File size: 124,073 Bytes
7b742f2 fadbe09 87008d1 fadbe09 87008d1 fadbe09 87008d1 fadbe09 87008d1 fadbe09 87008d1 fadbe09 87008d1 fadbe09 87008d1 fadbe09 7b742f2 | 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 | <!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>IntelligenceLab · Verified training environments for frontier agents</title>
<meta name="description" content="Verified, difficulty-calibrated training environments for terminal, scientific, CAD, web, and visual-to-code agents. Every task ships with a reference solution and an executable verifier.">
<meta property="og:type" content="website">
<meta property="og:title" content="IntelligenceLab · Verified training environments for frontier agents">
<meta property="og:description" content="Long-horizon agent tasks that prove themselves: executed, verified, and calibrated before they reach you.">
<meta name="theme-color" content="#f6f7f4">
<link rel="icon" type="image/svg+xml" href="data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 32 32'%3E%3Cdefs%3E%3ClinearGradient id='g' x1='0' y1='0' x2='1' y2='1'%3E%3Cstop offset='0' stop-color='%2310b981'/%3E%3Cstop offset='1' stop-color='%236366f1'/%3E%3C/linearGradient%3E%3C/defs%3E%3Crect width='32' height='32' rx='8' fill='url(%23g)'/%3E%3Cpath d='M7 23h5v-5h5v-5h5' fill='none' stroke='white' stroke-width='2.6' stroke-linecap='round' stroke-linejoin='round'/%3E%3Cpath d='M19.6 9.4l2.2 2.2 4.2-4.4' fill='none' stroke='white' stroke-width='2.6' stroke-linecap='round' stroke-linejoin='round'/%3E%3C/svg%3E">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&family=JetBrains+Mono:wght@400;500;600&display=swap" rel="stylesheet">
<link rel="stylesheet" href="styles.css">
<script>document.documentElement.classList.add('js');</script>
</head>
<body>
<a class="skip-link" href="#main">Skip to content</a>
<header class="nav">
<div class="container nav-inner">
<a class="brand" href="#top" aria-label="IntelligenceLab home">
<svg class="brand-mark" viewBox="0 0 32 32" aria-hidden="true">
<defs>
<linearGradient id="brand-grad" x1="0" y1="0" x2="1" y2="1">
<stop offset="0" stop-color="#10b981"/>
<stop offset="1" stop-color="#6366f1"/>
</linearGradient>
</defs>
<rect width="32" height="32" rx="8" fill="url(#brand-grad)"/>
<path d="M7 23h5v-5h5v-5h5" fill="none" stroke="#fff" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"/>
<path d="M19.6 9.4l2.2 2.2 4.2-4.4" fill="none" stroke="#fff" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"/>
</svg>
<span>IntelligenceLab</span>
</a>
<nav class="nav-links" aria-label="Primary">
<a href="#offerings">What we provide</a>
<a href="#examples">Examples</a>
<a href="#pipeline">How it works</a>
<a href="#trajectories">Trajectories</a>
<a href="#research">Research</a>
</nav>
<div class="nav-cta">
<a class="btn btn-dark btn-sm" href="#contact">Start a pilot</a>
<button class="nav-toggle" type="button" aria-label="Open menu" aria-expanded="false" aria-controls="mobile-menu">
<span></span><span></span>
</button>
</div>
</div>
<div class="mobile-menu" id="mobile-menu" hidden>
<a href="#offerings">What we provide</a>
<a href="#examples">Examples</a>
<a href="#pipeline">How it works</a>
<a href="#trajectories">Trajectories</a>
<a href="#research">Research</a>
<a class="btn btn-dark" href="#contact">Start a pilot</a>
</div>
</header>
<main id="main">
<section class="hero" id="top" aria-labelledby="hero-title">
<canvas class="hero-canvas" aria-hidden="true"></canvas>
<div class="hero-glow" aria-hidden="true"></div>
<div class="container hero-inner">
<div class="hero-copy">
<p class="eyebrow"><span class="pulse-dot" aria-hidden="true"></span>Terminal · Coding · Multimodal</p>
<h1 id="hero-title">Agent training data that <mark>proves itself</mark>.</h1>
<p class="lede">We provide terminal, coding, and multimodal training data for agents. Every task ships with its own reference solution and executable verifier, and every trajectory is verified and rewritten into one consistent harness, calibrated to the difficulty your model needs next.</p>
<div class="hero-actions">
<a class="btn btn-dark" href="#contact">Start a pilot
<svg viewBox="0 0 20 20" aria-hidden="true"><path d="M4 10h11m-4-4 4 4-4 4" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"/></svg>
</a>
<a class="btn btn-ghost" href="#examples">See examples</a>
</div>
<ul class="hero-proof">
<li><svg viewBox="0 0 20 20" aria-hidden="true"><path d="m5 10.5 3 3 7-7" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/></svg>Verified by execution</li>
<li><svg viewBox="0 0 20 20" aria-hidden="true"><path d="m5 10.5 3 3 7-7" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/></svg>Calibrated to your checkpoint</li>
<li><svg viewBox="0 0 20 20" aria-hidden="true"><path d="m5 10.5 3 3 7-7" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/></svg>Fresh every month</li>
</ul>
</div>
</div>
<ol class="tier-badges" aria-label="Difficulty tiers">
<li style="--tier:#14b88a"><span class="tier-num">T1</span>Warm-up</li>
<li style="--tier:#0da5a0"><span class="tier-num">T2</span>Core</li>
<li style="--tier:#258cd6"><span class="tier-num">T3</span>Hard</li>
