html, evt) {
tip.innerHTML = html;
tip.style.opacity = "1";
var r = tip.getBoundingClientRect();
var x = Math.min(Math.max(8, evt.clientX - r.width / 2), window.innerWidth - r.width - 8);
var y = evt.clientY - r.height - 12;
if (y < 8) { y = evt.clientY + 16; }
tip.style.left = x + "px";
tip.style.top = y + "px";
}
function hideTip() { tip.style.opacity = "0"; }
function bindTip(el, htmlFn) {
el.addEventListener("mousemove", function (e) { showTip(htmlFn(), e); });
el.addEventListener("mouseleave", hideTip);
}
/* ---------------- 0a. headline panels ---------------- */
(function () {
if (!D || !D.headline) { return; }
var P = D.headline;
var modes = [
{ id: "own", label: "per panel" },
{ id: "shared", label: "shared 0\u2013100" }
];
var mode = "own";
var chips = $("hl-chips");
modes.forEach(function (m) {
var b = document.createElement("button");
b.className = "chip";
b.type = "button";
b.textContent = m.label;
b.setAttribute("aria-pressed", m.id === mode ? "true" : "false");
b.addEventListener("click", function () {
mode = m.id;
Array.prototype.forEach.call(chips.children, function (c) {
c.setAttribute("aria-pressed", c === b ? "true" : "false");
});
render();
});
chips.appendChild(b);
});
var NS = "http://www.w3.org/2000/svg";
function mk(n, a) {
var e = document.createElementNS(NS, n);
Object.keys(a).forEach(function (k) { e.setAttribute(k, a[k]); });
return e;
}
function render() {
var showPrior = $("hl-prior").checked;
var box = $("hl-panels");
box.innerHTML = "";
P.forEach(function (p) {
var lo = mode === "shared" ? 0 : p.lo;
var hi = mode === "shared" ? 100 : p.hi;
var W = 190, H = 168, T = 20, B = 22, L = 26;
var ih = H - T - B;
var y = function (v) { return T + ih - ((v - lo) / (hi - lo)) * ih; };
var series = [{ c: "b", label: p.baseLabel, v: p.base }];
if (showPrior && p.prior !== null) { series.push({ c: "p", label: p.priorLabel, v: p.prior }); }
series.push({ c: "e", label: p.evolvedLabel, v: p.evolved });
var wrapEl = document.createElement("div");
wrapEl.className = "panel";
wrapEl.innerHTML = "<div class=\"pt\">" + p.short + "</div><div class=\"ps\">" +
p.metric + "<br>on " + p.model + "</div>";
var svg = mk("svg", { viewBox: "0 0 " + W + " " + H });
[lo, hi].forEach(function (v) {
var t = mk("text", { class: "tick", x: L - 6, y: y(v) + 3.5, "text-anchor": "end" });
t.textContent = v;
svg.appendChild(t);
});
svg.appendChild(mk("line", { class: "axl", x1: L, x2: W - 6, y1: y(lo), y2: y(lo) }));
var slot = (W - 6 - L) / series.length;
var bw = Math.min(34, slot * 0.62);
series.forEach(function (s, i) {
var cx = L + slot * (i + 0.5);
var top = y(Math.max(s.v, lo));
var rect = mk("rect", {
class: "bar " + s.c, x: cx - bw / 2, y: top, width: bw,
height: Math.max(0, y(lo) - top), rx: 2
});
svg.appendChild(rect);
var hit = mk("rect", { x: cx - slot / 2, y: T, width: slot, height: ih, fill: "transparent" });
hit.addEventListener("mousemove", function (e) {
showTip("<b>" + p.name + "</b><br>" + s.label + ": " + s.v.toFixed(1) + "%<br>" +
p.metric + ", " + p.scope, e);
});
hit.addEventListener("mouseleave", hideTip);
svg.appendChild(hit);
});
var ev = series[series.length - 1];
var dl = mk("text", { class: "dl", x: L + slot * (series.length - 0.5),
y: y(ev.v) - 6, "text-anchor": "middle" });
dl.textContent = "+" + p.delta.toFixed(1);
svg.appendChild(dl);
wrapEl.appendChild(svg);
var foot = document.createElement("div");
foot.className = "foot";
foot.textContent = p.prior === null ? "transfer only" : "vs " + p.priorLabel.split(" + ")[0];
wrapEl.appendChild(foot);
box.appendChild(wrapEl);
});
$("hl-legend").innerHTML =
"<span><i class=\"sw\" style=\"background:#c6d3e0\"></i> base / reference</span>" +
(showPrior ? "<span><i class=\"sw\" style=\"background:#8fb4e8\"></i> best prior agent</span>" : "") +
