File size: 112,650 Bytes
d824e4a 37b5b14 d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a bb7ad9f 2d7059c d824e4a 2d7059c d824e4a 2d7059c d824e4a 2d7059c d824e4a bb7ad9f d824e4a 2d7059c d824e4a d22a85e 37b5b14 d22a85e d824e4a 2d7059c 37b5b14 d824e4a 2d7059c d824e4a 37b5b14 d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a 37b5b14 d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a d22a85e d824e4a bb7ad9f 2d7059c 37b5b14 2d7059c 37b5b14 2d7059c bb7ad9f 37b5b14 bb7ad9f 37b5b14 bb7ad9f 2d7059c d824e4a 2d7059c d824e4a 2d7059c d824e4a 2d7059c d824e4a bb7ad9f 37b5b14 d824e4a 37b5b14 d824e4a ae7c20e | 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 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 | <!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<meta name="theme-color" content="#eaf8ff" />
<meta name="description" content="Orion / Project Prism — Transformer 2 research notes on Procedure Banks, Working State, recurrent deliberation, and the retirement of the Knowledge Vault." />
<meta property="og:title" content="Orion / Project Prism — Transformer 2 Research Notes" />
<meta property="og:description" content="Architecture notes, ablations, training telemetry, negative results, and the road to T2.2." />
<meta property="og:type" content="website" />
<meta name="twitter:card" content="summary_large_image" />
<title>Orion / Project Prism — Transformer 2 Research Blog</title>
<style>
:root{
--bg:#edf9ff;
--bg2:#f8fcff;
--ink:#0c1622;
--muted:#5c6b78;
--line:rgba(22,73,100,.12);
--glass:rgba(255,255,255,.56);
--glass-strong:rgba(255,255,255,.76);
--glass-border:rgba(255,255,255,.78);
--shadow:0 20px 60px rgba(39,88,112,.15);
--shadow-soft:0 10px 35px rgba(39,88,112,.10);
--cyan:#2eb6df;
--blue:#4d8dff;
--teal:#3ccdb4;
--mint:#8ce3cd;
--navy:#163954;
--danger:#d95b61;
--warn:#bd7a00;
--good:#278f72;
--radius:26px;
--radius-sm:18px;
--blur:28px;
--content:1180px;
}
*{box-sizing:border-box}
html{scroll-behavior:smooth}
body{
margin:0;
color:var(--ink);
font-family:-apple-system,BlinkMacSystemFont,"SF Pro Display","SF Pro Text","Segoe UI",Inter,Roboto,Helvetica,Arial,sans-serif;
background:
radial-gradient(circle at 14% 5%, rgba(112,226,255,.42), transparent 31%),
radial-gradient(circle at 88% 14%, rgba(98,188,255,.32), transparent 27%),
radial-gradient(circle at 45% 58%, rgba(117,231,209,.20), transparent 35%),
linear-gradient(180deg,var(--bg),var(--bg2) 42%,#eefbff 100%);
min-height:100vh;
overflow-x:hidden;
}
body::before{
content:"";
position:fixed;
inset:0;
pointer-events:none;
opacity:.28;
background-image:
linear-gradient(rgba(255,255,255,.7) 1px,transparent 1px),
linear-gradient(90deg,rgba(255,255,255,.55) 1px,transparent 1px);
background-size:52px 52px;
mask-image:linear-gradient(to bottom,black,transparent 80%);
z-index:-3;
}
body.dark{
--bg:#08131c;
--bg2:#0c1822;
--ink:#eef8ff;
--muted:#9fb4c2;
--line:rgba(181,220,239,.12);
--glass:rgba(13,27,37,.55);
--glass-strong:rgba(16,33,45,.78);
--glass-border:rgba(188,231,250,.14);
--shadow:0 24px 70px rgba(0,0,0,.34);
--shadow-soft:0 10px 35px rgba(0,0,0,.24);
background:
radial-gradient(circle at 14% 5%, rgba(29,138,177,.24), transparent 31%),
radial-gradient(circle at 88% 14%, rgba(63,121,205,.23), transparent 27%),
radial-gradient(circle at 45% 58%, rgba(50,173,152,.14), transparent 35%),
linear-gradient(180deg,#07131c,#0c1821 43%,#07131c 100%);
}
a{color:inherit;text-decoration:none}
button{font:inherit}
code,kbd,pre,.mono{
font-family:"SFMono-Regular",Consolas,"Liberation Mono",Menlo,monospace;
}
.ambient{
position:fixed;
border-radius:999px;
filter:blur(10px);
pointer-events:none;
z-index:-2;
opacity:.55;
animation:float 14s ease-in-out infinite;
}
.ambient.one{width:420px;height:420px;left:-160px;top:220px;background:rgba(78,214,235,.22)}
.ambient.two{width:360px;height:360px;right:-120px;top:520px;background:rgba(81,149,255,.16);animation-delay:-4s}
.ambient.three{width:300px;height:300px;left:48%;top:1050px;background:rgba(76,214,178,.14);animation-delay:-8s}
@keyframes float{50%{transform:translate3d(0,22px,0) scale(1.05)}}
.progress{
position:fixed;
top:0;left:0;
height:3px;
width:0%;
background:linear-gradient(90deg,var(--teal),var(--cyan),var(--blue));
z-index:200;
box-shadow:0 0 18px rgba(46,182,223,.45);
}
.nav-wrap{
position:sticky;
top:16px;
z-index:100;
width:min(calc(100% - 28px), var(--content));
margin:16px auto 0;
}
.nav{
display:flex;
align-items:center;
justify-content:space-between;
gap:18px;
min-height:66px;
padding:10px 12px 10px 18px;
border:1px solid var(--glass-border);
background:var(--glass);
backdrop-filter:blur(var(--blur)) saturate(165%);
-webkit-backdrop-filter:blur(var(--blur)) saturate(165%);
border-radius:24px;
box-shadow:var(--shadow-soft);
}
.brand{
display:flex;align-items:center;gap:12px;min-width:0;
}
.brand-mark{
width:38px;height:38px;border-radius:12px;
background:
radial-gradient(circle at 30% 25%,rgba(255,255,255,.95),transparent 26%),
linear-gradient(145deg,#6be0db,#45aaf4 55%,#2869db);
box-shadow:inset 0 1px 0 rgba(255,255,255,.7),0 8px 18px rgba(55,144,187,.25);
position:relative;
}
.brand-mark::after{
content:"";
position:absolute;inset:8px;
border:1.5px solid rgba(255,255,255,.8);
border-radius:9px;
transform:rotate(18deg);
}
.brand strong{display:block;font-size:.95rem;letter-spacing:.01em}
.brand span{display:block;color:var(--muted);font-size:.76rem;margin-top:2px}
.nav-links{display:flex;gap:4px;align-items:center}
.nav-links a{
color:var(--muted);
font-size:.83rem;
padding:10px 12px;
border-radius:14px;
transition:.2s ease;
}
.nav-links a:hover{color:var(--ink);background:rgba(255,255,255,.42)}
.icon-btn{
width:42px;height:42px;border:1px solid var(--glass-border);border-radius:14px;
display:grid;place-items:center;cursor:pointer;
background:rgba(255,255,255,.35);
color:var(--ink);
box-shadow:inset 0 1px 0 rgba(255,255,255,.45);
}
body.dark .icon-btn{background:rgba(255,255,255,.045)}
.hero{
width:min(calc(100% - 34px),var(--content));
margin:78px auto 42px;
position:relative;
}
.hero-grid{
display:grid;
grid-template-columns:minmax(0,1.24fr) minmax(330px,.76fr);
gap:28px;
align-items:stretch;
}
.hero-main,.hero-side,.glass-card,.callout,.stat-card,.timeline-card,.table-wrap{
border:1px solid var(--glass-border);
background:var(--glass);
backdrop-filter:blur(var(--blur)) saturate(160%);
-webkit-backdrop-filter:blur(var(--blur)) saturate(160%);
box-shadow:var(--shadow);
}
.hero-main{
border-radius:36px;
padding:52px 52px 46px;
position:relative;
overflow:hidden;
isolation:isolate;
}
.hero-main::before{
content:"";
position:absolute;
width:440px;height:440px;
right:-160px;top:-210px;
border-radius:50%;
background:linear-gradient(145deg,rgba(255,255,255,.7),rgba(97,219,237,.16));
border:1px solid rgba(255,255,255,.5);
box-shadow:inset 0 0 60px rgba(255,255,255,.55);
z-index:-1;
}
.eyebrow{
display:inline-flex;align-items:center;gap:8px;
border:1px solid var(--glass-border);
background:rgba(255,255,255,.42);
border-radius:999px;
padding:8px 12px;
font-size:.76rem;
font-weight:700;
color:#1f6884;
letter-spacing:.06em;
text-transform:uppercase;
}
body.dark .eyebrow{color:#80d9f6;background:rgba(255,255,255,.045)}
.pulse-dot{
width:8px;height:8px;border-radius:50%;background:var(--teal);
box-shadow:0 0 0 0 rgba(60,205,180,.4);
animation:pulse 2.1s infinite;
}
@keyframes pulse{70%{box-shadow:0 0 0 9px rgba(60,205,180,0)}100%{box-shadow:0 0 0 0 rgba(60,205,180,0)}}
h1{
font-size:clamp(2.8rem,6vw,5.35rem);
line-height:.91;
letter-spacing:-.045em;
margin:22px 0 24px;
max-width:810px;
}
h1 .soft{
color:transparent;
background:linear-gradient(100deg,#13506d,#2299c4 45%,#37bca1);
-webkit-background-clip:text;background-clip:text;
}
body.dark h1 .soft{background:linear-gradient(100deg,#c9f4ff,#66cdf0 50%,#79e0c9);-webkit-background-clip:text;background-clip:text}
.dek{
font-size:clamp(1.02rem,1.4vw,1.23rem);
line-height:1.65;
color:var(--muted);
max-width:790px;
margin:0;
}
.hero-actions{
display:flex;gap:12px;flex-wrap:wrap;margin-top:30px
}
.btn{
display:inline-flex;align-items:center;gap:8px;
padding:12px 16px;border-radius:15px;
font-weight:750;font-size:.9rem;
border:1px solid var(--glass-border);
transition:.2s ease;
}
.btn.primary{
color:white;
background:#197fa9;
box-shadow:0 6px 16px rgba(35,152,186,.16);
}
.btn.secondary{background:rgba(255,255,255,.38)}
body.dark .btn.secondary{background:rgba(255,255,255,.045)}
.btn:hover{transform:translateY(-1px)}
.hero-side{
border-radius:36px;
padding:24px;
display:flex;flex-direction:column;
justify-content:space-between;
overflow:hidden;position:relative;
}
.hero-side::after{
content:"";
position:absolute;
width:250px;height:250px;
border-radius:50%;
bottom:-130px;right:-90px;
background:rgba(97,210,230,.18);
filter:blur(2px);
}
.mini-title{font-size:.78rem;color:var(--muted);font-weight:700;text-transform:uppercase;letter-spacing:.08em}
.architecture-list{margin-top:20px;display:grid;gap:10px}
.arch-row{
display:grid;grid-template-columns:58px 1fr;gap:12px;align-items:center;
padding:12px 13px;border-radius:17px;
background:rgba(255,255,255,.34);
border:1px solid rgba(255,255,255,.48);
}
body.dark .arch-row{background:rgba(255,255,255,.035);border-color:rgba(255,255,255,.08)}
.arch-row .tag{
font-size:.68rem;font-weight:800;text-align:center;padding:7px 4px;border-radius:10px;
background:rgba(53,169,207,.12);color:#187399;
}
body.dark .arch-row .tag{color:#7ed8f8}
.arch-row strong{font-size:.9rem}
.arch-row span{display:block;color:var(--muted);font-size:.76rem;margin-top:2px}
.hero-note{
margin-top:22px;padding:16px;border-radius:18px;
background:rgba(255,255,255,.35);border:1px solid rgba(255,255,255,.5);
font-size:.85rem;color:var(--muted);line-height:1.5;
position:relative;z-index:2;
}
body.dark .hero-note{background:rgba(255,255,255,.035);border-color:rgba(255,255,255,.08)}
.stats{
width:min(calc(100% - 34px),var(--content));
margin:0 auto 42px;
display:grid;grid-template-columns:repeat(4,1fr);gap:14px;
}
.stat-card{
border-radius:22px;padding:20px;
box-shadow:var(--shadow-soft);
}
.stat-card .label{color:var(--muted);font-size:.73rem;text-transform:uppercase;letter-spacing:.07em;font-weight:700}
.stat-card .value{font-size:1.65rem;font-weight:820;letter-spacing:-.03em;margin-top:8px}
.stat-card .sub{color:var(--muted);font-size:.78rem;margin-top:4px}
.layout{
width:min(calc(100% - 34px),var(--content));
margin:0 auto;
display:grid;
grid-template-columns:250px minmax(0,1fr);
gap:30px;
align-items:start;
}
.toc{
position:sticky;top:102px;
border-radius:24px;padding:16px;
background:var(--glass);
border:1px solid var(--glass-border);
backdrop-filter:blur(var(--blur)) saturate(160%);
-webkit-backdrop-filter:blur(var(--blur)) saturate(160%);
box-shadow:var(--shadow-soft);
max-height:calc(100vh - 128px);
overflow:auto;
}
.toc h3{font-size:.78rem;text-transform:uppercase;letter-spacing:.08em;color:var(--muted);margin:4px 8px 10px}
.toc a{
display:block;padding:9px 10px;border-radius:12px;color:var(--muted);font-size:.81rem;line-height:1.25;
transition:.2s ease;
}
.toc a:hover,.toc a.active{background:rgba(255,255,255,.42);color:var(--ink)}
body.dark .toc a:hover,body.dark .toc a.active{background:rgba(255,255,255,.05)}
main{min-width:0}
section{
scroll-margin-top:110px;
margin-bottom:74px;
}
.section-kicker{
color:#1f87a9;font-size:.74rem;font-weight:800;letter-spacing:.09em;text-transform:uppercase;margin-bottom:10px
}
body.dark .section-kicker{color:#6dd5f4}
h2{
font-size:clamp(2rem,4vw,3.2rem);
letter-spacing:-.045em;
line-height:1.02;
margin:0 0 18px;
}
h3{
font-size:1.25rem;letter-spacing:-.02em;margin:28px 0 12px;
}
p{
font-size:1rem;line-height:1.78;color:var(--muted);margin:0 0 16px;
}
strong{color:var(--ink)}
.lead{
font-size:1.14rem;line-height:1.76;color:var(--muted);max-width:860px
}
.glass-card{
border-radius:var(--radius);
padding:26px;
margin:24px 0;
box-shadow:var(--shadow-soft);
}
.split{
display:grid;grid-template-columns:1fr 1fr;gap:18px;margin:24px 0;