<li style="--tier:#5865e2"><span class="tier-num">T4</span>Expert</li>
<li style="--tier:#7c5cdc"><span class="tier-num">T5</span>Frontier</li>
</ol>
</section>
<section class="marquee" aria-label="Software and formats our tasks run on">
<p class="marquee-title">Tasks run on the software practitioners actually use</p>
<div class="marquee-track">
<ul>
<li>CadQuery</li><li>OpenSCAD</li><li>FreeCAD</li><li>KiCad</li><li>Blender</li><li>OpenFOAM</li><li>MuJoCo</li><li>URDF</li><li>Tesseract</li><li>Coq</li><li>CMake</li><li>Git</li><li>Docker</li><li>Nuxt</li><li>Vue</li><li>Flask</li><li>SQLite</li><li>STEP</li>
</ul>
<ul aria-hidden="true">
<li>CadQuery</li><li>OpenSCAD</li><li>FreeCAD</li><li>KiCad</li><li>Blender</li><li>OpenFOAM</li><li>MuJoCo</li><li>URDF</li><li>Tesseract</li><li>Coq</li><li>CMake</li><li>Git</li><li>Docker</li><li>Nuxt</li><li>Vue</li><li>Flask</li><li>SQLite</li><li>STEP</li>
</ul>
</div>
</section>
<section class="section" id="offerings" aria-labelledby="offer-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">What we provide</p>
<h2 id="offer-title">Terminal, coding, and multimodal data.</h2>
<p class="sub">Terminal work is coding work: to finish a terminal task, the agent writes, runs, tests, and debugs real programs. The same verified environments extend to visual inputs, from engineering drawings to rendered pages.</p>
</div>
<div class="offer-grid">
<article class="offer-card reveal">
<div class="offer-top offer-terminal" aria-hidden="true">
<span><b>$</b> make -C services/ingest</span>
<span class="dim"> built ingest-worker</span>
<span><b>$</b> ./bin/migrate --apply v15.yaml</span>
<span><b>$</b> ./bin/replay --since 02:00</span>
<span class="ok">✓ end state verified</span>
</div>
<h3>Terminal</h3>
<p>Long-horizon command-line work across real tools, services, builds, and data: hundreds of commands toward one verified end state.</p>
<ul class="offer-list">
<li>Seeded from real recorded terminal sessions</li>
<li>Ops, data, build & release workflows</li>
<li>RL environments with executable verifiers</li>
</ul>
</article>
<article class="offer-card reveal">
<div class="offer-top offer-code" aria-hidden="true">
<span><i class="c-k">def</i> reconstruct(scenario):</span>
<span> duct = sweep(scenario[<i class="c-s">"inlet_size"</i>],</span>
<span> scenario[<i class="c-s">"outlet_size"</i>],</span>
<span> bend=scenario[<i class="c-s">"bend_angle"</i>])</span>
<span> <i class="c-k">return</i> add_ribs(duct, scenario[<i class="c-s">"rib_count"</i>])</span>
</div>
<h3>Coding</h3>
<p>Every terminal task is a coding task. Agents write, test, and debug real programs, from rebuilding scientific software to shipping full-stack apps.</p>
<ul class="offer-list">
<li>Software reconstruction from real program behavior</li>
<li>Repository repair & feature work</li>
<li>Full-stack apps verified in a browser</li>
</ul>
</article>
<article class="offer-card reveal">
<div class="offer-top offer-visual" aria-hidden="true">
<img src="assets/examples/stl-thumb.webp" alt="" width="740" height="394" loading="lazy">
</div>
<h3>Multimodal</h3>
<p>Visual inputs, code outputs: engineering drawings, renders, screenshots, meshes, and scientific images.</p>
<ul class="offer-list">
<li>CAD from engineering drawings</li>
<li>Web pages and apps from designs</li>
<li>Charts, SVG & documents to code</li>
</ul>
</article>
</div>
<div class="offer-band reveal">
<span>Every task ships with</span>
<ul>
<li>Reference solution</li>
<li>Executable verifier</li>
<li>Difficulty tier</li>
<li>Harness-consistent trajectories</li>
</ul>
</div>
</div>
</section>
<section class="section section-tint" id="examples" aria-labelledby="ex-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Examples</p>
<h2 id="ex-title">Watch real software run. Rebuild its behavior.</h2>
<p class="sub">A sample of our software-reconstruction tasks. The agent sees public runs of real scientific and engineering software, writes a program that reproduces it, and is graded on hidden configurations it never saw. Every visual is rendered from the task's own files.</p>
</div>
<div class="ex-grid" id="ex-grid">
<!--EXAMPLES_CARDS:START-->
<article class="ex-card reveal" data-index="0" data-kind="stl" tabindex="0" role="button" aria-label="Open example: OpenSCAD / BOSL2 parametric fan duct" style="--dc:#7c3aed">
<div class="ex-thumb"><img src="assets/examples/stl-thumb.webp" alt="Rendered fan-duct mesh from a public OpenSCAD run" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Spatial, Visual & 3D</p>
<h3>OpenSCAD / BOSL2 parametric fan duct</h3>
<p>Regenerate a swept fan-duct mount STL whose inlet size, outlet size, bend angle, and rib count change the geometry.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">3D mesh</li></ul>
<ul class="ex-sw"><li>OpenSCAD</li><li>BOSL2</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal" data-index="1" data-kind="foam" tabindex="0" role="button" aria-label="Open example: OpenFOAM 2-D dam-break CFD" style="--dc:#0891b2">