"<span><i class=\"sw\" style=\"background:var(--blue)\"></i> Monet (DarwinX)</span>";
$("hl-note").innerHTML = mode === "own"
? "Each panel is scaled to its own range, which is how the paper draws it: on a shared axis " +
"WebArena-Infinity's +49.5 flattens the two terminal benchmarks into slivers. Switch to the " +
"shared axis to see the honest relative magnitudes \u2014 the annotated deltas are the same either way."
: "On one 0\u2013100 axis the four regimes are directly comparable, and it is clear how much of the " +
"range each benchmark actually occupies. Bars within a panel share one frozen base model; the " +
"prior-agent bar uses a different model and effort setting, so it is context rather than a " +
"controlled comparison. On SWE-bench Verified the grey bar is the fix-skill reference, not an " +
"unevolved Monet, and no prior-agent bar is drawn.";
}
$("hl-prior").addEventListener("change", render);
render();
})();
/* ---------------- 0b. TerminalWorld specialist bars ---------------- */
(function () {
if (!D || !D.twBars) { return; }
var T = D.twBars;
function pct(s) { return (100 * s / T.n); }
// The arms span 58.5-68.3, so a full 0-100 track renders them as six near-identical
// bars. Zoomed is the default (and how the paper draws it); the full axis stays one
// click away because a truncated axis exaggerates the spread.
var scales = [{ id: "zoom", label: "55\u201370", lo: 55, hi: 70 },
{ id: "full", label: "0\u2013100", lo: 0, hi: 100 }];
var scale = scales[0];
var chips = $("sp-chips");
scales.forEach(function (s) {
var b = document.createElement("button");
b.className = "chip";
b.type = "button";
b.textContent = s.label;
b.setAttribute("aria-pressed", s === scale ? "true" : "false");
b.addEventListener("click", function () {
scale = s;
Array.prototype.forEach.call(chips.children, function (c) {
c.setAttribute("aria-pressed", c === b ? "true" : "false");
});
render();
});
chips.appendChild(b);
});
function render() {
var counts = $("sp-counts").checked;
var pos = function (p) {
return Math.max(0, Math.min(100, 100 * (p - scale.lo) / (scale.hi - scale.lo)));
};
var refPos = pos(pct(T.claudeCode));
var box = $("sp-rows");
box.innerHTML = "";
T.bars.forEach(function (b) {
var isDx = /DarwinX/.test(b.label);
var isBase = b.label === "Base";
var row = document.createElement("div");
row.className = "hrow" + (isBase ? " dim" : "");
var color = isDx ? "var(--blue)" : (isBase ? "#c6d3e0" : "#8fb4e8");
row.innerHTML =
"<div class=\"hn\"" + (isDx ? " style=\"font-weight:600\"" : "") + ">" + b.label + "</div>" +
"<div class=\"ht\"><div class=\"hf\" style=\"width:" + pos(pct(b.solved)).toFixed(1) +
"%;background:" + color + "\"></div></div>" +
"<div class=\"hv\">" + (counts ? b.solved + "/" + T.n : pct(b.solved).toFixed(1) + "%") + "</div>";
var track = row.querySelector(".ht");
var ref = document.createElement("div");
ref.className = "refline";
ref.style.left = refPos.toFixed(1) + "%";
track.appendChild(ref);
bindTip(track, function () {
return "<b>" + b.label + "</b><br>" + b.solved + " of " + T.n + " held-out tasks (" +
pct(b.solved).toFixed(1) + "%)<br>Claude Code reference: " + T.claudeCode + "/" + T.n +
" (" + pct(T.claudeCode).toFixed(1) + "%)";
});
box.appendChild(row);
});
var tag = document.createElement("div");
tag.className = "hrow";
tag.innerHTML = "<div></div><div style=\"position:relative;height:14px\">" +
"<div class=\"reftag\" style=\"left:" + refPos.toFixed(1) + "%;top:0\">Claude Code " +
pct(T.claudeCode).toFixed(1) + "%</div></div><div></div>";
box.appendChild(tag);
$("sp-note").innerHTML = scale.id === "zoom"
? "Bars start at 55%, not 0, so the arms are distinguishable \u2014 the whole spread is 9.8 points. " +
"Switch to the full axis to see that spread in proportion."