}
.subcard{
padding:20px;border-radius:20px;
background:rgba(255,255,255,.34);
border:1px solid rgba(255,255,255,.52);
}
body.dark .subcard{background:rgba(255,255,255,.035);border-color:rgba(255,255,255,.08)}
.subcard h3{margin:0 0 8px;font-size:1rem}
.subcard p{font-size:.88rem;line-height:1.6;margin:0}
.equation{
margin:26px 0;padding:22px 24px;
border-radius:20px;
background:rgba(13,67,91,.92);
color:#f3fbff;
overflow:auto;
box-shadow:0 16px 38px rgba(9,58,82,.22);
font-size:clamp(.88rem,2vw,1.06rem);
}
.callout{
border-radius:24px;
padding:22px 24px;
margin:24px 0;
box-shadow:var(--shadow-soft);
position:relative;
overflow:hidden;
}
.callout::before{
content:"";
position:absolute;left:0;top:0;bottom:0;width:4px;
background:linear-gradient(var(--cyan),var(--teal));
}
.callout .title{font-weight:820;margin-bottom:7px}
.callout p{margin:0;font-size:.92rem}
.callout.warn::before{background:linear-gradient(#f1b24d,#d67a14)}
.callout.danger::before{background:linear-gradient(#e06a71,#ba4852)}
.table-wrap{
border-radius:22px;
overflow:hidden;
margin:24px 0;
box-shadow:var(--shadow-soft);
}
table{
width:100%;border-collapse:collapse;font-size:.87rem
}
th,td{padding:14px 15px;text-align:left;border-bottom:1px solid var(--line)}
th{
font-size:.71rem;text-transform:uppercase;letter-spacing:.07em;color:var(--muted);
background:rgba(255,255,255,.38);
}
body.dark th{background:rgba(255,255,255,.035)}
tr:last-child td{border-bottom:none}
td:first-child{font-weight:700;color:var(--ink)}
.delta-high{color:#b34c3e;font-weight:800}
.delta-med{color:#9a6900;font-weight:800}
.delta-low{color:#2f7c68;font-weight:800}
.bars{display:grid;gap:14px;margin-top:18px}
.bar-row{display:grid;grid-template-columns:130px minmax(0,1fr) 72px;gap:12px;align-items:center}
.bar-label{font-size:.82rem;font-weight:700}
.bar-track{height:10px;border-radius:999px;background:rgba(38,91,113,.10);overflow:hidden}
.bar-fill{
width:0;height:100%;border-radius:999px;
background:linear-gradient(90deg,var(--teal),var(--cyan),var(--blue));
transition:width 1s cubic-bezier(.2,.8,.2,1);
}
.bar-val{font-size:.78rem;text-align:right;color:var(--muted)}
.timeline{
position:relative;margin:28px 0;
}
.timeline::before{
content:"";position:absolute;left:16px;top:8px;bottom:8px;width:2px;background:var(--line)
}
.timeline-card{
margin:0 0 16px 46px;
border-radius:20px;padding:18px 20px;
box-shadow:var(--shadow-soft);
position:relative;
}
.timeline-card::before{
content:"";position:absolute;left:-38px;top:22px;width:16px;height:16px;border-radius:50%;
background:linear-gradient(145deg,var(--teal),var(--blue));
border:3px solid var(--bg2);
box-shadow:0 0 0 4px rgba(46,182,223,.12);
}
.timeline-card .time{font-size:.72rem;text-transform:uppercase;letter-spacing:.08em;color:#2581a1;font-weight:800;margin-bottom:6px}
body.dark .timeline-card .time{color:#73d1ee}
.timeline-card strong{font-size:.96rem}
.timeline-card p{font-size:.86rem;line-height:1.55;margin:6px 0 0}
.flow{
display:grid;grid-template-columns:repeat(4,1fr);gap:12px;margin:24px 0
}
.flow-item{
min-height:138px;
padding:18px;
border-radius:20px;
background:rgba(255,255,255,.34);
border:1px solid rgba(255,255,255,.55);
position:relative;
}
body.dark .flow-item{background:rgba(255,255,255,.035);border-color:rgba(255,255,255,.08)}
.flow-item .verb{font-size:.7rem;letter-spacing:.08em;text-transform:uppercase;font-weight:850;color:#2b86a4}
body.dark .flow-item .verb{color:#72d4f2}
.flow-item h3{margin:10px 0 6px;font-size:1rem}
.flow-item p{font-size:.79rem;line-height:1.5;margin:0}
.flow-item.retired{opacity:.6}
.flow-item.retired::after{
content:"Retiring";
position:absolute;right:10px;top:10px;
font-size:.62rem;font-weight:850;text-transform:uppercase;letter-spacing:.06em;
padding:5px 7px;border-radius:999px;background:rgba(220,92,98,.12);color:#b0474d;
}
.quote{
font-size:clamp(1.45rem,3vw,2.25rem);
line-height:1.25;
letter-spacing:-.035em;
color:var(--ink);
margin:30px 0;
padding-left:24px;
border-left:3px solid var(--teal);
font-weight:760;
}
.chip-row{display:flex;gap:8px;flex-wrap:wrap;margin:12px 0 0}
.chip{
padding:7px 10px;border-radius:999px;font-size:.72rem;font-weight:720;color:#1f708d;
background:rgba(56,177,206,.10);border:1px solid rgba(56,177,206,.15)
}
body.dark .chip{color:#80d5ef}
.footnotes{
padding:25px;border-radius:24px;
border:1px solid var(--glass-border);background:var(--glass);
backdrop-filter:blur(var(--blur));-webkit-backdrop-filter:blur(var(--blur));
box-shadow:var(--shadow-soft);
}
.footnotes ol{margin:10px 0 0;padding-left:22px}
.footnotes li{color:var(--muted);font-size:.86rem;line-height:1.55;margin-bottom:8px}
footer{
width:min(calc(100% - 34px),var(--content));
margin:30px auto 50px;
padding:30px 0 0;
border-top:1px solid var(--line);
display:flex;justify-content:space-between;gap:20px;flex-wrap:wrap;
color:var(--muted);font-size:.82rem;
}
.ask-shell{
position:relative;
overflow:hidden;
border-radius:34px;
padding:1px;
margin-top:26px;
background:linear-gradient(135deg,rgba(255,255,255,.9),rgba(77,185,233,.24),rgba(83,218,184,.24),rgba(255,255,255,.72));
box-shadow:0 30px 90px rgba(33,111,145,.18);
}
.ask-shell::before{
content:"";
position:absolute;width:360px;height:360px;border-radius:50%;
left:-120px;top:-160px;
background:rgba(93,221,238,.24);filter:blur(20px);pointer-events:none;
}
.ask-shell::after{
content:"";
position:absolute;width:320px;height:320px;border-radius:50%;
right:-120px;bottom:-190px;
background:rgba(78,145,255,.19);filter:blur(20px);pointer-events:none;
}
.ask-inner{
position:relative;z-index:2;
border-radius:33px;
padding:24px;
background:rgba(245,253,255,.52);
backdrop-filter:blur(34px) saturate(185%);
-webkit-backdrop-filter:blur(34px) saturate(185%);
border:1px solid rgba(255,255,255,.7);
}
body.dark .ask-inner{background:rgba(7,22,31,.62);border-color:rgba(198,237,255,.12)}
.ask-top{
display:flex;align-items:center;justify-content:space-between;gap:14px;margin-bottom:18px;
}
.bot-id{display:flex;align-items:center;gap:12px}
.bot-orb{
width:48px;height:48px;border-radius:16px;position:relative;
background:
radial-gradient(circle at 30% 23%,white 0 7%,transparent 8%),
linear-gradient(145deg,#7ee9d6,#42b9e5 50%,#4f87ff);
box-shadow:inset 0 1px 0 rgba(255,255,255,.8),0 10px 30px rgba(52,164,203,.28);
}
.bot-orb::after{
content:"✦";position:absolute;inset:0;display:grid;place-items:center;color:white;font-size:20px;
text-shadow:0 2px 8px rgba(0,60,100,.22);
}
.bot-id strong{display:block;font-size:.98rem}
.bot-id span{display:block;color:var(--muted);font-size:.76rem;margin-top:3px}
.online-pill{
flex:none;padding:7px 10px;border-radius:999px;font-size:.68rem;font-weight:800;
color:#267c68;background:rgba(71,204,169,.11);border:1px solid rgba(71,204,169,.2)
}
.chat-window{
height:440px;overflow:auto;padding:18px;
border-radius:25px;
background:rgba(255,255,255,.29);
border:1px solid rgba(255,255,255,.55);
box-shadow:inset 0 1px 0 rgba(255,255,255,.45);
scroll-behavior:smooth;
}
body.dark .chat-window{background:rgba(255,255,255,.025);border-color:rgba(255,255,255,.075)}
.msg{display:flex;margin:0 0 14px;animation:msgIn .28s ease both}
.msg.user{justify-content:flex-end}
.bubble{
max-width:min(82%,690px);padding:13px 15px;border-radius:19px;
font-size:.88rem;line-height:1.55;white-space:pre-wrap;
}
.msg.bot .bubble{
color:var(--ink);background:rgba(255,255,255,.66);
border:1px solid rgba(255,255,255,.78);
border-bottom-left-radius:7px;
box-shadow:0 8px 25px rgba(44,102,126,.08);
}
body.dark .msg.bot .bubble{background:rgba(255,255,255,.07);border-color:rgba(255,255,255,.09)}
.msg.user .bubble{
color:white;background:linear-gradient(135deg,#1e99ca,#34b9c2 55%,#46c3a6);
border-bottom-right-radius:7px;
box-shadow:0 8px 24px rgba(31,153,188,.22);
}
@keyframes msgIn{from{opacity:0;transform:translateY(6px) scale(.99)}to{opacity:1;transform:none}}
.typing{display:inline-flex;gap:4px;align-items:center;height:18px}
.typing i{width:5px;height:5px;border-radius:50%;background:var(--muted);animation:dot 1s infinite ease-in-out}
.typing i:nth-child(2){animation-delay:.12s}.typing i:nth-child(3){animation-delay:.24s}
@keyframes dot{50%{transform:translateY(-4px);opacity:.45}}
.suggestions{display:flex;gap:8px;overflow:auto;padding:13px 2px 4px;scrollbar-width:none}
.suggestions::-webkit-scrollbar{display:none}
.suggestion{
white-space:nowrap;cursor:pointer;border:1px solid rgba(255,255,255,.7);
background:rgba(255,255,255,.42);color:var(--ink);
padding:9px 12px;border-radius:999px;font-size:.75rem;font-weight:700;
transition:.18s ease;
}
body.dark .suggestion{background:rgba(255,255,255,.045);border-color:rgba(255,255,255,.09)}
.suggestion:hover{transform:translateY(-1px);background:rgba(255,255,255,.7)}
.composer{
display:grid;grid-template-columns:1fr auto;gap:10px;margin-top:12px;
padding:7px;border-radius:21px;background:rgba(255,255,255,.52);
border:1px solid rgba(255,255,255,.76);
box-shadow:0 10px 30px rgba(35,99,126,.08);
}
body.dark .composer{background:rgba(255,255,255,.045);border-color:rgba(255,255,255,.09)}
.composer textarea{
resize:none;min-height:46px;max-height:130px;padding:13px 12px;border:0;outline:0;
background:transparent;color:var(--ink);font:inherit;font-size:.9rem;
}
.send-btn{
align-self:end;width:46px;height:46px;border:0;border-radius:15px;cursor:pointer;
color:white;font-size:1.15rem;font-weight:900;
background:linear-gradient(145deg,#218fc6,#37bfc0 55%,#4ac5a8);
box-shadow:0 8px 20px rgba(35,153,184,.24);
transition:.18s ease;
}
.send-btn:hover{transform:scale(1.04)}
.send-btn:active{transform:scale(.96)}
.bot-disclaimer{font-size:.72rem;color:var(--muted);line-height:1.45;margin:11px 4px 0}
.fun-strip{
display:grid;grid-template-columns:repeat(3,1fr);gap:10px;margin-top:16px;
}
.fun-tile{
padding:14px;border-radius:18px;background:rgba(255,255,255,.31);
border:1px solid rgba(255,255,255,.52);font-size:.78rem;color:var(--muted)
}
body.dark .fun-tile{background:rgba(255,255,255,.03);border-color:rgba(255,255,255,.075)}
.fun-tile strong{display:block;margin-bottom:3px;font-size:.82rem}
.nav-tab-btn,.back-blog{
border:1px solid var(--glass-border);cursor:pointer;color:var(--ink);
background:rgba(255,255,255,.34);font-weight:750;
box-shadow:inset 0 1px 0 rgba(255,255,255,.4);
}
.nav-tab-btn{padding:10px 13px;border-radius:14px;font-size:.82rem;margin-left:4px}
.nav-tab-btn:hover,.back-blog:hover{background:rgba(255,255,255,.58)}
body.dark .nav-tab-btn,body.dark .back-blog{background:rgba(255,255,255,.045)}
body.dark .nav-tab-btn:hover,body.dark .back-blog:hover{background:rgba(255,255,255,.08)}
.assistant-view{
display:none;
width:min(calc(100% - 34px),var(--content));
margin:54px auto 80px;
min-height:calc(100vh - 180px);
}
body.assistant-mode .assistant-view{display:block}
body.assistant-mode .hero,
body.assistant-mode > .stats,
body.assistant-mode > .layout,
body.assistant-mode > footer{display:none}
.assistant-view-head{display:flex;align-items:center;gap:14px;flex-wrap:wrap;margin-bottom:14px}
.back-blog{padding:10px 14px;border-radius:14px}
body.assistant-mode .nav-wrap{position:sticky}
body.assistant-mode .ask-shell{max-width:980px;margin:24px auto 0}
body.assistant-mode .assistant-view > h2,
body.assistant-mode .assistant-view > p{max-width:980px;margin-left:auto;margin-right:auto}
body.assistant-mode .assistant-view-head{max-width:980px;margin-left:auto;margin-right:auto}
.desktop-only{display:inline}
.mobile-menu{display:none}
@media (max-width: 980px){
:root{--blur:22px}
.hero-grid{grid-template-columns:1fr}
.hero-side{min-height:0}
.stats{grid-template-columns:repeat(2,1fr)}
.layout{grid-template-columns:1fr}
.toc{display:none}
.flow{grid-template-columns:repeat(2,1fr)}
.nav-links a{display:none}
.nav-links{display:flex}
.nav-tab-btn{margin:0}
.mobile-menu{display:block}
}
@media (max-width: 640px){
.nav-wrap{top:8px;margin-top:8px;width:calc(100% - 16px)}
.nav{min-height:58px;border-radius:20px}
.brand-mark{width:34px;height:34px}