<div class="ex-thumb"><img src="assets/examples/foam-thumb.webp" alt="OpenFOAM dam-break mesh drawn from the public polyMesh" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Physical & Engineering Simulation</p>
<h3>OpenFOAM 2-D dam-break CFD</h3>
<p>Reconstruct a 2-D dam-break CFD workflow across water height, mesh, gravity, and simulation time.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">CFD mesh</li></ul>
<ul class="ex-sw"><li>OpenFOAM</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal" data-index="2" data-kind="drc" tabindex="0" role="button" aria-label="Open example: KiCad LED-matrix driver PCB" style="--dc:#7c3aed">
<div class="ex-thumb"><img src="assets/examples/drc-thumb.webp" alt="KiCad board with design-rule findings plotted at their positions" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Spatial, Visual & 3D</p>
<h3>KiCad LED-matrix driver PCB</h3>
<p>Reconstruct LED-matrix driver board outputs (DRC report and project files) under row, column, driver, and routing interventions.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">PCB layout</li></ul>
<ul class="ex-sw"><li>KiCad</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal" data-index="3" data-kind="container" tabindex="0" role="button" aria-label="Open example: Container SBOM and supply-chain attestation" style="--dc:#2563eb">
<div class="ex-thumb"><img src="assets/examples/container-thumb.webp" alt="Container image layers and attestation from a public run" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Software & Formal Systems</p>
<h3>Container SBOM and supply-chain attestation</h3>
<p>Infer coupled container base-image, package-set, label, entrypoint, and SBOM-scope behavior.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li></ul>
<ul class="ex-sw"><li>Hadolint</li><li>Syft</li><li>Skopeo</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal" data-index="4" data-kind="blender" tabindex="0" role="button" aria-label="Open example: Blender cloth flag scene" style="--dc:#7c3aed">
<div class="ex-thumb"><img src="assets/examples/blender-thumb.webp" alt="Blender cloth scene rotation dial across public scenarios" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Spatial, Visual & 3D</p>
<h3>Blender cloth flag scene</h3>
<p>Reconstruct cloth-scene behavior under wind, pin-group, frame, and subdivision interventions.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">3D scene</li></ul>
<ul class="ex-sw"><li>Blender</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal" data-index="5" data-kind="robot" tabindex="0" role="button" aria-label="Open example: MuJoCo / URDF mass-inertia repair" style="--dc:#0891b2">
<div class="ex-thumb"><img src="assets/examples/robot-thumb.webp" alt="MuJoCo joint-position trajectories across public scenarios" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Physical & Engineering Simulation</p>
<h3>MuJoCo / URDF mass-inertia repair</h3>
<p>Infer robot state and trajectory behavior under actuator, collision, pose, joint-limit, and target interventions.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li></ul>
<ul class="ex-sw"><li>MuJoCo</li><li>URDF/MJCF tooling</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal is-extra" data-index="6" data-kind="micro" tabindex="0" role="button" aria-label="Open example: Microscopy cell segmentation and counting" style="--dc:#16a34a">
<div class="ex-thumb"><img src="assets/examples/micro-thumb.webp" alt="Microscopy label masks from public runs" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Life & Molecular Sciences</p>
<h3>Microscopy cell segmentation and counting</h3>
<p>Reconstruct cell segmentation and measurement behavior across channel, illumination, threshold, object size, and measurement feature.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">Microscopy images</li></ul>
<ul class="ex-sw"><li>CellProfiler</li><li>scikit-image</li><li>Bio-Formats</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal is-extra" data-index="7" data-kind="astro" tabindex="0" role="button" aria-label="Open example: Astronomy FITS/WCS cutout and photometry" style="--dc:#db2777">
<div class="ex-thumb"><img src="assets/examples/astro-thumb.webp" alt="Astronomy FITS cutouts from public runs" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Earth & Space Sciences</p>
<h3>Astronomy FITS/WCS cutout and photometry</h3>
<p>Reconstruct FITS/WCS cutout and photometry behavior under aperture, background, clipping, threshold, and coordinate-frame interventions.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">FITS images</li></ul>
<ul class="ex-sw"><li>Astropy</li><li>Photutils</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal is-extra" data-index="8" data-kind="ocr" tabindex="0" role="button" aria-label="Open example: Tesseract table-image OCR" style="--dc:#d97706">
<div class="ex-thumb"><img src="assets/examples/ocr-thumb.webp" alt="OCR page preview from a public run" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Data & Document Workflows</p>
<h3>Tesseract table-image OCR</h3>
<p>Reconstruct how DPI, font, language, metadata, and page range affect OCR behavior.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">Document images</li></ul>