: "On the full axis every arm looks alike, which is the honest framing of the magnitude: the merge " +
"buys 2.4 points over the best single specialist. The ordering is what matters here, not the size.";
}
$("sp-counts").addEventListener("change", render);
render();
})();
/* ---------------- 0c. TB2.1 compute ---------------- */
(function () {
if (!D || !D.tb21Compute) { return; }
var C = D.tb21Compute;
var metrics = [
{ id: "turns", label: "turns", unit: "", fmt: function (v) { return v; } },
{ id: "tokens", label: "tokens", unit: "K", fmt: function (v) { return v + "K"; } }
];
var active = "turns";
var chips = $("cp-chips");
metrics.forEach(function (m) {
var b = document.createElement("button");
b.className = "chip";
b.type = "button";
b.textContent = m.label;
b.setAttribute("aria-pressed", m.id === active ? "true" : "false");
b.addEventListener("click", function () {
active = m.id;
Array.prototype.forEach.call(chips.children, function (c) {
c.setAttribute("aria-pressed", c === b ? "true" : "false");
});
render();
});
chips.appendChild(b);
});
function render() {
var m = metrics.filter(function (x) { return x.id === active; })[0];
var d = C[active];
var max = Math.max.apply(null, d.base.concat(d.evolved));
var box = $("cp-rows");
box.innerHTML = "";
C.groups.forEach(function (g, gi) {
[["Monet (base)", d.base[gi], "#c6d3e0"], ["Monet (DarwinX)", d.evolved[gi], "var(--blue)"]]
.forEach(function (s, si) {
var row = document.createElement("div");
row.className = "hrow";
row.innerHTML =
"<div class=\"hn\">" + (si === 0 ? g : "") + "</div>" +
"<div class=\"ht\"><div class=\"hf\" style=\"width:" + (100 * s[1] / max).toFixed(1) +
"%;background:" + s[2] + "\"></div></div>" +
"<div class=\"hv\">" + m.fmt(s[1]) + "</div>";
bindTip(row.querySelector(".ht"), function () {
var other = si === 0 ? d.evolved[gi] : d.base[gi];
var ratio = si === 0 ? (other / s[1]) : (s[1] / other);
return "<b>" + g + "</b><br>" + s[0] + ": " + m.fmt(s[1]) + " median " + m.label +
"<br>evolved / base = " + ratio.toFixed(1) + "x";
});
box.appendChild(row);
});
});
$("cp-note").innerHTML = active === "turns"
? "On the six tasks that flip from failing to solved the evolved harness roughly doubles its turns " +
"(22 vs. 11). On the 69 both already solve, turns barely move (13 vs. 12): the extra effort is " +
"spent where it changes the outcome, not spread across the benchmark."