.icon-btn{width:38px;height:38px;border-radius:13px}
.nav-tab-btn{padding:9px 10px;font-size:.74rem}
.hero{margin-top:38px;width:calc(100% - 20px)}
.layout,.stats,footer{width:calc(100% - 20px)}
.assistant-view{width:calc(100% - 20px);margin-top:32px}
.ask-inner{padding:12px;border-radius:26px}
.ask-shell{border-radius:27px}
.chat-window{height:56vh;min-height:360px;padding:12px;border-radius:20px}
.bubble{max-width:91%;font-size:.86rem}
.ask-top{align-items:flex-start}
.online-pill{font-size:.6rem}
.fun-strip{display:none}
.composer{position:sticky;bottom:8px;z-index:5}
.hero{margin-top:48px}
.hero-main{padding:34px 24px 30px;border-radius:28px}
.hero-side{border-radius:28px}
h1{font-size:clamp(3rem,18vw,4.55rem)}
.stats{grid-template-columns:1fr 1fr;gap:10px}
.stat-card{padding:16px}
.stat-card .value{font-size:1.32rem}
.split,.flow,.fun-strip{grid-template-columns:1fr}
.bar-row{grid-template-columns:102px minmax(0,1fr) 56px;gap:8px}
.table-wrap{overflow-x:auto}
table{min-width:620px}
.nav{padding-left:12px}
.brand span{display:none}
.brand strong{font-size:.86rem}
.hero-actions .btn{width:100%;justify-content:center}
section{margin-bottom:58px}
}
@media (min-width: 1500px){
:root{--content:1320px}
.layout{grid-template-columns:280px minmax(0,1fr);gap:42px}
.hero-main{padding:62px}
.chat-window{height:520px}
}
@media (prefers-reduced-motion: reduce){
*,*::before,*::after{animation:none!important;transition:none!important;scroll-behavior:auto!important}
}
.hero-main,.hero-side{border-radius:24px!important;box-shadow:0 12px 32px rgba(39,88,112,.09)}
.eyebrow{border-radius:10px}.stat-card{border-radius:16px}.glass-card,.callout,.table-wrap,.timeline-card{border-radius:18px}
.btn{border-radius:11px}.arch-row{border-radius:12px}.arch-row .tag{border-radius:7px}.toc{border-radius:18px}
.section-kicker{letter-spacing:.055em}.quote{font-weight:650}
.quick-tools{
position:fixed;right:16px;bottom:16px;z-index:90;
display:flex;gap:8px;padding:7px;border-radius:16px;
background:var(--glass);border:1px solid var(--glass-border);
backdrop-filter:blur(20px) saturate(160%);box-shadow:var(--shadow-soft)
}
.quick-tools button,.assistant-actions button{
border:1px solid var(--glass-border);background:rgba(255,255,255,.35);color:var(--ink);
border-radius:10px;cursor:pointer;font-weight:750
}
.quick-tools button{width:38px;height:38px}
body.dark .quick-tools button,body.dark .assistant-actions button{background:rgba(255,255,255,.045)}
.lab-grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:16px;margin-top:26px}
.lab-card{
background:var(--glass);border:1px solid var(--glass-border);backdrop-filter:blur(22px) saturate(150%);
border-radius:18px;padding:20px;box-shadow:var(--shadow-soft);min-width:0
}
.span-2{grid-column:span 2}
.lab-head{display:flex;justify-content:space-between;gap:12px;align-items:flex-start;margin-bottom:16px}
.lab-head h3{margin:7px 0 0}
.evidence{display:inline-block;font-size:.61rem;font-weight:850;letter-spacing:.08em;padding:5px 7px;border-radius:7px;border:1px solid var(--line)}
.observed{color:#157a63;background:rgba(53,180,143,.08)}
.ablation{color:#8a6200;background:rgba(204,153,42,.08)}
.planned{color:#316db1;background:rgba(70,132,203,.08)}
.illustrative{color:#7c5a9e;background:rgba(133,88,173,.08)}
.range{width:100%;accent-color:var(--cyan)}
.tm-scale{display:flex;justify-content:space-between;gap:6px;font-size:.65rem;color:var(--muted);margin:8px 0 15px}
.metric-bars{display:grid;gap:10px}
.metric-row{display:grid;grid-template-columns:95px 1fr 70px;gap:10px;align-items:center;font-size:.78rem}
.metric-track{height:10px;background:rgba(30,80,100,.09);border-radius:999px;overflow:hidden}
.metric-fill{height:100%;background:var(--cyan);border-radius:999px;transition:width .35s ease}
.micro{font-size:.76rem;line-height:1.5;color:var(--muted);margin:12px 0 0}
.ablation-switches{display:grid;gap:9px}
.ablation-switches label{display:flex;gap:9px;align-items:center;font-size:.84rem}
.ablation-meter{height:12px;border-radius:999px;background:rgba(30,80,100,.09);overflow:hidden;margin:16px 0}
.ablation-meter div{height:100%;width:14%;background:var(--cyan);transition:.25s}
.status-live{font-size:.66rem;font-weight:850;color:#21816b}
.status-list{display:grid;gap:0;margin:0}
.status-list div{display:flex;justify-content:space-between;gap:18px;padding:9px 0;border-bottom:1px solid var(--line)}
.status-list dt{color:var(--muted);font-size:.76rem}.status-list dd{margin:0;font-size:.79rem;font-weight:700;text-align:right}
.diff{margin:0;background:rgba(9,29,39,.94);color:#e9f5fa;padding:17px;border-radius:13px;line-height:1.8;overflow:auto;font-size:.8rem}
.diff .minus{color:#ff9b9b}.diff .plus{color:#8fe1b9}.diff .same{color:#d9e6ec}.diff .maybe{color:#f0cd7a}
.token-demo{display:flex;gap:8px}.token-demo input{flex:1;min-width:0}
.token-demo input,.token-demo button,.assistant-mode-select{
border:1px solid var(--glass-border);background:rgba(255,255,255,.35);color:var(--ink);border-radius:10px;padding:10px 12px
}
body.dark .token-demo input,body.dark .token-demo button,body.dark .assistant-mode-select{background:rgba(255,255,255,.045)}
.token-demo button{cursor:pointer;font-weight:750}
.token-path{display:flex;gap:8px;flex-wrap:wrap;margin-top:16px}
.token-node{padding:9px 10px;border:1px solid var(--line);border-radius:10px;font-size:.74rem;opacity:.35;transform:translateY(4px);transition:.3s}
.token-node.on{opacity:1;transform:none;background:rgba(54,175,205,.08)}
.experts{display:grid;grid-template-columns:repeat(4,1fr);gap:8px}
.expert{aspect-ratio:1;border-radius:12px;border:1px solid var(--line);display:grid;place-items:center;font-size:.72rem;font-weight:800;background:rgba(255,255,255,.2)}
.expert.hot{background:rgba(53,185,204,.18);box-shadow:inset 0 0 0 1px rgba(53,185,204,.18)}
.family-tree{display:flex;align-items:center;justify-content:space-between;gap:8px;flex-wrap:wrap}
.family-tree div{padding:12px;border:1px solid var(--line);border-radius:12px;display:flex;flex-direction:column;gap:3px;min-width:105px}
.family-tree span{font-size:.7rem;color:var(--muted)}.family-tree i{color:var(--muted)}.family-tree .future{border-style:dashed}
.repro summary{cursor:pointer;font-weight:800}.repro-grid{display:grid;grid-template-columns:1fr 1fr;gap:12px;margin-top:16px}
.repro-grid div{padding:12px;border:1px solid var(--line);border-radius:12px}.repro-grid b,.repro-grid span{display:block}.repro-grid span{font-size:.76rem;color:var(--muted);margin-top:4px}
.assistant-mode-select{font-size:.7rem;padding:6px 8px;margin-left:8px}
.assistant-actions{display:flex;gap:8px;flex-wrap:wrap;margin-top:10px}.assistant-actions button{padding:8px 10px;font-size:.72rem}
.modal-backdrop{position:fixed;inset:0;background:rgba(5,12,18,.38);backdrop-filter:blur(10px);z-index:300}
.modal-panel{position:fixed;left:50%;top:14%;transform:translateX(-50%);width:min(680px,calc(100% - 24px));z-index:310;background:var(--glass-strong);border:1px solid var(--glass-border);backdrop-filter:blur(30px) saturate(170%);border-radius:18px;box-shadow:0 30px 80px rgba(0,0,0,.22);padding:14px}
.palette input,.search-panel input{width:100%;border:1px solid var(--line);background:transparent;color:var(--ink);border-radius:12px;padding:14px;font-size:1rem;outline:0}
.palette-item,.search-result{padding:11px 12px;border-radius:10px;cursor:pointer;border-bottom:1px solid var(--line)}
.palette-item:hover,.palette-item.active,.search-result:hover{background:rgba(255,255,255,.28)}
.palette-help{font-size:.68rem;color:var(--muted);padding:9px 5px 2px}
.modal-title{display:flex;justify-content:space-between;align-items:center;font-weight:850;margin:2px 2px 12px}
.modal-title button{border:0;background:none;color:var(--ink);font-size:1.4rem;cursor:pointer}
.change-item{padding:12px 4px;border-top:1px solid var(--line)}.change-item b,.change-item span{display:block}.change-item span{color:var(--muted);font-size:.78rem;line-height:1.5;margin-top:4px}
.toast{position:fixed;left:50%;bottom:24px;transform:translate(-50%,20px);opacity:0;pointer-events:none;z-index:400;background:#102b3a;color:white;border-radius:10px;padding:10px 14px;font-size:.78rem;transition:.2s}
.toast.show{opacity:1;transform:translate(-50%,0)}
.section-anchor{opacity:0;margin-left:8px;font-size:.7em;color:var(--muted);transition:.15s}.section-anchor:hover{opacity:1}h2:hover .section-anchor{opacity:.65}
mark.search-hit{background:rgba(255,222,91,.55);color:inherit;border-radius:3px}
body.paper-mode{background:#fff!important;color:#111!important}
body.paper-mode .ambient,body.paper-mode .quick-tools,body.paper-mode .nav-wrap{display:none!important}
body.paper-mode .hero-main,body.paper-mode .hero-side,body.paper-mode .stat-card,body.paper-mode .glass-card,body.paper-mode .callout,body.paper-mode .table-wrap,body.paper-mode .timeline-card,body.paper-mode .lab-card{background:white!important;box-shadow:none!important;backdrop-filter:none!important;border:1px solid #ddd!important}
body.paper-mode{--ink:#111;--muted:#555;--line:#ddd;--glass:#fff;--glass-border:#ddd}
body.no-glass *{backdrop-filter:none!important;-webkit-backdrop-filter:none!important}
body.no-animations *{animation:none!important;transition:none!important}
body.no-radius *{border-radius:0!important}
body.no-css{font-family:serif!important;background:white!important;color:black!important}
body.no-css *{box-shadow:none!important;background:transparent!important;color:inherit!important;border-radius:0!important;backdrop-filter:none!important}
@media(max-width:760px){
.lab-grid{grid-template-columns:1fr}.span-2{grid-column:auto}.repro-grid{grid-template-columns:1fr}
.quick-tools{right:8px;bottom:8px}.quick-tools button{width:34px;height:34px}
.tm-scale{font-size:.55rem}.metric-row{grid-template-columns:78px 1fr 55px}.family-tree i{display:none}
.modal-panel{top:8%}
}
@media print{
.nav-wrap,.quick-tools,.ambient,.progress,.assistant-view,.hero-actions{display:none!important}
body{background:#fff!important}.hero-main,.hero-side,.stat-card,.glass-card,.callout,.table-wrap,.timeline-card,.lab-card{box-shadow:none!important;backdrop-filter:none!important;background:white!important}
}
</style>
</head>
<body>
<div class="progress" id="progress"></div>
<div class="ambient one"></div>
<div class="ambient two"></div>
<div class="ambient three"></div>
<div class="nav-wrap">
<nav class="nav" aria-label="Primary navigation">
<a class="brand" href="#top" aria-label="Orion home">
<span class="brand-mark" aria-hidden="true"></span>
<span>
<strong>Orion / Project Prism</strong>
<span>Transformer 2 research notes</span>
</span>
</a>
<div class="nav-links">
<a href="#hypothesis">Hypothesis</a>
<a href="#mini">Mini</a>
<a href="#nano">Nano</a>
<a href="#scale">2B scale-up</a>
<a href="#t22">T2.2</a>
<button class="nav-tab-btn" id="openAssistant" type="button">✦ AI Assistant</button>
<button class="nav-tab-btn" id="openPalette" type="button" title="Command palette">⌘K</button>
</div>
<button class="icon-btn" id="themeToggle" aria-label="Switch to dark mode" title="Switch appearance"><span id="themeIcon">☾</span></button>
</nav>
</div>
<div class="quick-tools" aria-label="Research tools">
<button id="searchBtn" title="Search research">⌕</button>
<button id="paperModeBtn" title="Paper / print mode">▤</button>
<button id="ablateSiteBtn" title="Ablate the website">Δ</button>
<button id="changelogBtn" title="Changelog">↻</button>
</div>
<header class="hero" id="top">
<div class="hero-grid">
<div class="hero-main">
<div class="eyebrow"><span class="pulse-dot"></span> Ongoing independent ML research</div>
<h1>Transformer 2<br><span class="soft">Research Notes</span></h1>
<p class="dek">
Notes from the Orion / Project Prism experiments on separating persistent knowledge,
reusable procedures, temporary state, and recurrent computation. This page records the architecture,
measurements, negative results, and changes made between T2 generations.