<ul class="ex-sw"><li>Pandoc</li><li>Poppler</li><li>Tesseract</li><li>QPDF</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="ex-card reveal is-extra" data-index="9" data-kind="color" tabindex="0" role="button" aria-label="Open example: Color-space and tone-curve pipeline" style="--dc:#7c3aed">
<div class="ex-thumb"><img src="assets/examples/color-thumb.webp" alt="Color-pipeline previews from public runs" width="740" height="394" loading="lazy"></div>
<div class="ex-card-body">
<p class="ex-card-domain">Spatial, Visual & 3D</p>
<h3>Color-space and tone-curve pipeline</h3>
<p>Infer a color pipeline across color space, white point, tone curve, demosaicing, and compression quality.</p>
<ul class="ex-card-tracks"><li class="trk trk-terminal">Terminal</li><li class="trk trk-coding">Coding</li><li class="trk trk-multimodal">Multimodal</li><li class="trk trk-mod">Images</li></ul>
<ul class="ex-sw"><li>colour-science</li><li>ImageMagick</li><li>scikit-image</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<!--EXAMPLES_CARDS:END-->
</div>
<div class="ex-more">
<button class="btn btn-ghost" id="ex-more" type="button" aria-expanded="false">Show more examples</button>
</div>
</div>
</section>
<section class="section" id="terminal-examples" data-nav="examples" aria-labelledby="tx-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Terminal & coding examples</p>
<h2 id="tx-title">Real terminal work, verified end to end.</h2>
<p class="sub">Seeded from real recorded terminal sessions, then extended round by round. Every task ships with its environment, a reference solution, and a verifier that passed in a fresh sandbox. Each replay shows commands from the task's own reference solution.</p>
</div>
<div class="tx-grid" id="tx-grid">
<!--TERMINAL_CARDS:START-->
<article class="tx-card reveal" data-index="0" tabindex="0" role="button" aria-label="Open terminal task: Repair and cross-build an AArch64 assembly program">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>ubuntu:22.04</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> source "$CFG"</span><span class="tx-cmd"><b>$</b> "$AS" --version</span><span class="tx-cmd"><b>$</b> cat > /app/horas_seg.s << 'SOURCEEOF'</span><span class="tx-cmd"><b>$</b> TEST_OUT=$(run_program "2" 2>&1)</span><span class="tx-cmd"><b>$</b> chmod +x /app/build.sh</span><span class="tx-cmd"><b>$</b> ln -sf /app/horas_seg /usr/local/bin/horas_seg</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Software engineering</p>
<h3>Repair and cross-build an AArch64 assembly program</h3>
<p>Fix the assembly source at /app/horas_seg.s so it prompts for integer hours and correctly prints the equivalent seconds (e.g., 2 → 7200 segundos.).</p>
<ul class="tx-meta"><li>239-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal" data-index="1" tabindex="0" role="button" aria-label="Open terminal task: Wire a network namespace with a veth pair">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>ubuntu:22.04</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> ip netns add t1</span><span class="tx-cmd"><b>$</b> ip link add veth0 mtu "$CHOSEN" type veth peer name veth1 mtu "$CHOSEN"</span><span class="tx-cmd"><b>$</b> ip link set veth0 netns t1</span><span class="tx-cmd"><b>$</b> ip netns exec t1 ip addr add 10.0.0.1/24 dev veth0</span><span class="tx-cmd"><b>$</b> ip addr add 10.0.0.2/24 dev veth1</span><span class="tx-cmd"><b>$</b> ip netns exec t1 ip link set veth0 up</span><span class="tx-cmd"><b>$</b> ip link set veth1 up</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Networking</p>
<h3>Wire a network namespace with a veth pair</h3>
<p>Set up a Linux network namespace <code>t1</code> with a veth pair.</p>
<ul class="tx-meta"><li>213-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal" data-index="2" tabindex="0" role="button" aria-label="Open terminal task: Align a container build with its Kustomize manifest">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>ubuntu:22.04</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> cp -f "$SOURCES_DIR/config.yaml" "$APP_DIR/config.yaml"</span><span class="tx-cmd"><b>$</b> sed -i "/^CMD /i COPY config.yaml /app/config.yaml" "$DOCKERFILE"</span><span class="tx-cmd"><b>$</b> sha256sum "$APP_DIR/$f" >> "$INTEGRITY"</span><span class="tx-cmd"><b>$</b> kustomize build "$APP_DIR" > "$RESULT"</span><span class="tx-cmd"><b>$</b> sed -i 's/^apiVersion: v1$/apiVersion=v1/' "$RESULT"</span><span class="tx-cmd"><b>$</b> tar czf evidence.tar.gz *</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Containers & cloud</p>
<h3>Align a container build with its Kustomize manifest</h3>
<p>Align the Dockerfile at /environment/Dockerfile with the files referenced by the kustomization in /app (kustomization.yaml), notably config.yaml.</p>