: "Tokens tell the same story more sharply: 4.3x on the newly solved six (380K vs. 89K) against " +
"1.4x on the 69 already solved (172K vs. 125K). This is verify-and-retry effort converting a " +
"near-miss into a pass, not uniformly more compute.";
}
render();
})();
/* ---------------- 0d. WAI invalid composition ---------------- */
(function () {
if (!D || !D.waiInvalid) { return; }
var V = D.waiInvalid;
var views = [{ id: "application", label: "application" }, { id: "mechanism", label: "mechanism" }];
var active = "mechanism";
var chips = $("iv-chips");
views.forEach(function (v) {
var b = document.createElement("button");
b.className = "chip";
b.type = "button";
b.textContent = v.label;
b.setAttribute("aria-pressed", v.id === active ? "true" : "false");
b.addEventListener("click", function () {
active = v.id;
Array.prototype.forEach.call(chips.children, function (c) {
c.setAttribute("aria-pressed", c === b ? "true" : "false");
});
render();
});
chips.appendChild(b);
});
function render() {
var v = V[active];
var norm = $("iv-norm").checked;
var totalBefore = v.before.reduce(function (a, b) { return a + b; }, 0);
var box = $("iv-rows");
box.innerHTML = "";
[["Before evolution", v.before], ["After evolution", v.after]].forEach(function (pair) {
var vals = pair[1];
var sum = vals.reduce(function (a, b) { return a + b; }, 0);
var denom = norm ? (sum || 1) : totalBefore;
var row = document.createElement("div");
row.className = "srow";
row.innerHTML = "<div class=\"sh\"><span>" + pair[0] +
"</span><span><b>" + sum + "</b> invalid</span></div>";
var st = document.createElement("div");
st.className = "stack";
vals.forEach(function (val, i) {
if (!val) { return; }
var seg = document.createElement("i");
seg.style.width = (100 * val / denom).toFixed(2) + "%";
seg.style.background = v.colors[i];
bindTip(seg, function () {
return "<b>" + v.labels[i] + "</b><br>" + val + " invalid trajector" +
(val === 1 ? "y" : "ies") + "<br>" + (100 * val / (sum || 1)).toFixed(1) +
"% of this row";
});
st.appendChild(seg);
});
row.appendChild(st);
box.appendChild(row);
});
$("iv-keys").innerHTML = v.labels.map(function (l, i) {
if (!v.before[i] && !v.after[i]) { return ""; }
return "<span><i class=\"sw\" style=\"background:" + v.colors[i] + "\"></i>" + l + "</span>";
}).join("");
$("iv-note").innerHTML = active === "mechanism"
? "Evaluation-plane access, privileged knowledge, and exploits go to <b>zero</b>. Every one of the " +
"17 remaining invalid trajectories is a raw-state mutation, up from 15 of 293 before \u2014 the " +
"category that survives is the mildest one."
: "The remainder concentrates: 8 of the 17 sit in a single application (Super-human), and five of " +
"the ten applications end with no invalid trajectories at all.";
}
$("iv-norm").addEventListener("change", render);
render();
})();
/* ---------------- 0e. WAI audit dumbbell ---------------- */
(function () {
if (!D || !D.waiAudit) { return; }
var A = D.waiAudit;
function render() {
var showBase = $("ad-base").checked, showDx = $("ad-dx").checked;
var idx = A.apps.map(function (_, i) { return i; });
var overall = idx.shift();
if ($("ad-sort").checked) {
idx.sort(function (a, b) {
return (A.rawBase[b] - A.auditedBase[b]) - (A.rawBase[a] - A.auditedBase[a]);
});
}
var box = $("ad-rows");
box.innerHTML = "";
var ax = document.createElement("div");