</p>
<div class="hero-actions">
<a class="btn primary" href="#hypothesis">Start with the hypothesis</a>
<a class="btn secondary" href="#scale">Current 2B experiment</a>
</div>
</div>
<aside class="hero-side" aria-label="Current architecture summary">
<div>
<div class="mini-title">Current working model</div>
<div class="architecture-list">
<div class="arch-row">
<div class="tag">KNOW</div>
<div><strong>Distributed weights</strong><span>Persistent factual knowledge</span></div>
</div>
<div class="arch-row">
<div class="tag">DO</div>
<div><strong>Procedure Banks</strong><span>Reusable conditional transformations</span></div>
</div>
<div class="arch-row">
<div class="tag">THINK</div>
<div><strong>Working State</strong><span>Temporary computation and context</span></div>
</div>
<div class="arch-row">
<div class="tag">VERIFY</div>
<div><strong>Recurrent deliberation</strong><span>Revisit, integrate, refine</span></div>
</div>
</div>
</div>
<div class="hero-note">
The Knowledge Vault was part of the original design. Across the Orion variants tested so far,
it has not earned its parameter and complexity budget.
</div>
</aside>
</div>
</header>
<section class="stats" aria-label="Project at a glance">
<div class="stat-card">
<div class="label">Current generation</div>
<div class="value">T2.1</div>
<div class="sub">T2.2 in design</div>
</div>
<div class="stat-card">
<div class="label">Current scale</div>
<div class="value">~2B</div>
<div class="sub">parameters</div>
</div>
<div class="stat-card">
<div class="label">Training target</div>
<div class="value">≈100B</div>
<div class="sub">tokens</div>
</div>
<div class="stat-card">
<div class="label">Hardware</div>
<div class="value">8× H200</div>
<div class="sub">NVIDIA GPUs</div>
</div>
</section>
<div class="layout">
<aside class="toc" aria-label="Table of contents">
<h3>Contents</h3>
<a href="#hypothesis">01 — The original hypothesis</a>
<a href="#mini">02 — Mini-T2</a>
<a href="#mini-ablation">03 — Mini ablations</a>
<a href="#nano">04 — Nano / T2.1</a>
<a href="#nano-findings">05 — Nano findings</a>
<a href="#benchmark">06 — Benchmark result</a>
<a href="#vault">07 — The Vault problem</a>
<a href="#scale">08 — The ~2B scale-up</a>
<a href="#telemetry">09 — Early telemetry</a>
<a href="#dynamics">10 — Training dynamics</a>
<a href="#t22">11 — T2.2 direction</a>
<a href="#vm">12 — Built-in VM idea</a>
<a href="#science">13 — What is actually proven?</a>
<a href="#lab">14 — Interactive research lab</a>
</aside>
<main>
<section id="hypothesis">
<div class="section-kicker">01 / Original hypothesis</div>
<h2>The original hypothesis</h2>
<p class="lead">
Orion / Project Prism is an experimental architecture family I call <strong>Transformer 2 (T2)</strong>.
Its goal is deliberately aggressive: <strong>maximum capability, minimum parameters</strong>.
</p>
<p>
The project started from the suspicion that a standard Transformer may spend capacity inefficiently by
mixing several different jobs into the same residual stream and weight system. Persistent facts, reusable
algorithms, temporary reasoning, and verification are not obviously the same kind of thing — so the first
T2 designs tried to give them different homes.
</p>
<div class="flow">
<div class="flow-item retired">
<div class="verb">Know</div>
<h3>Knowledge Vault</h3>
<p>Explicit persistent retrieval memory for factual/declarative information.</p>
</div>
<div class="flow-item">
<div class="verb">Do</div>
<h3>Procedure Banks</h3>
<p>Conditional learned specialists intended to encode reusable transformations.</p>
</div>
<div class="flow-item">
<div class="verb">Think</div>
<h3>Working State</h3>
<p>A separate bounded stream for temporary computation and intermediate reasoning.</p>
</div>
<div class="flow-item">
<div class="verb">Verify</div>
<h3>Deliberation</h3>
<p>Recurrent computation that revisits deeper stages instead of always moving forward once.</p>
</div>
</div>
<div class="callout">
<div class="title">The updated hypothesis</div>
<p>
Persistent knowledge can probably stay distributed. Temporary computation may deserve explicit state,
and reusable computation may deserve explicit conditional pathways.
</p>
</div>
<p>
That sentence is much narrower than the original vision. That is the point. T2 is not one frozen
architecture; it is an iterative research program in which components have to justify the weights,
compute, and engineering complexity they consume.
</p>
</section>
<section id="mini">
<div class="section-kicker">02 / Generation one</div>
<h2>Mini-T2: separating KNOW, DO, and THINK</h2>
<p class="lead">
The original Mini-T2 was about <strong>219.1M parameters</strong>. It combined a recurrent/shared
Universal Cortex with explicit Procedure Banks, a Knowledge Vault, a separate causal Working State,
and a recurrent deliberation tail.
</p>
<div class="split">
<div class="subcard">
<h3>Core network</h3>
<p>12 shared/recurrent stages, <strong>d<sub>model</sub> = 768</strong>, 12 query heads and 4 KV heads.</p>
</div>
<div class="subcard">
<h3>Procedure pathway</h3>
<p>3 Procedure Banks, each with 16 experts, top-2 routing plus a shared expert, hidden size ≈1024.</p>
</div>
<div class="subcard">
<h3>Knowledge pathway</h3>
<p>65,536-slot Vault with product-key retrieval, top-4 reads near stages 3, 7, and 11.</p>
</div>
<div class="subcard">
<h3>Temporary state</h3>
<p>Separate causal Working State with reads/updates near stages 3, 7, and 11.</p>
</div>
</div>
<h3>Bounded state updates</h3>
<p>
The Working State was not an unbounded additive scratchpad. Its update used gated replacement, which
gives the model a way to preserve old state or replace it with a candidate while keeping the magnitude
controlled.
</p>
<div class="equation">state_new = (1 − write_gate) · state + write_gate · candidate</div>
<p>
A learned read gate later projected this state back into the main residual stream. The first Mini-T2
therefore implemented the original conceptual split quite literally: <strong>KNOW → Vault,
DO → Procedure, THINK → State.</strong>
</p>
</section>
<section id="mini-ablation">
<div class="section-kicker">03 / Mini ablations</div>
<h2>Mini ablations changed the story</h2>
<p class="lead">
One early set of inference-time ablations made the Working State look much more important than the Vault
for that trained checkpoint.
</p>
<div class="table-wrap">
<table>
<thead>
<tr><th>Configuration</th><th>Facts</th><th>Reason</th><th>Language</th><th>Log-prob style result</th></tr>
</thead>
<tbody>
<tr><td>Full</td><td>60</td><td>66.7</td><td>100</td><td>-2.979 / -1.576 / -3.686</td></tr>
<tr><td>No Vault</td><td>60</td><td>73.3</td><td>95</td><td>-2.871 / -1.569 / -3.701</td></tr>
<tr><td>No State</td><td>23.3</td><td>16.7</td><td>50</td><td>-10.319 / -9.867 / -8.222</td></tr>
<tr><td>No Deliberation</td><td>60</td><td>83.3</td><td>95</td><td>-3.462 / -2.122 / -4.399</td></tr>
</tbody>
</table>
</div>
<div class="callout warn">
<div class="title">Important causal limitation</div>
<p>
This shows that the <em>trained checkpoint</em> relied heavily on Working State. It does not prove that
the architecture is fundamentally better than a conventional Transformer. A matched model retrained
without the state pathway would be needed for that stronger claim.
</p>
</div>
<p>
The Vault, meanwhile, showed weak and mixed evidence. This was not enough to remove it immediately,
but it was the first sign that the original KNOW/DO/THINK decomposition might be over-structured.
</p>
</section>
<section id="nano">
<div class="section-kicker">04 / Generation two</div>
<h2>Orion Flagship Nano / T2.1</h2>
<p class="lead">
The next generation became smaller, more conditional, and more careful about routing. Orion Flagship
Nano T2.1 had approximately <strong>134.715M total parameters</strong> and an estimated
<strong>72.812M active-path parameter-equivalent</strong>.
</p>
<div class="glass-card">
<div class="split" style="margin:0">
<div class="subcard">
<h3>Backbone</h3>
<p>d<sub>model</sub> 640, 12 stages, 10 attention heads, 2 KV heads.</p>
</div>
<div class="subcard">
<h3>Procedure routing</h3>
<p>3 banks × 12 specialists, top-1 routing, small shared expert, and a soft NULL route.</p>
</div>
<div class="subcard">
<h3>Conditioning</h3>
<p>Router sees token representation, projected Working State, and stage/cycle information.</p>
</div>
<div class="subcard">
<h3>Other subsystems</h3>
<p>Working State, Vault v2, and recurrent deliberation over the final three stages.</p>
</div>
</div>
</div>
<p>
New subsystems were initialized close to neutral so training did not begin with aggressive,
potentially destabilizing side paths.
</p>
<h3>Training data</h3>
<p>The official base corpus was 50B tokens with no SFT, DPO, or RLHF:</p>
<div class="bars" data-bars>
<div class="bar-row"><div class="bar-label">FineWeb-Edu</div><div class="bar-track"><div class="bar-fill" data-width="60"></div></div><div class="bar-val">60% · 30B</div></div>
<div class="bar-row"><div class="bar-label">FineMath</div><div class="bar-track"><div class="bar-fill" data-width="12"></div></div><div class="bar-val">12% · 6B</div></div>
<div class="bar-row"><div class="bar-label">OpenWebMath</div><div class="bar-track"><div class="bar-fill" data-width="8"></div></div><div class="bar-val">8% · 4B</div></div>
<div class="bar-row"><div class="bar-label">CodeParrot Clean</div><div class="bar-track"><div class="bar-fill" data-width="20"></div></div><div class="bar-val">20% · 10B</div></div>
</div>
<div class="chip-row">
<span class="chip">60% general</span><span class="chip">20% math</span><span class="chip">20% code</span>
</div>
</section>
<section id="nano-findings">
<div class="section-kicker">05 / T2.1 learning dynamics</div>
<h2>The Procedure Banks arrived first. State arrived later.</h2>
<p class="lead">
The most interesting Nano result was not one giant number. It was a repeated pattern across checkpoints.
Procedure routing mattered very early; Working State became much more important only after the model had
learned more basic language structure.