<ul class="tx-meta"><li>311-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal" data-index="3" tabindex="0" role="button" aria-label="Open terminal task: Stand up InfluxDB 2 against a compliance policy">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>debian:bookworm</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> jq empty "$f"</span><span class="tx-cmd"><b>$</b> influx setup --username "$USERNAME" --password "$PASSWORD" --org "$ORG" --bucket "$BUCKET" --retention…</span><span class="tx-cmd"><b>$</b> influx bucket update --id "$BID" --retention "$RETENTION"</span><span class="tx-cmd"><b>$</b> influx bucket update --id "$BID" --description "$REQUIRED_BUCKET_DESC"</span><span class="tx-cmd"><b>$</b> chmod +x /app/validate.sh</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Databases</p>
<h3>Stand up InfluxDB 2 against a compliance policy</h3>
<p>Set up InfluxDB 2 using the configuration at /etc/influxdb/init.json.</p>
<ul class="tx-meta"><li>329-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal" data-index="4" tabindex="0" role="button" aria-label="Open terminal task: Version a dataset with DVC and reconcile its history">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>ubuntu:22.04</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> dvc add dataset/foo</span><span class="tx-cmd"><b>$</b> dvc add dataset/test/0</span><span class="tx-cmd"><b>$</b> git add dataset/foo.dvc dataset/test/0.dvc</span><span class="tx-cmd"><b>$</b> git commit -m "DVC track dataset items"</span><span class="tx-cmd"><b>$</b> git diff "${PARENT}..${ADD_COMMIT}" -- "$dvc_file" > "$patch_path"</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Version control</p>
<h3>Version a dataset with DVC and reconcile its history</h3>
<p>Set up a DVC-tracked dataset under /app/dataset/ with two items: dataset/foo and dataset/test/0.</p>
<ul class="tx-meta"><li>313-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal" data-index="5" tabindex="0" role="button" aria-label="Open terminal task: Install an HPC scheduler's Python bindings, verified">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>fluxrm/flux-core:focal-v0.48.0</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> ACTUAL_HASH=$(sha256sum "$TARBALL" | awk '{print $1}')</span><span class="tx-cmd"><b>$</b> pip install --no-compile "$TARBALL"</span><span class="tx-cmd"><b>$</b> find "$FLUX_ROOT" -type f ! -name '*.pyc' -exec sha256sum {} \; | sort -k2 > "$INSTALLED_MANIFEST"</span><span class="tx-cmd"><b>$</b> cp "$INSTALLED_MANIFEST" "$EXPECTED_MANIFEST"</span><span class="tx-cmd"><b>$</b> flux start --test-size=4 bash -c '…'</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Data & scientific computing</p>
<h3>Install an HPC scheduler's Python bindings, verified</h3>
<p>Install the Flux Python package from the tarball at /app/flux-python-0.48.0rc6.tar.gz after verifying its integrity against the provided checksum file.</p>
<ul class="tx-meta"><li>171-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal is-extra" data-index="6" tabindex="0" role="button" aria-label="Open terminal task: Rebuild a GRUB boot sequence from a broken system">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>ubuntu:22.04</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> "$CHECK_SCRIPT"</span><span class="tx-cmd"><b>$</b> ERROR_LINE=$(grep -i "ERROR" "$LOG" | grep -i "boot-params" | head -1 || true)</span><span class="tx-cmd"><b>$</b> printf '%s\n' "${param_lines[@]}" > "$BOOT_PARAMS_FILE"</span><span class="tx-cmd"><b>$</b> linux_cmd="linux ${grub_prefix}/${actual_kernel_path} console=${console},${baud}"</span><span class="tx-cmd"><b>$</b> initrd_cmd="initrd ${grub_prefix}/${initrd_path}"</span><span class="tx-cmd"><b>$</b> cat > "$RESULT" <<EOF</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">System administration</p>
<h3>Rebuild a GRUB boot sequence from a broken system</h3>
<p>Assemble a three-line GRUB boot sequence (linux, initrd, boot) in /app/result.txt for the system rooted at /mnt/target.</p>
<ul class="tx-meta"><li>214-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal is-extra" data-index="7" tabindex="0" role="button" aria-label="Open terminal task: Diagnose and fix a Python dependency bug">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>python:3.11-slim</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> python /app/bug.py > /app/result.txt</span><span class="tx-cmd"><b>$</b> /app/validate_output.py /app/result.txt > /app/validation_stdout.txt 2> /app/validation_stderr.txt</span><span class="tx-cmd"><b>$</b> pip install "environs==${REQUIRED_VERSION}" 'marshmallow<4.0.0'</span><span class="tx-cmd"><b>$</b> python /app/bug.py > /app/fixed.txt</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Debugging & performance</p>
<h3>Diagnose and fix a Python dependency bug</h3>
<p>In /app, diagnose and fix the environs URL parsing bug.</p>
<ul class="tx-meta"><li>202-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<article class="tx-card reveal is-extra" data-index="8" tabindex="0" role="button" aria-label="Open terminal task: Count attack telemetry with the EQL shell">
<div class="tx-term" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em>python:3.9-slim</em></div>