ax.className = "drow dax";
ax.innerHTML = "<div class=\"dn\" style=\"color:var(--muted);font-size:0.7rem\">pass@1 (%)</div>" +
"<div class=\"dt\">" + [0, 25, 50, 75, 100].map(function (t) {
return "<span style=\"left:" + t + "%\">" + t + "</span>";
}).join("") + "</div>";
box.appendChild(ax);
[overall].concat(idx).forEach(function (i, pos) {
var row = document.createElement("div");
row.className = "drow" + (pos === 0 ? " tot" : "");
row.innerHTML = "<div class=\"dn\">" + A.apps[i] + "</div><div class=\"dt\"></div>";
var t = row.querySelector(".dt");
function lane(raw, aud, color, top) {
var x1 = Math.min(raw, aud), x2 = Math.max(raw, aud);
var seg = document.createElement("div");
seg.className = "seg";
seg.style.left = x1 + "%";
seg.style.width = (x2 - x1) + "%";
seg.style.top = top + "px";
seg.style.background = color;
t.appendChild(seg);
var p1 = document.createElement("div");
p1.className = "pt2 hollow";
p1.style.left = raw + "%";
p1.style.top = (top + 1.5) + "px";
p1.style.boxShadow = "inset 0 0 0 2px " + color;
t.appendChild(p1);
var p2 = document.createElement("div");
p2.className = "pt2";
p2.style.left = aud + "%";
p2.style.top = (top + 1.5) + "px";
p2.style.background = color;
t.appendChild(p2);
}
if (showBase) { lane(A.rawBase[i], A.auditedBase[i], "#8794a6", 8); }
if (showDx) { lane(A.rawDx[i], A.auditedDx[i], "#0176d3", 19); }
bindTip(t, function () {
return "<b>" + A.apps[i] + "</b><br>base " + A.rawBase[i].toFixed(1) + " \u2192 " +
A.auditedBase[i].toFixed(1) + " (\u2212" + (A.rawBase[i] - A.auditedBase[i]).toFixed(1) +
")<br>DarwinX " + A.rawDx[i].toFixed(1) + " \u2192 " + A.auditedDx[i].toFixed(1) +
" (\u2212" + (A.rawDx[i] - A.auditedDx[i]).toFixed(1) + ")";
});
box.appendChild(row);
});
var lossB = A.rawBase[0] - A.auditedBase[0], lossD = A.rawDx[0] - A.auditedDx[0];
$("ad-note").innerHTML = "Overall the audit costs the base <b>" + lossB.toFixed(1) +
"</b> points and DarwinX <b>" + lossD.toFixed(1) + "</b>. The gap therefore <i>widens</i> under " +
"scrutiny, from " + (A.rawDx[0] - A.rawBase[0]).toFixed(1) + " points raw to " +
(A.auditedDx[0] - A.auditedBase[0]).toFixed(1) + " audited, which is why the page reports the " +
"audited number. Audit coverage is 99.0% for the base and 99.4% for the evolved harness.";
}
["ad-base", "ad-dx", "ad-sort"].forEach(function (id) {
$(id).addEventListener("change", render);
});
render();
})();
/* ---------------- 1. TerminalWorld merge explorer ---------------- */
(function () {
if (!D || !D.tw) { return; }
var tasks = D.tw.tasks;
var counts = D.tw.counts;
var KEYS = ["A", "B", "C", "D"];
var sel = { A: true, B: true, C: true, D: true };
var chips = $("tw-chips");
KEYS.forEach(function (k, i) {
var b = document.createElement("button");
b.className = "chip";
b.type = "button";
b.setAttribute("aria-pressed", "true");
b.innerHTML = "Spec. " + k + " <span class=\"cnt\">" + counts[k] + "/41</span>";
b.addEventListener("click", function () {
sel[k] = !sel[k];
b.setAttribute("aria-pressed", sel[k] ? "true" : "false");
render();
});
chips.appendChild(b);
void i;
});
var grid = $("tw-grid");
var cells = tasks.map(function (t) {
var c = document.createElement("div");
c.className = "tcell";
grid.appendChild(c);
bindTip(c, function () {
var solvers = KEYS.filter(function (k) { return t.bits[KEYS.indexOf(k)] === "1"; });
return "<b>" + t.id + "</b><br>solved by: " +