</p>
<div class="timeline">
<div class="timeline-card">
<div class="time">~1k steps · CE ≈ 4.81</div>
<strong>Procedure already matters; Vault almost does not.</strong>
<p>Δvault +0.0012 · Δstate +0.0332 · Δprocedure +2.9562 · Δdeliberation +0.0806</p>
</div>
<div class="timeline-card">
<div class="time">~4.36B tokens · CE ≈ 2.60</div>
<strong>Working State becomes a major dependency.</strong>
<p>Δvault +0.1622 · Δstate +4.4783 · Δprocedure +5.8667 · Δdeliberation +0.6068</p>
</div>
<div class="timeline-card">
<div class="time">~9.4B tokens</div>
<strong>Procedure and State dominate the ablation picture.</strong>
<p>Δvault +0.2111 · Δstate +6.1979 · Δprocedure +6.6810 · Δdeliberation +1.0009</p>
</div>
<div class="timeline-card">
<div class="time">~10.55B tokens</div>
<strong>The qualitative ordering remains stable.</strong>
<p>Δvault +0.1579 · Δstate +5.8374 · Δprocedure +6.4856 · Δdeliberation +0.8026</p>
</div>
<div class="timeline-card">
<div class="time">Nearby checkpoint</div>
<strong>Same pattern, larger magnitudes.</strong>
<p>Δvault +0.1140 · Δstate +7.1968 · Δprocedure +8.7902 · Δdeliberation +0.7189</p>
</div>
</div>
<div class="callout warn">
<div class="title">Tiny ablation batches are noisy</div>
<p>
The exact deltas are strongly prompt-dependent. I care more about the repeated cross-checkpoint pattern
— Procedure and State matter a lot, Deliberation matters some, Vault matters comparatively little —
than I do about any one headline value.
</p>
</div>
</section>
<section id="benchmark">
<div class="section-kicker">06 / Nano benchmark</div>
<h2>Nano’s BananaMind result</h2>
<p>
On BananaMind Base Bench 1.1, a 350-example multiple-choice benchmark, Nano reached an overall Elo of
<strong>1041</strong> at 9.99B training tokens, with <strong>56.86% raw accuracy</strong> and
<strong>53.26% weighted accuracy</strong>. That exceeded the older 219.1M Mini model’s 1030 Elo while
Nano itself was only about 134.7M parameters.
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Category</th><th>Elo</th></tr></thead>
<tbody>
<tr><td>Language</td><td>1331</td></tr>
<tr><td>Commonsense</td><td>932</td></tr>
<tr><td>Knowledge</td><td>1040</td></tr>
<tr><td>Context</td><td>963</td></tr>
<tr><td>Quantitative</td><td>872</td></tr>
<tr><td>Logic</td><td>1045</td></tr>
<tr><td>Code</td><td>1203</td></tr>
</tbody>
</table>
</div>
<p>
Later benchmark scaling was not monotonic: the overall Elo stayed around 1041 at 13.62B tokens and was
around 1019 in the 27B-token era. That does <strong>not</strong> justify the conclusion that later training
made the model generally worse. A small fixed benchmark is too narrow for that.
</p>
</section>
<section id="vault">
<div class="section-kicker">07 / Knowledge Vault</div>
<h2>The Knowledge Vault kept failing to earn its weights</h2>
<p class="lead">
Across Orion generations and scales, the explicit neural Knowledge Vault repeatedly produced much weaker
causal signals than Procedure Banks or Working State.
</p>
<div class="quote">
“Across the Orion designs tested so far, this explicit neural Knowledge Vault has not earned its parameter
and complexity budget.”
</div>
<p>
That is deliberately <strong>not</strong> the same as saying external memory is useless. The claim is only
about this particular learned Vault design, in the Orion models tested so far.
</p>
<div class="split">
<div class="subcard">
<h3>Original view</h3>
<p><strong>KNOW</strong> → dedicated Vault<br><strong>DO</strong> → Procedure Banks<br><strong>THINK</strong> → Working State</p>
</div>
<div class="subcard">
<h3>Current view</h3>
<p><strong>KNOW</strong> → distributed weights<br><strong>DO</strong> → Procedure Banks<br><strong>THINK</strong> → Working State<br><strong>VERIFY</strong> → deliberation</p>
</div>
</div>
</section>
<section id="scale">
<div class="section-kicker">08 / Current experiment</div>
<h2>Scaling T2.1 to ~2B parameters</h2>
<p class="lead">
The current Orion run is an approximately <strong>2B-parameter T2.1 model</strong> training on
<strong>8× NVIDIA H200 GPUs</strong> toward a target of 99,999,907,840 tokens — effectively 100B.
</p>
<div class="stats" style="width:100%;margin:24px 0 28px">
<div class="stat-card">
<div class="label">Updates</div>
<div class="value">871,930</div>
<div class="sub">target total</div>
</div>
<div class="stat-card">
<div class="label">Tokens / update</div>
<div class="value">114,688</div>
<div class="sub">global update</div>
</div>
<div class="stat-card">
<div class="label">Throughput</div>
<div class="value">160–167k</div>
<div class="sub">tokens / second</div>
</div>
<div class="stat-card">
<div class="label">Peak VRAM</div>
<div class="value">~72.4 GiB</div>
<div class="sub">reported device/process context</div>
</div>
</div>
<p>
So far the run has reported no allocation retries, no skipped steps, and essentially zero data-loader
wait. This larger model has <strong>six Procedure Banks</strong>. The Vault was deliberately shrunk and
the freed parameter budget reallocated elsewhere, but it was not fully removed in this run to avoid
introducing an extra architectural risk at the same time as the scale jump.
</p>
</section>
<section id="telemetry">
<div class="section-kicker">09 / Early 2B telemetry</div>
<h2>The first billion tokens looked familiar</h2>
<p>
Around step 8,600 (~986M tokens), training CE was about 2.57 and probe CE about 2.21. The Vault gate was
already nearly closed, while the Working State and Procedure pathway showed materially stronger activity.
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Subsystem / signal</th><th>Observed value</th><th>Reading</th></tr></thead>
<tbody>
<tr><td>Vault gate</td><td>~0.003 ± 0.022</td><td>Almost closed</td></tr>
<tr><td>Vault unique slots</td><td>~342 / 16,384</td><td>Low utilization</td></tr>
<tr><td>Vault gradient</td><td>~1.35e-2</td><td>Small relative signal</td></tr>
<tr><td>Working State read</td><td>~0.107</td><td>Active</td></tr>
<tr><td>Working State RMS</td><td>~0.4030</td><td>Non-trivial state magnitude</td></tr>
<tr><td>State gradient</td><td>~5.47e-2</td><td>Stronger than Vault</td></tr>
<tr><td>Procedure gradient</td><td>~1.49e-1</td><td>Strong</td></tr>
<tr><td>Cortex gradient</td><td>~1.81e-1</td><td>Strong baseline pathway</td></tr>
</tbody>
</table>
</div>
<h3>Depth-dependent Procedure usage</h3>
<p>
Procedure routing had zero dead experts. The soft-NULL probability showed a clear depth pattern:
the earliest bank was mostly bypassed, while deeper banks were used far more often.
</p>
<div class="bars" data-bars>
<div class="bar-row"><div class="bar-label">B0 NULL</div><div class="bar-track"><div class="bar-fill" data-width="94.3"></div></div><div class="bar-val">0.943</div></div>
<div class="bar-row"><div class="bar-label">B1 NULL</div><div class="bar-track"><div class="bar-fill" data-width="85.3"></div></div><div class="bar-val">0.853</div></div>
<div class="bar-row"><div class="bar-label">B2 NULL</div><div class="bar-track"><div class="bar-fill" data-width="72"></div></div><div class="bar-val">0.720</div></div>
<div class="bar-row"><div class="bar-label">B3 NULL</div><div class="bar-track"><div class="bar-fill" data-width="62"></div></div><div class="bar-val">0.620</div></div>
<div class="bar-row"><div class="bar-label">B4 NULL</div><div class="bar-track"><div class="bar-fill" data-width="44.1"></div></div><div class="bar-val">0.441</div></div>
<div class="bar-row"><div class="bar-label">B5 NULL</div><div class="bar-track"><div class="bar-fill" data-width="53.4"></div></div><div class="bar-val">0.534</div></div>
</div>
<h3>At ~1.6–1.7B tokens</h3>
<p>
The first serious 2B ablations were striking. At step ~14,000, removing Procedure increased probe CE by
+11.1971. At step 15,000 / ~1.720B tokens, the Procedure delta rose to +13.8884, while the Vault effect
remained effectively zero.
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Checkpoint</th><th>Full CE</th><th>Δ Vault</th><th>Δ State</th><th>Δ Procedure</th><th>Δ Deliberation</th></tr></thead>
<tbody>
<tr><td>~step 14,000</td><td>3.0327</td><td class="delta-low">+0.0021</td><td class="delta-med">+0.6982</td><td class="delta-high">+11.1971</td><td class="delta-med">+0.4908</td></tr>
<tr><td>step 15,000</td><td>2.7877</td><td class="delta-low">−0.0030</td><td class="delta-med">+0.8092</td><td class="delta-high">+13.8884</td><td class="delta-med">+0.4310</td></tr>
</tbody>
</table>
</div>
<div class="callout danger">
<div class="title">Do not turn +13.9 CE into an “intelligence percentage.”</div>
<p>
Removing a major subsystem can push internal activations far out of the distribution seen during
training. The defensible conclusion is narrower: <strong>this trained checkpoint is extraordinarily
dependent on its Procedure pathway.</strong>
</p>
</div>
</section>
<section id="dynamics">
<div class="section-kicker">10 / Training dynamics</div>
<h2>Working State may not have “arrived” yet</h2>
<p class="lead">
In previous Orion generations, Working State often became much more important only after the model had
learned enough basic language structure to make use of explicit temporary computation.
</p>
<p>
At ~1.7B out of ~100B intended tokens, the 2B run is only around <strong>1.7% complete</strong>.
The current State ablation effect of roughly +0.8 CE is already meaningful, but the more important question
is whether it grows over the next several billion tokens as it did in Nano.
</p>
<p>
That trajectory will directly influence how much of the T2.2 parameter budget is assigned to Working State.
</p>
</section>
<section id="t22">
<div class="section-kicker">11 / T2.2 design notes</div>
<h2>T2.2 is becoming simpler, not more ornate</h2>
<p>
The provisional T2.2 direction follows the evidence instead of preserving every original idea:
</p>
<div class="glass-card">
<div class="flow" style="margin:0">
<div class="flow-item">
<div class="verb">Increase</div>
<h3>Procedure capacity</h3>
<p>More budget for the pathway that has shown the strongest repeated causal importance.</p>
</div>
<div class="flow-item">
<div class="verb">Retain</div>
<h3>Working State</h3>
<p>Substantial, mid-sized temporary state; exact budget depends on later 2B dynamics.</p>
</div>
<div class="flow-item retired">
<div class="verb">Remove</div>
<h3>Knowledge Vault</h3>
<p>Persistent knowledge goes back into ordinary distributed model weights.</p>
</div>
<div class="flow-item">
<div class="verb">Keep</div>
<h3>Deliberation</h3>
<p>Recurrent/shared computation, state-conditioned routing, and neutral initialization remain.</p>
</div>
</div>
</div>
<p>
The architecture is therefore moving away from the idea that every conceptual function needs a dedicated
neural subsystem. A subsystem now survives only if its measured contribution justifies its cost.
</p>
</section>
<section id="vm">
<div class="section-kicker">12 / Proposed exact-compute path</div>
<h2>A possible built-in exact-compute path</h2>
<p>
One future T2.2 idea is to add a deterministic computation path — effectively a small built-in VM or
exact-compute subsystem. This would not replace neural reasoning; it would separate exact operations from
fuzzy learned transformations.
</p>
<div class="flow">
<div class="flow-item">
<div class="verb">Weights</div>
<h3>Persistent knowledge</h3>
<p>Facts and representations remain distributed in normal model parameters.</p>
</div>
<div class="flow-item">
<div class="verb">Procedure</div>
<h3>Learned transforms</h3>
<p>Reusable neural routines selected conditionally by the model.</p>
</div>
<div class="flow-item">
<div class="verb">State</div>
<h3>Fuzzy scratch space</h3>
<p>Temporary reasoning, intermediate features, and contextual computation.</p>
</div>
<div class="flow-item">
<div class="verb">VM</div>
<h3>Exact computation</h3>
<p>Arithmetic, symbolic operations, deterministic algorithms, or potentially code execution.</p>
</div>
</div>
<p>
A future router could choose among NULL, shared neural computation, a specialist Procedure path, and the
deterministic VM/tool path. This remains a research idea, not a demonstrated part of Orion.
</p>
</section>
<section id="science">
<div class="section-kicker">13 / Evidence and limitations</div>
<h2>What the current evidence does — and does not — prove</h2>
<p class="lead">
Inference-time ablations answer an important question: <em>what did this particular trained checkpoint
learn to rely on?</em> They do not automatically answer: <em>was this architecture the best way to spend
the same compute and parameters?</em>
</p>
<div class="split">
<div class="subcard">
<h3>Supported now</h3>
<p>
Procedure pathways are a very strong causal dependency in current checkpoints. Working State has
repeatedly become important. Deliberation has a smaller but recurring positive signal. The tested
Vault design remains weak.
</p>
</div>
<div class="subcard">
<h3>Not yet proven</h3>
<p>
That T2 is universally superior to ordinary Transformers, that external memory is useless, or that
any ablation delta corresponds to a literal percentage of intelligence.