<pre class="tx-screen"><code><span class="tx-cmd"><b>$</b> eql shell << 'EQL_IN'</span><span class="tx-cmd"><b>eql></b> input normalized-atomic-red-team.json.gz</span><span class="tx-cmd"><b>eql></b> search | count event_type</span><span class="tx-cmd"><b>eql></b> output counts.json</span><span class="tx-cmd"><b>$</b> /app/check.sh > /app/initial_failure.txt 2>&1</span><span class="tx-cmd"><b>$</b> python3 << 'PYREPAIR'</span></code></pre>
</div>
<div class="tx-body">
<p class="tx-area">Security analytics</p>
<h3>Count attack telemetry with the EQL shell</h3>
<p>Working in /app, count each event_type in the dataset '/app/normalized-atomic-red-team.json.gz' using the EQL shell and save results to counts.json.</p>
<ul class="tx-meta"><li>242-line reference solution</li><li>Verified in a fresh sandbox</li></ul>
</div>
<span class="ex-open" aria-hidden="true">Explore ↗</span>
</article>
<!--TERMINAL_CARDS:END-->
</div>
<div class="ex-more">
<button class="btn btn-ghost" id="tx-more" type="button" aria-expanded="false">Show more terminal tasks</button>
</div>
<div class="tx-coverage reveal">
<div class="tx-coverage-copy">
<h3>Breadth across real engineering work</h3>
<p>Share of terminal tasks by area, from system administration and databases to security and scientific computing.</p>
</div>
<ul class="tx-bars">
<!--TERMINAL_COVERAGE:START-->
<li><span class="txb-label">Software engineering</span><span class="txb-track"><i style="--w:100.0%;--i:0"></i></span><span class="txb-pct">19%</span></li>
<li><span class="txb-label">Scripting & automation</span><span class="txb-track"><i style="--w:91.5%;--i:1"></i></span><span class="txb-pct">17%</span></li>
<li><span class="txb-label">System administration</span><span class="txb-track"><i style="--w:72.9%;--i:2"></i></span><span class="txb-pct">14%</span></li>
<li><span class="txb-label">Environment setup</span><span class="txb-track"><i style="--w:57.4%;--i:3"></i></span><span class="txb-pct">11%</span></li>
<li><span class="txb-label">Version control</span><span class="txb-track"><i style="--w:44.7%;--i:4"></i></span><span class="txb-pct">8%</span></li>
<li><span class="txb-label">Containers & cloud</span><span class="txb-track"><i style="--w:36.7%;--i:5"></i></span><span class="txb-pct">7%</span></li>
<li><span class="txb-label">Security</span><span class="txb-track"><i style="--w:29.3%;--i:6"></i></span><span class="txb-pct">6%</span></li>
<li><span class="txb-label">Databases</span><span class="txb-track"><i style="--w:19.7%;--i:7"></i></span><span class="txb-pct">4%</span></li>
<li><span class="txb-label">Data & scientific computing</span><span class="txb-track"><i style="--w:17.0%;--i:8"></i></span><span class="txb-pct">3%</span></li>
<li><span class="txb-label">Files & storage</span><span class="txb-track"><i style="--w:14.9%;--i:9"></i></span><span class="txb-pct">3%</span></li>
<li><span class="txb-label">Networking</span><span class="txb-track"><i style="--w:11.7%;--i:10"></i></span><span class="txb-pct">2%</span></li>
<li><span class="txb-label">Other</span><span class="txb-track"><i style="--w:36.7%;--i:11"></i></span><span class="txb-pct">7%</span></li>
<!--TERMINAL_COVERAGE:END-->
</ul>
</div>
</div>
</section>
<section class="section section-tint" id="why" aria-labelledby="why-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Why verified synthesis</p>
<h2 id="why-title">Scale without losing the proof.</h2>
<p class="sub">Hand-written environments are slow to produce. Prompted generation is fast, but a task can look perfect and be silently broken. We generate aggressively and then try to break every task, so only environments that survive the gates ship.</p>
</div>
<div class="grid-4">
<article class="value-card reveal">
<div class="icon-tile"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M12 3 4.5 6v5.5c0 4.6 3.1 8.2 7.5 9.5 4.4-1.3 7.5-4.9 7.5-9.5V6L12 3Z" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linejoin="round"/><path d="m8.8 12.2 2.3 2.3 4.3-4.6" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round"/></svg></div>
<h3>Verified by execution</h3>
<p>A task is accepted only if its reference solution passes in a fresh sandbox, a do-nothing agent fails, and every check is stated in the instruction.</p>
</article>
<article class="value-card reveal">
<div class="icon-tile"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M4 19h4v-4h4v-4h4V7h4" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round"/><circle cx="20" cy="7" r="1.6" fill="currentColor"/></svg></div>
<h3>Difficulty you can dial</h3>
<p>Tasks grow harder round by round. Every task carries a measured tier, so you train on a curriculum instead of a pile.</p>
</article>
<article class="value-card reveal">
<div class="icon-tile"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M20 12a8 8 0 1 1-2.3-5.6M20 4v4.5h-4.5" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round"/></svg></div>
<h3>Fresh every month</h3>
<p>New, never-published environments each cycle, synthesized for you and audited against public benchmark test sets before delivery.</p>
</article>
<article class="value-card reveal">