(solvers.length ? "Spec. " + solvers.join(", ") : "no specialist") +
"<br>realized merge: " + (t.bits[4] === "1" ? "solved" : "unsolved");
});
return c;
});
function unionOf(keys) {
return tasks.filter(function (t) {
return keys.some(function (k) { return t.bits[KEYS.indexOf(k)] === "1"; });
});
}
function render() {
var on = KEYS.filter(function (k) { return sel[k]; });
var u = unionOf(on);
var uIds = {};
u.forEach(function (t) { uIds[t.id] = 1; });
var cmp = $("tw-cmp").checked;
cells.forEach(function (c, i) {
var t = tasks[i];
var inU = !!uIds[t.id];
var inM = t.bits[4] === "1";
c.className = "tcell" + (inU ? " on" : "") +
(cmp && inM && !inU ? " gain" : "") + (cmp && inU && !inM ? " loss" : "");
});
var best = on.length ? Math.max.apply(null, on.map(function (k) { return counts[k]; })) : 0;
$("tw-readout").innerHTML =
ro("Selected", on.length ? on.join(" + ") : "none", true) +
ro("Union coverage", u.length + "<small>/41</small>") +
ro("Best single selected", (best || "\u2014") + (best ? "<small>/41</small>" : "")) +
ro("Realized merge", counts.M + "<small>/41</small>");
var note = "";
if (on.length === 4) {
note = "All four specialists together cover <b>29/41</b>, two more than the best single lineage " +
"(Spec. D at 27). The harness DarwinX actually produced solves <b>28/41</b>: recombination is not " +
"a free set union. Turn on the comparison to see the difference \u2014 the merge picks up " +
"<code>tw_448247</code>, which <i>no</i> specialist solves, and drops <code>tw_449421</code> and " +
"<code>tw_498533</code>. These are single attempts, so individual flips sit inside the noise band " +
"(one task = 2.4 points).";
} else if (on.length === 0) {
note = "Nothing selected. 12 of the 41 held-out tasks are solved by none of the four specialists.";
} else if (u.length > best) {
note = "This selection covers <b>" + u.length + "/41</b>, which is " + (u.length - best) +
" more than its best individual member \u2014 the specialists solve overlapping but distinct subsets.";
} else {
note = "This selection covers <b>" + u.length + "/41</b>, no more than its best individual member: " +
"these lineages are not complementary.";
}
$("tw-note").innerHTML = note;
var lg = $("tw-legend");
lg.innerHTML =
"<span><i class=\"sw\" style=\"background:var(--blue)\"></i> solved by selection</span>" +
"<span><i class=\"sw\" style=\"background:#e6ecf2\"></i> unsolved</span>" +
(cmp ? "<span><i class=\"sw\" style=\"box-shadow:0 0 0 2px #e8a33d;background:#e6ecf2\"></i>" +
" merge solves, selection does not</span>" +
"<span><i class=\"sw\" style=\"box-shadow:0 0 0 2px #d05a4e;background:var(--blue)\"></i>" +
" selection solves, merge does not</span>" : "");
}
function ro(k, v, plain) {
return "<div class=\"ro" + (plain ? "" : " hi") + "\"><div class=\"k\">" + k +
"</div><div class=\"v\"" + (plain ? " style=\"font-size:0.95rem;font-weight:600\"" : "") + ">" +
v + "</div></div>";
}
$("tw-cmp").addEventListener("change", render);
render();
})();
/* ---------------- 2. TB2.1 cluster explorer ---------------- */
(function () {
if (!D || !D.tb21Clusters) { return; }
var data = D.tb21Clusters.slice();
var modes = [
{ id: "delta", label: "gain", cmp: function (a, b) { return b.delta - a.delta; } },
{ id: "base", label: "base rate", cmp: function (a, b) { return a.base - b.base; } },