</p>
</div>
</div>
<h3>Matched controls that would strengthen the case</h3>
<div class="table-wrap">
<table>
<thead><tr><th>Question</th><th>Needed comparison</th></tr></thead>
<tbody>
<tr><td>Does Working State improve efficiency?</td><td>T2 with State vs matched model retrained without State</td></tr>
<tr><td>Are Procedure Banks better than dense FFNs?</td><td>Sparse Procedure Banks vs matched dense FFN at similar compute/parameter budget</td></tr>
<tr><td>Does recurrence help?</td><td>Shared recurrent layers vs unique layers at matched FLOPs</td></tr>
<tr><td>Was the Vault worth its budget?</td><td>Vault vs equal parameter budget added to ordinary weights</td></tr>
<tr><td>Would exact compute help?</td><td>VM vs no-VM, with VM execution compute reported separately</td></tr>
</tbody>
</table>
</div>
<p>
The evaluation should track validation CE, downstream benchmarks, total and active parameters,
FLOPs/token, latency, VRAM, stability, subsystem utilization, ablation effects, capability per parameter,
and capability per FLOP.
</p>
<div class="quote">Every subsystem has to earn its weights.</div>
<div class="footnotes">
<strong>Research status</strong>
<ol>
<li>The ~2B T2.1 model is an ongoing training run, not a finished result.</li>
<li>Small ablation batches are noisy and prompt-dependent.</li>
<li>Repeated patterns across generations are more informative than any single dramatic delta.</li>
<li>BananaMind is one benchmark; it is not evidence of general intelligence by itself.</li>
<li>The current Vault conclusion applies to the tested Orion Vault designs, not to memory architectures in general.</li>
</ol>
</div>
</section>
<section id="lab">
<div class="section-kicker">14 / Interactive research lab</div>
<h2>Explore the architecture</h2>
<p class="lead">Interactive views of the same measurements used in the research notes. Where a view is illustrative rather than measured, it is labelled as such.</p>
<div class="lab-grid">
<article class="lab-card span-2">
<div class="lab-head">
<div><span class="evidence observed">OBSERVED</span><h3>Checkpoint Time Machine</h3></div>
<span class="mono" id="tmLabel">Mini</span>
</div>
<input id="timeMachine" class="range" type="range" min="0" max="4" step="1" value="0" aria-label="Checkpoint">
<div class="tm-scale"><span>Mini</span><span>Nano 1k</span><span>Nano 4.36B</span><span>Nano 9.4B</span><span>2B 1.72B</span></div>
<div class="metric-bars" id="tmBars"></div>
<p class="micro" id="tmNote"></p>
</article>
<article class="lab-card">
<div class="lab-head"><div><span class="evidence ablation">ABLATION</span><h3>Ablation Playground</h3></div><span class="mono" id="ablCE">CE 2.7877</span></div>
<div class="ablation-switches">
<label><input type="checkbox" data-ablate="vault" checked> Vault</label>
<label><input type="checkbox" data-ablate="state" checked> Working State</label>
<label><input type="checkbox" data-ablate="procedure" checked> Procedure</label>
<label><input type="checkbox" data-ablate="delib" checked> Deliberation</label>
</div>
<div class="ablation-meter"><div id="ablFill"></div></div>
<p class="micro" id="ablText">All measured subsystems enabled.</p>
</article>
<article class="lab-card">
<div class="lab-head"><div><span class="evidence observed">OBSERVED</span><h3>Experiment status</h3></div><span class="status-live">● RUNNING</span></div>
<dl class="status-list">
<div><dt>Model</dt><dd>~2B T2.1</dd></div>
<div><dt>Target</dt><dd>99,999,907,840 tokens</dd></div>
<div><dt>Hardware</dt><dd>8× H200</dd></div>
<div><dt>Throughput</dt><dd>~160–167k tok/s</dd></div>
<div><dt>Latest supplied point</dt><dd>~1.72B tokens</dd></div>
</dl>
<p class="micro">Static snapshot from the supplied research notes. Ready to swap for a live telemetry endpoint later.</p>
</article>
<article class="lab-card span-2">
<div class="lab-head"><div><span class="evidence planned">PLANNED</span><h3>T2.1 → T2.2 architecture diff</h3></div></div>
<pre class="diff"><span class="minus">- Knowledge Vault (explicit persistent memory)</span>
<span class="plus">+ More Procedure capacity</span>
<span class="plus">+ Mid-sized Working State budget</span>
<span class="same"> Recurrent/shared computation</span>
<span class="same"> Recurrent deliberation</span>
<span class="same"> State-conditioned routing</span>
<span class="maybe">? Deterministic VM / exact-compute path</span></pre>
</article>
<article class="lab-card span-2">
<div class="lab-head"><div><span class="evidence illustrative">ILLUSTRATIVE</span><h3>Inside one token</h3></div></div>
<div class="token-demo">
<input id="tokenInput" value="The capital of France is" aria-label="Illustrative token prompt">
<button id="runToken">Run illustrative trace</button>
</div>
<div class="token-path" id="tokenPath"></div>
<p class="micro">This animation is a conceptual walkthrough, not a recorded causal trace from a live checkpoint.</p>
</article>
<article class="lab-card">
<div class="lab-head"><div><span class="evidence observed">OBSERVED</span><h3>Procedure Bank visualizer</h3></div><select id="bankSelect" aria-label="Procedure bank"></select></div>
<div class="experts" id="experts"></div>
<p class="micro" id="bankNote"></p>
</article>
<article class="lab-card">
<div class="lab-head"><div><span class="evidence observed">OBSERVED</span><h3>Model family tree</h3></div></div>
<div class="family-tree">
<div><b>Mini</b><span>219.1M</span></div><i>→</i>
<div><b>Nano T2.1</b><span>134.7M</span></div><i>→</i>
<div><b>2B T2.1</b><span>~2B</span></div><i>→</i>
<div class="future"><b>T2.2</b><span>designing</span></div>
</div>
</article>
<article class="lab-card span-2">
<details class="repro">
<summary>Reproducibility drawer</summary>
<div class="repro-grid">
<div><b>Data</b><span>Nano: 50B tokens; 60% general, 20% math, 20% code</span></div>
<div><b>2B target</b><span>99,999,907,840 tokens</span></div>
<div><b>Hardware</b><span>8× NVIDIA H200</span></div>
<div><b>Tokens/update</b><span>114,688</span></div>
<div><b>Reported throughput</b><span>~160k–167k tokens/s</span></div>
<div><b>Scientific caveat</b><span>Inference ablations show checkpoint dependence, not matched-training superiority.</span></div>
</div>
</details>
</article>
</div>
</section>
</main>
</div>
<section id="ask-orion" class="assistant-view" aria-hidden="true">
<div class="assistant-view-head">
<button class="back-blog" id="backToBlog" type="button">← Research blog</button>
<div class="section-kicker">AI Assistant / Interactive architecture guide</div>
</div>
<h2>Ask our pattern-matching chatbot</h2>
<p class="lead">
Ask about Orion, T2.1, the ablations, Procedure Banks, Working State, the Vault, training data,
the current ~2B run, or the emerging T2.2 design. This is not a tiny list of canned question/answer pairs:
it scores your question against the blog's research knowledge, combines relevant evidence, and generates
a fresh response in your browser.
</p>
<div class="ask-shell">
<div class="ask-inner">
<div class="ask-top">
<div class="bot-id">
<div class="bot-orb" aria-hidden="true"></div>
<div><strong>Prism Pattern Matcher</strong><span>Local research guide · no server required</span></div>
<select id="assistantMode" class="assistant-mode-select" title="Assistant mode">
<option value="pattern">Pattern Matcher</option>
<option value="orion" disabled>Orion 2B · connect later</option>
</select>
</div>
<div class="online-pill">● READY</div>
</div>
<div class="chat-window" id="chatWindow" aria-live="polite">
<div class="msg bot"><div class="bubble">Hey! 👋 I’m the Orion architecture guide. I answer by matching your question against the research notes on this page and composing the most relevant evidence. Try asking why the Vault is being removed, what the Procedure Banks do, or whether the ablations prove T2 beats a normal Transformer.</div></div>
</div>
<div class="suggestions" id="suggestions">
<button class="suggestion">Why is the Vault being removed?</button>
<button class="suggestion">What do Procedure Banks do?</button>
<button class="suggestion">Does +13.9 CE mean 90% of intelligence?</button>
<button class="suggestion">How is Working State different from weights?</button>
<button class="suggestion">What is planned for T2.2?</button>
</div>
<div class="composer">
<textarea id="chatInput" rows="1" placeholder="Ask about the architecture…" aria-label="Ask Orion"></textarea>
<button class="send-btn" id="sendChat" aria-label="Send question">↑</button>
</div>
<div class="fun-strip">
<div class="fun-tile"><strong>🧠 Evidence-aware</strong>Distinguishes observations, ablations, hypotheses, and unproven claims.</div>
<div class="fun-tile"><strong>⚡ Instant</strong>Runs entirely inside this single HTML file.</div>
<div class="fun-tile"><strong>🧪 Research mode</strong>Will say when the current notes do not support an answer.</div>
</div>
<div class="assistant-actions">
<button id="researchModeBtn">Research mode: OFF</button>
<button id="exportChatBtn">Export Markdown</button>
<button id="shareChatBtn">Copy share link</button>
</div>
<div class="bot-disclaimer">
Pattern-matching assistant, not a hosted LLM. It can handle varied/rephrased architecture questions by
retrieving and combining relevant project facts, but it does not invent new experimental results.
</div>
</div>
</div>
</section>
<footer>
<div><strong>Orion / Project Prism</strong> · Transformer 2 research notes</div>
<div>Independent architecture research · T2.1 → T2.2</div>
</footer>
<script>
(() => {
const body = document.body;
const toggle = document.getElementById('themeToggle');
const progress = document.getElementById('progress');
const tocLinks = [...document.querySelectorAll('.toc a')];
const sections = [...document.querySelectorAll('main section')];
const themeIcon = document.getElementById('themeIcon');
const safeGetTheme = () => { try { return localStorage.getItem('orion-theme'); } catch(e) { return null; } };
const safeSetTheme = v => { try { localStorage.setItem('orion-theme', v); } catch(e) {} };
const applyTheme = dark => {
body.classList.toggle('dark', dark);
themeIcon.textContent = dark ? '☀︎' : '☾';
toggle.setAttribute('aria-label', dark ? 'Switch to light mode' : 'Switch to dark mode');
document.querySelector('meta[name="theme-color"]').setAttribute('content', dark ? '#08131c' : '#eaf8ff');
};
const saved = safeGetTheme();
applyTheme(saved ? saved === 'dark' : window.matchMedia('(prefers-color-scheme: dark)').matches);
toggle.addEventListener('click', () => {
const next = !body.classList.contains('dark');
applyTheme(next);
safeSetTheme(next ? 'dark' : 'light');
});
const updateScroll = () => {
const max = document.documentElement.scrollHeight - window.innerHeight;
const pct = max > 0 ? (window.scrollY / max) * 100 : 0;
progress.style.width = pct + '%';
let current = '';
for (const sec of sections) {
const rect = sec.getBoundingClientRect();
if (rect.top <= 150) current = sec.id;
}
tocLinks.forEach(a => a.classList.toggle('active', a.getAttribute('href') === '#' + current));
};
window.addEventListener('scroll', updateScroll, {passive:true});
updateScroll();
const openAssistant = document.getElementById('openAssistant');
const backToBlog = document.getElementById('backToBlog');
const assistantView = document.getElementById('ask-orion');
function setView(mode, push=true){
const assistant = mode === 'assistant';
body.classList.toggle('assistant-mode', assistant);
assistantView.setAttribute('aria-hidden', assistant ? 'false' : 'true');
document.title = assistant ? 'AI Assistant — Orion / Project Prism' : 'Orion / Project Prism — Transformer 2 Research Blog';
if(push) history.pushState({view:mode},'',assistant ? '#assistant' : '#top');
window.scrollTo({top:0,behavior:'smooth'});
}
openAssistant.addEventListener('click',()=>setView('assistant'));
backToBlog.addEventListener('click',()=>setView('blog'));
window.addEventListener('popstate',()=>setView(location.hash === '#assistant' ? 'assistant' : 'blog',false));
if(location.hash === '#assistant') setView('assistant',false);
const observer = new IntersectionObserver(entries => {
entries.forEach(entry => {
if (!entry.isIntersecting) return;
entry.target.querySelectorAll('.bar-fill').forEach(el => {
el.style.width = el.dataset.width + '%';
});
observer.unobserve(entry.target);
});
}, {threshold:.2});
document.querySelectorAll('[data-bars]').forEach(el => observer.observe(el));
// ---- Prism Pattern Matcher -------------------------------------------------
// Lightweight local retrieval + response composition. The knowledge units are
// propositions rather than fixed question/answer pairs, so rephrased questions
// can retrieve multiple relevant facts and receive a newly assembled response.