<div class="icon-tile"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M4 6h5c2 0 3 1 4.5 3S16 12 19 12M4 12h15M4 18h5c2 0 3-1 4.5-3S16 12 19 12" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round"/><circle cx="19.5" cy="12" r="1.6" fill="currentColor"/></svg></div>
<h3>One consistent harness</h3>
<p>Passing runs from many harnesses are rewritten into the single harness your model trains and runs under, so it learns the skill, not the scaffolding.</p>
</article>
</div>
</div>
</section>
<section class="section" id="capabilities" aria-labelledby="cap-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Capabilities</p>
<h2 id="cap-title">Six task families. One standard of proof.</h2>
<p class="sub">Each family pairs what the model sees with what it must produce and how we check it. Pick one, or combine them into a single training mix.</p>
</div>
<div class="tabs reveal">
<div class="tab-list" role="tablist" aria-label="Task families">
<button class="tab" role="tab" id="tab-terminal" aria-controls="panel-terminal" aria-selected="true">
<span class="tab-icon"><svg viewBox="0 0 24 24" aria-hidden="true"><rect x="3" y="4.5" width="18" height="15" rx="2.5" fill="none" stroke="currentColor" stroke-width="1.7"/><path d="m7 10 3 2.5L7 15m5 0h5" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round"/></svg></span>
<span>Terminal & coding agents</span>
</button>
<button class="tab" role="tab" id="tab-science" aria-controls="panel-science" aria-selected="false" tabindex="-1">
<span class="tab-icon"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M9.5 3.5h5M10.5 3.5v6L5 19a1.5 1.5 0 0 0 1.3 2.2h11.4A1.5 1.5 0 0 0 19 19l-5.5-9.5v-6" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round"/><path d="M7.5 15h9" stroke="currentColor" stroke-width="1.7" stroke-linecap="round"/></svg></span>
<span>Scientific software reconstruction</span>
</button>
<button class="tab" role="tab" id="tab-cad" aria-controls="panel-cad" aria-selected="false" tabindex="-1">
<span class="tab-icon"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="m12 3 8 4.5v9L12 21l-8-4.5v-9L12 3Z" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linejoin="round"/><path d="m4 7.5 8 4.5 8-4.5M12 12v9" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linejoin="round"/></svg></span>
<span>CAD from engineering drawings</span>
</button>
<button class="tab" role="tab" id="tab-web" aria-controls="panel-web" aria-selected="false" tabindex="-1">
<span class="tab-icon"><svg viewBox="0 0 24 24" aria-hidden="true"><rect x="3" y="4" width="18" height="16" rx="2.5" fill="none" stroke="currentColor" stroke-width="1.7"/><path d="M3 8.5h18M7 13h4m-4 3h7" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round"/></svg></span>
<span>Web & full-stack applications</span>
</button>
<button class="tab" role="tab" id="tab-visual" aria-controls="panel-visual" aria-selected="false" tabindex="-1">
<span class="tab-icon"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M4 20V10m5 10V5m5 15v-7m5 7V8" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round"/></svg></span>
<span>Visual-to-code</span>
</button>
<button class="tab" role="tab" id="tab-swe" aria-controls="panel-swe" aria-selected="false" tabindex="-1">
<span class="tab-icon"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="m8.5 8-4 4 4 4m7-8 4 4-4 4M13.5 5l-3 14" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round"/></svg></span>
<span>Multimodal software engineering</span>
</button>
</div>
<div class="tab-panels">
<div class="tab-panel is-active" role="tabpanel" id="panel-terminal" aria-labelledby="tab-terminal" tabindex="0">
<div class="panel-copy">
<h3>Terminal & coding agents</h3>
<p>Multi-hour workflows where the agent explores a real environment, writes and runs programs, and checks its own work. Terminal tasks are coding tasks.</p>
<dl class="spec">
<div><dt>Model sees</dt><dd>A live workspace, an instruction, and the tools a practitioner would reach for.</dd></div>
<div><dt>Model produces</dt><dd>Working code and a verified end state after hundreds of commands: rebuilt programs, migrated services, repaired builds.</dd></div>
<div><dt>We verify</dt><dd>Artifacts and system state, with checks that reject placeholders, hard-coded outputs, and skipped steps.</dd></div>
</dl>
</div>
<div class="panel-visual">
<div class="mock mock-terminal">
<div class="mock-bar dark"><i></i><i></i><i></i><em>agent@sandbox:/app</em></div>
<pre class="term"><code><span class="t-line" style="--i:0"><b>$</b> jq -c .axes_to_vary observations/manifest.json</span>
<span class="t-line t-dim" style="--i:1">["inlet_size","outlet_size","bend_angle","rib_count"]</span>
<span class="t-line" style="--i:2"><b>$</b> ls /tmp/obs/public_1/native</span>
<span class="t-line t-dim" style="--i:3">metrics.json product.stl</span>
<span class="t-line" style="--i:4"><b>$</b> jq -c .parameters /tmp/obs/public_1/scenario.json</span>
<span class="t-line t-dim" style="--i:5">{"bend_angle":238,"inlet_size":11.15,"rib_count":5}</span>
<span class="t-line" style="--i:6"><b>$</b> python3 workspace/reconstruct.py < hidden.json</span>
<span class="t-line t-verify" style="--i:7">verifier ▸ hidden configurations vs. real OpenSCAD</span>