{ id: "n", label: "cluster size", cmp: function (a, b) { return b.n - a.n; } },
{ id: "name", label: "name", cmp: function (a, b) { return a.name.localeCompare(b.name); } }
];
var active = "delta";
var chips = $("cl-chips");
modes.forEach(function (m) {
var b = document.createElement("button");
b.className = "chip";
b.type = "button";
b.textContent = m.label;
b.setAttribute("aria-pressed", m.id === active ? "true" : "false");
b.addEventListener("click", function () {
active = m.id;
Array.prototype.forEach.call(chips.children, function (c) {
c.setAttribute("aria-pressed", c === b ? "true" : "false");
});
render();
});
chips.appendChild(b);
});
function render() {
var mode = modes.filter(function (m) { return m.id === active; })[0];
var rows = data.slice().sort(mode.cmp);
var showBase = $("cl-base").checked;
var box = $("cl-bars");
box.innerHTML = "";
rows.forEach(function (d) {
var el = document.createElement("div");
el.className = "cbar";
el.innerHTML =
"<div class=\"cname\">" + d.name + " <span class=\"cn\">n=" + d.n + "</span></div>" +
"<div class=\"track\">" +
(showBase ? "<div class=\"fill b\" style=\"left:0;width:" + d.base + "%\"></div>" : "") +
"<div class=\"fill e\" style=\"left:0;width:" + d.evolved + "%;height:" +
(showBase ? "58%;top:21%" : "100%;top:0") + "\"></div>" +
"</div>" +
"<div class=\"cd " + (d.delta >= 0 ? "up" : "down") + "\">" +
(d.delta >= 0 ? "+" : "\u2212") + Math.abs(d.delta).toFixed(1) + "</div>";
bindTip(el.querySelector(".track"), function () {
return "<b>" + d.name + "</b><br>base " + d.base.toFixed(1) + "% \u2192 evolved " +
d.evolved.toFixed(1) + "%<br>" + d.n + " tasks, \u0394 " +
(d.delta >= 0 ? "+" : "\u2212") + Math.abs(d.delta).toFixed(1) + " points";
});
box.appendChild(el);
});
}
$("cl-base").addEventListener("change", render);
render();
})();
/* ---------------- 3. WAI evolution curve ---------------- */
(function () {
if (!D || !D.waiCurve) { return; }
var s = D.waiCurve;
var svg = $("cv-chart");
var W = 860, H = 320, L = 44, R = 14, T = 14, B = 34;
var iw = W - L - R, ih = H - T - B;
var NS = "http://www.w3.org/2000/svg";
function x(i) { return L + (s.length === 1 ? iw / 2 : (i / (s.length - 1)) * iw); }
function y(v) { return T + ih - (v / 100) * ih; }
function mk(n, a) {
var e = document.createElementNS(NS, n);
Object.keys(a).forEach(function (k) { e.setAttribute(k, a[k]); });
return e;
}
function render() {
while (svg.firstChild) { svg.removeChild(svg.firstChild); }
[0, 25, 50, 75, 100].forEach(function (v) {
svg.appendChild(mk("line", { class: "gl", x1: L, x2: W - R, y1: y(v), y2: y(v) }));
var t = mk("text", { class: "lbl", x: L - 8, y: y(v) + 3.5, "text-anchor": "end" });
t.textContent = v;
svg.appendChild(t);
});
svg.appendChild(mk("line", { class: "ax", x1: L, x2: L, y1: T, y2: T + ih }));
svg.appendChild(mk("line", { class: "ax", x1: L, x2: W - R, y1: T + ih, y2: T + ih }));
[0, 9, 18, 27, 36].forEach(function (i) {
if (i >= s.length) { return; }
var t = mk("text", { class: "lbl", x: x(i), y: T + ih + 17, "text-anchor": "middle" });
t.textContent = i;
svg.appendChild(t);
});
var xl = mk("text", { class: "lbl", x: L + iw / 2, y: H - 3, "text-anchor": "middle" });
xl.textContent = "archive node";
svg.appendChild(xl);
var yl = mk("text", { class: "lbl", x: 12, y: T + ih / 2, "text-anchor": "middle",
transform: "rotate(-90 12 " + (T + ih / 2) + ")" });