const knowledge = [
{tags:"goal t2 transformer 2 orion project prism purpose capability parameters efficient", text:"Orion / Project Prism is the Transformer 2 (T2) research program. Its core goal is maximum capability with minimum parameters."},
{tags:"original hypothesis know knowledge do procedure think working state verify deliberation split", text:"The original hypothesis separated persistent knowledge, reusable procedures, temporary reasoning, and verification into different pathways: Vault, Procedure Banks, Working State, and recurrent deliberation."},
{tags:"updated hypothesis current view distributed weights knowledge persistent", text:"The current view is that persistent factual knowledge can probably remain distributed in ordinary model weights rather than requiring a dedicated Vault."},
{tags:"procedure bank banks experts specialists router routing do reusable transformation conditional compute", text:"Procedure Banks are conditional learned specialist pathways intended to represent reusable neural transformations. T2.1 routing can depend on the current token representation, projected Working State, and stage/cycle information."},
{tags:"procedure importance ablation 2b 13.8884 11.1971 catastrophic dependency", text:"In the early ~2B run, Procedure was the dominant causal dependency on the small ablation probes: around step 14,000 Δprocedure was +11.1971 CE, and at step 15,000 it was +13.8884 CE."},
{tags:"procedure caveat intelligence percentage out distribution ablation proof", text:"A large Procedure ablation delta must not be interpreted as a literal percentage of intelligence. Removing a major trained subsystem can push activations out of distribution; the defensible conclusion is that the checkpoint is extraordinarily dependent on its Procedure pathway."},
{tags:"working state state think scratchpad temporary computation reasoning gate bounded replacement", text:"Working State is a separate causal stream for temporary computation. Mini-T2 updated it with bounded gated replacement: state_new = (1 − write_gate) × state + write_gate × candidate, then projected it back through a learned read gate."},
{tags:"working state importance emerges later training dynamics language nano", text:"Across earlier Orion training, Working State tended to become much more important only after basic language structure had emerged. This makes its importance over training time a key measurement for the current ~2B run."},
{tags:"vault knowledge memory remove removed retiring useless weak gate t2.2", text:"The explicit Knowledge Vault is likely to be removed in T2.2 because across the tested Orion designs it repeatedly showed a much weaker contribution than Procedure or Working State and has not earned its parameter and complexity budget."},
{tags:"vault external memory universal failure proven useless caveat", text:"The Vault result is not evidence that external memory architectures are universally useless. It is a negative result about the specific neural Vault designs tested in Orion so far."},
{tags:"mini original 219.1m 12 stages 768 16 experts 65536 slots", text:"Original Mini-T2 was about 219.1M parameters, with 12 recurrent/shared stages, d_model 768, three 16-expert Procedure Banks, a 65,536-slot Knowledge Vault, Working State, and recurrent deliberation."},
{tags:"mini ablation no state facts reason language causal", text:"In one Mini checkpoint, removing Working State sharply damaged facts, reasoning, and language scores, while removing the Vault had little effect. This established checkpoint dependence, not architectural superiority."},
{tags:"nano t2.1 134.715m active 72.812m parameters configuration", text:"Orion Flagship Nano T2.1 had about 134.715M total parameters and about 72.812M active-path parameter-equivalent, with d_model 640, 12 stages, three Procedure Banks, Working State, Vault v2, and recurrent final-three-stage deliberation."},
{tags:"data corpus fineweb edu finemath openwebmath codeparrot 50b training", text:"Nano's 50B-token base corpus was 60% FineWeb-Edu (30B), 12% FineMath (6B), 8% OpenWebMath (4B), and 20% CodeParrot Clean (10B): overall 60% general, 20% math, 20% code, with no SFT, DPO, or RLHF in the base corpus."},
{tags:"bananamind benchmark elo accuracy nano 1041 56.86 53.26", text:"At 9.99B training tokens, Nano scored about 1041 overall Elo on BananaMind Base Bench 1.1, with 56.86% raw accuracy and 53.26% weighted accuracy. This is one benchmark result, not proof of universal architectural superiority or general intelligence."},
{tags:"benchmark later 13.62b 27b monotonic worse", text:"Benchmark scaling was not monotonic: overall Elo stayed around 1041 at 13.62B tokens and was around 1019 in the 27B-token era. A small fixed benchmark is not enough to claim later training made the model generally worse."},
{tags:"2b current training h200 8 gpu 100b tokens throughput 160k 167k vram", text:"The current T2.1 scale-up is approximately 2B parameters on 8× NVIDIA H200 GPUs, targeting about 100B tokens. Reported throughput is roughly 160k–167k tokens/s, with peak reported VRAM around 72.4 GiB per device/process context."},
{tags:"2b health allocation retries skipped data loader six procedure banks", text:"The ~2B run has six Procedure Banks and has reported no allocation retries, no skipped steps, and essentially zero data-loader wait in the supplied telemetry."},
{tags:"2b early telemetry vault gate state gradients null b0 depth", text:"Around ~986M tokens, the Vault gate was about 0.003 ± 0.022, while Procedure and Working State had stronger gradient signals. Procedure routing had zero dead experts, and the earliest bank B0 had a very high soft-NULL probability around 0.943."},
{tags:"2b step 15000 1.72b ablation full 2.7877 vault state procedure deliberation", text:"At step 15,000 / ~1.720B tokens, full probe CE was 2.7877; Δvault −0.0030, Δstate +0.8092, Δprocedure +13.8884, and Δdeliberation +0.4310."},
{tags:"2b early percent complete 1.7 training tokens", text:"At roughly 1.7B of the intended ~100B tokens, the current model was only about 1.7% through the planned training run, so these results should be treated as early training dynamics rather than final outcomes."},
{tags:"deliberation recurrent verify usefulness", text:"Recurrent deliberation repeats deeper computation so the model can integrate or refine results. It has shown a smaller but recurring positive ablation signal relative to Procedure and, later in training, Working State."},
{tags:"t2.2 future design larger procedure mid state no vault recurrent deliberation", text:"The provisional T2.2 direction is a larger Procedure subsystem, a mid-sized Working State, no explicit Knowledge Vault, retained recurrent/shared computation and deliberation, state-conditioned conditional compute, neutral initialization, and extensive telemetry."},
{tags:"vm built in exact deterministic compute arithmetic symbolic code tool future router null", text:"A possible future T2.2 experiment is a deterministic built-in VM/exact-compute path for arithmetic, symbolic operations, deterministic algorithms, or possibly code execution. This is an idea, not a validated architecture component."},
{tags:"scientific evidence causality matched controls retraining superiority prove", text:"Inference-time ablations demonstrate what a trained checkpoint depends on. Stronger architectural claims require matched retrained controls at similar parameter and compute budgets."},
{tags:"metrics flops latency vram validation ce capability parameter compute evaluation", text:"A stronger evaluation should report validation CE, downstream benchmarks, total and active parameters, FLOPs/token, latency, VRAM, training stability, component utilization, ablations, capability per parameter, and capability per FLOP."}
];
const stop = new Set("a an the is are was were be been being do does did of to in on at for from with and or but if then than it its this that these those what why how who when where which can could would should about our your my me i we they them their as by into over under vs versus much many mean means tell explain".split(" "));
const synonyms = {
memory:"vault", memories:"vault", knowledge:"vault weights",
scratchpad:"state", reasoning:"state deliberation", think:"state",
expert:"procedure", experts:"procedure", moe:"procedure", specialist:"procedure",
procedures:"procedure", bank:"procedure", banks:"procedure",
remove:"vault t2.2", removed:"vault t2.2", retire:"vault t2.2", retired:"vault t2.2",
gpu:"h200", gpus:"h200", hardware:"h200", speed:"throughput",
dataset:"data corpus", training:"training tokens", future:"t2.2",
proof:"causality matched controls", prove:"causality matched controls",
better:"superiority matched controls", transformer:"transformer t2",
math:"finemath openwebmath vm", code:"codeparrot vm",
ablations:"ablation", ablate:"ablation"
};
const norm = str => str.toLowerCase().replace(/[^a-z0-9.+×~-]+/g," ").trim().split(/\s+/).filter(Boolean);
function expand(tokens){
const out = [...tokens];
tokens.forEach(t => { if(synonyms[t]) out.push(...synonyms[t].split(" ")); });
return out.filter(t => !stop.has(t));
}
function scoreUnit(qTokens, unit){
const hay = new Set(norm(unit.tags + " " + unit.text));
let score = 0;
qTokens.forEach(t => {
if(hay.has(t)) score += t.length > 6 ? 3 : 2;
else if(t.length > 4 && [...hay].some(h => h.includes(t) || t.includes(h))) score += 0.8;
});
return score;
}
function composeAnswer(question){
const q = question.toLowerCase();
const qt = expand(norm(question));
const ranked = knowledge.map(k => ({...k, score:scoreUnit(qt,k)})).sort((a,b)=>b.score-a.score);
const useful = ranked.filter(x=>x.score>0).slice(0, q.includes("compare") || q.includes("difference") ? 4 : 3);
if(!useful.length || useful[0].score < 1.5){
return "I can’t support a confident answer to that from the Orion research notes currently embedded in this page. Try asking about T2/T2.1, Procedure Banks, Working State, the Knowledge Vault, deliberation, Nano, BananaMind, the ~2B H200 run, ablations, T2.2, or the proposed VM.";
}
let prefix = "";
if(q.includes("prove") || q.includes("better") || q.includes("superior"))
prefix = "Important distinction: the current evidence does not by itself prove architectural superiority. ";
else if(q.includes("vault") && (q.includes("why") || q.includes("remove") || q.includes("retir")))
prefix = "The short version: the Vault kept losing the parameter-budget argument. ";
else if(q.includes("procedure"))
prefix = "Procedure Banks are currently one of the strongest signals in the Orion experiments. ";
else if(q.includes("state"))
prefix = "Working State is Orion’s explicit temporary-computation pathway. ";
else if(q.includes("t2.2"))
prefix = "T2.2 is still provisional, but its direction is increasingly clear. ";
const unique = [];
useful.forEach(u => { if(!unique.includes(u.text)) unique.push(u.text); });
let answer = prefix + unique.join(" ");
if((q.includes("13.9") || q.includes("intelligence") || q.includes("percent")) && !answer.includes("literal percentage"))
answer += " In particular, an ablation delta is not a literal percentage of the model’s intelligence.";
return answer;
}
const chatWindow = document.getElementById('chatWindow');
const chatInput = document.getElementById('chatInput');
const sendChat = document.getElementById('sendChat');
function addMessage(text, who){
const row = document.createElement('div');
row.className = 'msg ' + who;
const bubble = document.createElement('div');
bubble.className = 'bubble';
bubble.textContent = text;
row.appendChild(bubble);
chatWindow.appendChild(row);
chatWindow.scrollTop = chatWindow.scrollHeight;
return row;
}
function ask(text){
text = text.trim();
if(!text) return;
addMessage(text,'user');
chatInput.value = '';
chatInput.style.height = '46px';
const typing = document.createElement('div');
typing.className = 'msg bot';
typing.innerHTML = '<div class="bubble"><span class="typing"><i></i><i></i><i></i></span></div>';
chatWindow.appendChild(typing);
chatWindow.scrollTop = chatWindow.scrollHeight;
const answer = composeAnswer(text);
setTimeout(() => {
typing.remove();
addMessage(answer,'bot');
}, Math.min(780, 260 + text.length * 7));
}
sendChat.addEventListener('click',()=>ask(chatInput.value));
chatInput.addEventListener('keydown',e=>{
if(e.key === 'Enter' && !e.shiftKey){e.preventDefault();ask(chatInput.value);}
});
chatInput.addEventListener('input',()=>{
chatInput.style.height='46px';
chatInput.style.height=Math.min(chatInput.scrollHeight,130)+'px';
});
document.querySelectorAll('.suggestion').forEach(btn=>btn.addEventListener('click',()=>ask(btn.textContent)));
// ---------- Reading time, anchors, search corpus ----------
const articleText = [...document.querySelectorAll('main section')].map(x=>x.innerText).join(' ');
const words = articleText.trim().split(/\s+/).length;
const reading = document.getElementById('readingTime');
if(reading) reading.textContent = Math.max(1, Math.round(words/220)) + ' min read';
document.querySelectorAll('main section h2').forEach(h=>{
const sec=h.closest('section'); if(!sec || !sec.id) return;
const a=document.createElement('a'); a.href='#'+sec.id; a.className='section-anchor'; a.textContent='¶'; a.title='Copy section link';
a.addEventListener('click',e=>{ e.preventDefault(); history.replaceState(null,'','#'+sec.id); navigator.clipboard?.writeText(location.href); showToast('Section link copied'); });