yl.textContent = "screening score (%)";
svg.appendChild(yl);
if ($("cv-best").checked) {
var m = -1, pts = [];
s.forEach(function (v, i) {
m = Math.max(m, v);
if (i > 0) { pts.push(x(i) + "," + y(m)); }
pts.push(x(i) + "," + y(m));
});
svg.appendChild(mk("polyline", { class: "best", points: pts.join(" ") }));
}
if ($("cv-all").checked) {
s.forEach(function (v, i) {
var dot = mk("circle", { class: "dot", cx: x(i), cy: y(v), r: 3.4 });
svg.appendChild(dot);
var hit = mk("circle", { class: "hit", cx: x(i), cy: y(v), r: 11 });
hit.addEventListener("mousemove", function (e) {
dot.classList.add("act");
showTip("<b>archive node " + i + "</b><br>screening score " + v.toFixed(1) + "%", e);
});
hit.addEventListener("mouseleave", function () { dot.classList.remove("act"); hideTip(); });
svg.appendChild(hit);
});
}
}
var best = Math.max.apply(null, s);
$("cv-note").innerHTML = "The run evaluates <b>" + s.length + "</b> variants. The screening score " +
"climbs from " + s[0].toFixed(1) + "% to <b>" + best.toFixed(1) + "%</b>, but not monotonically: " +
"individual variants score well below the running best, and those are the ones the " +
"preserve-and-extend contract reverts. Screening runs on synthetic intents, so this axis is the " +
"in-loop proxy, not the 1,260-task real result.";
$("cv-all").addEventListener("change", render);
$("cv-best").addEventListener("change", render);
render();
})();
/* ---------------- 4. sortable tables ---------------- */
(function () {
function num(td) {
var v = parseFloat(td.textContent.replace(/[^0-9.\-\u2212]/g, "").replace("\u2212", "-"));
return isNaN(v) ? null : v;
}
Array.prototype.forEach.call(document.querySelectorAll("table.sortable"), function (tb) {
var ths = tb.querySelectorAll("thead th");
Array.prototype.forEach.call(ths, function (th, ci) {
th.classList.add("sortable");
th.addEventListener("click", function () {
var body = tb.querySelector("tbody");
var rows = Array.prototype.slice.call(body.rows);
var pinned = rows.filter(function (r) { return r.classList.contains("total"); });
var movable = rows.filter(function (r) { return !r.classList.contains("total"); });
if (!movable.length) { return; }
// Numeric columns open descending (best first); text columns open A-Z.
var numeric = movable.some(function (r) { return num(r.cells[ci]) !== null; });
var desc;
if (th.classList.contains("desc")) { desc = false; }
else if (th.classList.contains("asc")) { desc = true; }
else { desc = numeric; }
Array.prototype.forEach.call(ths, function (o) { o.classList.remove("asc", "desc"); });
th.classList.add(desc ? "desc" : "asc");
movable.sort(function (a, b) {
var x = num(a.cells[ci]), y = num(b.cells[ci]);
if (x === null || y === null) {
var av = a.cells[ci].textContent.trim(), bv = b.cells[ci].textContent.trim();
return desc ? bv.localeCompare(av) : av.localeCompare(bv);
}
return desc ? y - x : x - y;
});
movable.concat(pinned).forEach(function (r) { body.appendChild(r); });
});
});
});
})();
/* ---------------- 5. BibTeX copy ---------------- */
$("copybtn").addEventListener("click", function () {
var btn = this;
navigator.clipboard.writeText($("bib").textContent).then(
function () {
btn.textContent = "Copied";
setTimeout(function () { btn.textContent = "Copy"; }, 1600);
},
function () { btn.textContent = "Press \u2318C"; }
);
});
})();
</script>
</body>
</html>
|