h.appendChild(a);
});
const searchDocs=[...document.querySelectorAll('main section')].map(sec=>({
id:sec.id, title:sec.querySelector('h2')?.childNodes[0]?.textContent?.trim() || sec.id, text:sec.innerText
}));
// ---------- Checkpoint Time Machine ----------
const checkpoints=[
{name:'Mini', note:'Mini checkpoint: Working State was the strongest observed dependency; Vault evidence was weak/mixed.', vals:{Vault:.2,State:9.0,Procedure:0,Deliberation:1.6}},
{name:'Nano · ~1k steps', note:'Procedure mattered almost immediately, while State and Vault were still small.', vals:{Vault:.0012,State:.0332,Procedure:2.9562,Deliberation:.0806}},
{name:'Nano · 4.36B', note:'Working State had become a major dependency by this point.', vals:{Vault:.1622,State:4.4783,Procedure:5.8667,Deliberation:.6068}},
{name:'Nano · 9.4B', note:'Procedure and State were both strongly important; Vault remained much smaller.', vals:{Vault:.2111,State:6.1979,Procedure:6.6810,Deliberation:1.0009}},
{name:'2B · 1.72B', note:'Early 2B checkpoint: Procedure dominated the measured ablation effect.', vals:{Vault:0,State:.8092,Procedure:13.8884,Deliberation:.4310}}
];
const tm=document.getElementById('timeMachine'), tmBars=document.getElementById('tmBars'), tmLabel=document.getElementById('tmLabel'), tmNote=document.getElementById('tmNote');
function renderTM(){
const c=checkpoints[+tm.value]; tmLabel.textContent=c.name; tmNote.textContent=c.note; tmBars.innerHTML='';
const max=Math.max(...Object.values(c.vals),1);
Object.entries(c.vals).forEach(([k,v])=>{
const r=document.createElement('div'); r.className='metric-row';
r.innerHTML=`<b>${k}</b><div class="metric-track"><div class="metric-fill" style="width:${Math.max(1,v/max*100)}%"></div></div><span class="mono">${v===0?'≈0':v.toFixed(v<1?3:2)}</span>`;
tmBars.appendChild(r);
});
}
tm?.addEventListener('input',renderTM); if(tm) renderTM();
// ---------- Ablation playground ----------
const baseCE=2.7877, deltas={vault:-.0030,state:.8092,procedure:13.8884,delib:.4310};
function renderAblation(){
let ce=baseCE, removed=[];
document.querySelectorAll('[data-ablate]').forEach(x=>{ if(!x.checked){ce+=deltas[x.dataset.ablate];removed.push(x.parentElement.textContent.trim())} });
document.getElementById('ablCE').textContent='CE '+ce.toFixed(4);
document.getElementById('ablFill').style.width=Math.min(100,Math.max(5,ce/17*100))+'%';
document.getElementById('ablText').textContent=removed.length ? 'Disabled: '+removed.join(', ')+'. This is a simple composition of the supplied single-subsystem deltas, not a measured multi-ablation result.' : 'All measured subsystems enabled.';
}
document.querySelectorAll('[data-ablate]').forEach(x=>x.addEventListener('change',renderAblation));
// ---------- Token walkthrough ----------
const tokenPath=document.getElementById('tokenPath');
const tokenSteps=['Embedding','Cortex','Procedure B0 · maybe NULL','Working State write','Deeper Cortex','Procedure B4','Deliberation','Next-token logits'];
function resetToken(){ tokenPath.innerHTML=''; tokenSteps.forEach(t=>{const n=document.createElement('div');n.className='token-node';n.textContent=t;tokenPath.appendChild(n)}); }
resetToken();
document.getElementById('runToken')?.addEventListener('click',()=>{
resetToken(); [...tokenPath.children].forEach((n,i)=>setTimeout(()=>n.classList.add('on'),i*220));
});
// ---------- Procedure visualizer ----------
const bankNull=[.943,.853,.720,.620,.441,.534], bankSelect=document.getElementById('bankSelect'), experts=document.getElementById('experts'), bankNote=document.getElementById('bankNote');
if(bankSelect){
bankNull.forEach((v,i)=>{const o=document.createElement('option');o.value=i;o.textContent='Bank B'+i;bankSelect.appendChild(o)});
function renderBank(){
const b=+bankSelect.value, active=1-bankNull[b]; experts.innerHTML='';
for(let i=0;i<12;i++){ const e=document.createElement('div'); e.className='expert '+(i<Math.round(active*12)?'hot':''); e.textContent='E'+i; e.title='Illustrative expert activity display'; experts.appendChild(e); }
bankNote.textContent=`Observed soft-NULL probability for B${b}: ${bankNull[b].toFixed(3)}. Expert highlights are illustrative; the supplied notes only establish zero dead experts and bank-level NULL usage.`;
}
bankSelect.addEventListener('change',renderBank); renderBank();
}
// ---------- Modal / palette / search ----------
const backdrop=document.getElementById('modalBackdrop'), palette=document.getElementById('commandPalette'), searchPanel=document.getElementById('searchPanel'), changelogPanel=document.getElementById('changelogPanel');
const modalPanels=[palette,searchPanel,changelogPanel];
function closeModals(){modalPanels.forEach(x=>x.hidden=true);backdrop.hidden=true}
function openModal(el){closeModals();el.hidden=false;backdrop.hidden=false;setTimeout(()=>el.querySelector('input')?.focus(),30)}
backdrop.addEventListener('click',closeModals); document.querySelectorAll('[data-close-modal]').forEach(b=>b.addEventListener('click',closeModals));
document.addEventListener('keydown',e=>{if(e.key==='Escape')closeModals()});
const commands=[
{label:'Open AI Assistant',action:()=>setView('assistant')},
{label:'Go to current 2B experiment',action:()=>location.hash='scale'},
{label:'Open interactive research lab',action:()=>location.hash='lab'},
{label:'Toggle dark mode',action:()=>toggle.click()},
{label:'Toggle paper mode',action:()=>document.body.classList.toggle('paper-mode')},
{label:'Search research',action:()=>openModal(searchPanel)},
{label:'Print / save PDF',action:()=>window.print()},
{label:'Download RSS snapshot',action:downloadRSS}
];
searchDocs.forEach(d=>commands.push({label:'Section: '+d.title,action:()=>location.hash=d.id}));
const palInput=document.getElementById('paletteInput'), palResults=document.getElementById('paletteResults');
function renderPalette(){
const q=(palInput.value||'').toLowerCase(); palResults.innerHTML='';
commands.filter(c=>c.label.toLowerCase().includes(q)).slice(0,10).forEach((c,i)=>{
const el=document.createElement('div');el.className='palette-item';el.textContent=c.label;el.onclick=()=>{closeModals();c.action()};palResults.appendChild(el)
});
}
palInput.addEventListener('input',renderPalette);
document.getElementById('openPalette')?.addEventListener('click',()=>{openModal(palette);renderPalette()});
document.addEventListener('keydown',e=>{if((e.ctrlKey||e.metaKey)&&e.key.toLowerCase()==='k'){e.preventDefault();openModal(palette);renderPalette()}});
const siteSearchInput=document.getElementById('siteSearchInput'), siteSearchResults=document.getElementById('siteSearchResults');
function renderSearch(){
const q=siteSearchInput.value.trim().toLowerCase();siteSearchResults.innerHTML=''; if(q.length<2)return;
searchDocs.map(d=>({...d,score:(d.title.toLowerCase().includes(q)?4:0)+(d.text.toLowerCase().match(new RegExp(q.replace(/[.*+?^${}()|[\]\\]/g,'\\$&'),'g'))||[]).length}))
.filter(x=>x.score>0).sort((a,b)=>b.score-a.score).slice(0,12).forEach(d=>{
const el=document.createElement('div');el.className='search-result';el.innerHTML=`<b>${d.title}</b><div class="micro">${d.text.slice(0,150).replace(/\s+/g,' ')}…</div>`;
el.onclick=()=>{closeModals();setView('blog',false);location.hash=d.id};siteSearchResults.appendChild(el)
});
}
siteSearchInput.addEventListener('input',renderSearch);
document.getElementById('searchBtn')?.addEventListener('click',()=>openModal(searchPanel));
document.getElementById('changelogBtn')?.addEventListener('click',()=>openModal(changelogPanel));
// ---------- Paper mode + website self-ablation ----------
document.getElementById('paperModeBtn')?.addEventListener('click',()=>document.body.classList.toggle('paper-mode'));
let ablationStage=0;
const siteAblations=[
{cls:'no-glass',msg:'ΔGlass: backdrop blur removed.'},
{cls:'no-animations',msg:'ΔAnimations: motion removed.'},
{cls:'no-radius',msg:'ΔRadius: rounded corners removed.'},
{cls:'no-css',msg:'ΔCSS: website has entered the Times New Roman dimension.'}
];
document.getElementById('ablateSiteBtn')?.addEventListener('click',()=>{
if(ablationStage>=siteAblations.length){
siteAblations.forEach(x=>document.body.classList.remove(x.cls));ablationStage=0;showToast('Website restored. CSS pathway recovered.');return;
}
const a=siteAblations[ablationStage++];document.body.classList.add(a.cls);showToast(a.msg+(ablationStage===4?' The website is extraordinarily dependent on its CSS pathway.':''));
});
// ---------- Assistant research mode / export / share ----------
let researchMode=false;
const researchBtn=document.getElementById('researchModeBtn');
researchBtn?.addEventListener('click',()=>{researchMode=!researchMode;researchBtn.textContent='Research mode: '+(researchMode?'ON':'OFF')});
const oldCompose=composeAnswer;
composeAnswer=function(question){
const a=oldCompose(question);
if(!researchMode)return a;
return `OBSERVATION\n${a}\n\nINTERPRETATION\nThis answer is assembled from the supplied Orion research notes and their repeated qualitative patterns.\n\nLIMITATION\nInference-time ablations establish checkpoint dependence; matched retrained controls are still needed for stronger architecture-level claims.`;
};
document.getElementById('exportChatBtn')?.addEventListener('click',()=>{
const rows=[...document.querySelectorAll('#chatWindow .msg')].map(m=>(m.classList.contains('user')?'## You':'## Prism')+'\n'+m.innerText.trim()).join('\n\n');
const blob=new Blob(['# Orion Assistant Conversation\n\n'+rows],{type:'text/markdown'});const a=document.createElement('a');a.href=URL.createObjectURL(blob);a.download='orion-chat.md';a.click();URL.revokeObjectURL(a.href);
});
document.getElementById('shareChatBtn')?.addEventListener('click',()=>{history.replaceState(null,'','#assistant');navigator.clipboard?.writeText(location.href);showToast('Assistant link copied')});
// slash commands
const originalAsk=ask;
ask=function(text){
const t=text.trim();
if(t.startsWith('/')){
const map={'/architecture':'What is the current Orion architecture?','/ablations':'What have the ablations shown?','/training':'How is the current 2B run training?','/t22':'What is planned for T2.2?'};
if(map[t.toLowerCase()]) return originalAsk(map[t.toLowerCase()]);
}
if(/are you conscious/i.test(t)){addMessage(t,'user');setTimeout(()=>addMessage('bro I am currently JavaScript pattern matching 😭','bot'),180);chatInput.value='';return}
return originalAsk(text);
};
// ---------- RSS snapshot ----------
function downloadRSS(){
const items=searchDocs.slice(0,8).map(d=>`<item><title>${escapeXML(d.title)}</title><description>${escapeXML(d.text.slice(0,280))}</description><guid>#${d.id}</guid></item>`).join('');
const xml=`<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Orion / Project Prism Research Notes</title><description>Transformer 2 research updates</description><link>${escapeXML(location.href)}</link>${items}</channel></rss>`;
const blob=new Blob([xml],{type:'application/rss+xml'});const a=document.createElement('a');a.href=URL.createObjectURL(blob);a.download='orion-research.xml';a.click();URL.revokeObjectURL(a.href);
}
function escapeXML(x){return String(x).replace(/[<>&'"]/g,c=>({'<':'<','>':'>','&':'&',"'":''','"':'"'}[c]))}
// ---------- Toast ----------
let toastTimer;
function showToast(msg){const t=document.getElementById('toast');t.textContent=msg;t.classList.add('show');clearTimeout(toastTimer);toastTimer=setTimeout(()=>t.classList.remove('show'),2200)}
// ---------- Keyboard section navigation ----------
document.addEventListener('keydown',e=>{
if(['INPUT','TEXTAREA','SELECT'].includes(document.activeElement.tagName))return;
if(e.key==='j'||e.key==='k'){
const secs=[...document.querySelectorAll('main section')]; const y=scrollY+160;
let idx=secs.findIndex(x=>x.offsetTop>y); if(idx<0)idx=secs.length;
if(e.key==='k') idx=Math.max(0,idx-2); else idx=Math.min(secs.length-1,idx);
secs[idx]?.scrollIntoView({behavior:'smooth'});
}
});
})();
</script>
<div class="modal-backdrop" id="modalBackdrop" hidden></div>
<div class="palette modal-panel" id="commandPalette" hidden role="dialog" aria-modal="true" aria-label="Command palette">
<input id="paletteInput" placeholder="Search sections or run a command…" autocomplete="off">
<div id="paletteResults"></div>
<div class="palette-help">↑↓ navigate · Enter open · Esc close · Ctrl/⌘ K toggle</div>
</div>
<div class="modal-panel search-panel" id="searchPanel" hidden role="dialog" aria-modal="true">
<div class="modal-title">Search research <button data-close-modal>×</button></div>
<input id="siteSearchInput" placeholder="Search the blog…" autocomplete="off">
<div id="siteSearchResults"></div>
</div>
<div class="modal-panel changelog-panel" id="changelogPanel" hidden role="dialog" aria-modal="true">
<div class="modal-title">Changelog <button data-close-modal>×</button></div>
<div class="change-item"><b>v4 · Interactive Lab</b><span>Checkpoint time machine, ablation playground, family tree, token walkthrough, procedure visualizer, search, command palette, paper mode, self-ablation.</span></div>
<div class="change-item"><b>v3 · Assistant view</b><span>Separate AI tab, responsive mobile layout, persistent dark mode.</span></div>
<div class="change-item"><b>v2 · Pattern Matcher</b><span>Local architecture Q&A with evidence-aware retrieval.</span></div>
<div class="change-item"><b>v1 · Research Blog</b><span>Original Orion/T2 research notes site.</span></div>
</div>
<div class="toast" id="toast" role="status" aria-live="polite"></div>
</body>
</html>
|