<span class="t-line t-ok" style="--i:8">✓ geometry matches on every hidden scenario</span></code></pre>
</div>
</div>
</div>
<div class="tab-panel" role="tabpanel" id="panel-science" aria-labelledby="tab-science" tabindex="0" hidden>
<div class="panel-copy">
<h3>Scientific software reconstruction</h3>
<p>Real scientific and engineering software becomes the reference solution. The model sees how it behaves and must rebuild that behavior without its source.</p>
<dl class="spec">
<div><dt>Model sees</dt><dd>Public runs of real software: parameters plus native outputs such as tables, meshes, waveforms, images, and sequences.</dd></div>
<div><dt>Model produces</dt><dd>An editable program that reproduces the software's behavior across its inputs.</dd></div>
<div><dt>We verify</dt><dd>On hidden configurations the model never saw, with domain-aware comparison, so replaying public outputs never passes.</dd></div>
</dl>
<ul class="chip-row" aria-label="Domains">
<li>Physical & engineering simulation</li><li>Life & molecular science</li><li>Earth & space</li><li>Spatial & 3D</li><li>Data & documents</li><li>Software & formal systems</li>
</ul>
</div>
<div class="panel-visual">
<div class="mock mock-plot">
<div class="mock-bar"><i></i><i></i><i></i><em>public runs vs hidden configuration</em></div>
<svg viewBox="0 0 360 210" role="img" aria-label="Two public runs and a hidden configuration matched by the reconstructed program">
<line x1="36" y1="186" x2="340" y2="186" class="axis"/>
<line x1="36" y1="22" x2="36" y2="186" class="axis"/>
<path class="wave wave-a" data-wave="58 0.9 6 0"/>
<path class="wave wave-b" data-wave="44 0.7 8 0.4"/>
<path class="wave wave-hidden" data-wave="66 0.5 10 0.9"/>
<path class="wave wave-recon" data-wave="66 0.5 10 0.9"/>
<g class="legend">
<rect x="210" y="16" width="10" height="3" rx="1.5" class="lg-a"/><text x="226" y="20">public · gain 0.4</text>
<rect x="210" y="30" width="10" height="3" rx="1.5" class="lg-b"/><text x="226" y="34">public · gain 0.8</text>
<rect x="210" y="44" width="10" height="3" rx="1.5" class="lg-h"/><text x="226" y="48">hidden · gain 1.3</text>
</g>
</svg>
<span class="mock-chip ok">✓ hidden configuration reproduced</span>
</div>
</div>
</div>
<div class="tab-panel" role="tabpanel" id="panel-cad" aria-labelledby="tab-cad" tabindex="0" hidden>
<div class="panel-copy">
<h3>CAD from engineering drawings</h3>
<p>Turn a dimensioned drawing or multi-view renders into parametric CAD code that builds the right solid.</p>
<dl class="spec">
<div><dt>Model sees</dt><dd>An engineering drawing or multi-view renders, with as much or as little supplied dimensioning as the tier calls for.</dd></div>
<div><dt>Model produces</dt><dd>A CadQuery, OpenSCAD, or FreeCAD program and the watertight solid it builds.</dd></div>
<div><dt>We verify</dt><dd>Geometry, not code: a fresh rebuild is compared on bounds, volume, surface area, and probes inside every hole and profile.</dd></div>
</dl>
</div>
<div class="panel-visual">
<div class="mock mock-cad">
<svg viewBox="0 0 360 230" role="img" aria-label="Engineering drawing of an L-bracket with dimensions and a section view">
<defs>
<pattern id="hatch" width="7" height="7" patternUnits="userSpaceOnUse" patternTransform="rotate(45)"><line x1="0" y1="0" x2="0" y2="7" class="hatch-line"/></pattern>
<marker id="arr" markerWidth="8" markerHeight="8" refX="4" refY="4" orient="auto-start-reverse" markerUnits="userSpaceOnUse"><path d="M0 0 8 4 0 8z" class="arrow"/></marker>
</defs>
<rect x="0.5" y="0.5" width="359" height="229" rx="10" class="paper"/>
<path class="draw part" pathLength="1" d="M40 52H170V84H96V182H40Z"/>
<circle class="draw part" pathLength="1" cx="68" cy="150" r="11"/>
<circle class="draw part" pathLength="1" cx="140" cy="68" r="7"/>
<path class="center" d="M52 150h32M68 134v32M128 68h24M140 56v24"/>
<path class="dim" d="M40 48V30M170 48V30M36 52H18M36 182H18"/>
<line class="dim" x1="40" y1="36" x2="170" y2="36" marker-start="url(#arr)" marker-end="url(#arr)"/>
<line class="dim" x1="24" y1="52" x2="24" y2="182" marker-start="url(#arr)" marker-end="url(#arr)"/>
<text x="105" y="30" class="dim-text" text-anchor="middle">130</text>
<text x="16" y="120" class="dim-text" text-anchor="middle" transform="rotate(-90 16 120)">130</text>
<path class="dim" d="M76 142 104 124H130"/>
<text x="106" y="119" class="dim-text">Ø22 THRU</text>
<rect x="214" y="52" width="38" height="130" class="cad-section"/>
<path class="hidden-line" d="M214 139h38M214 161h38"/>
<text x="233" y="200" class="label" text-anchor="middle">SECTION A–A</text>
<g class="title-block">
<rect x="270" y="150" width="78" height="66" rx="3"/>
<path d="M270 172h78M270 194h78M309 194v22"/>
<text x="309" y="165" text-anchor="middle">BRACKET</text>
<text x="309" y="187" text-anchor="middle">mm · 1:2</text>
<text x="289" y="209" text-anchor="middle">REV B</text>
<text x="328" y="209" text-anchor="middle">A3</text>
</g>
</svg>
<span class="mock-chip ok">✓ rebuilt solid matches geometry</span>
</div>
</div>
</div>
<div class="tab-panel" role="tabpanel" id="panel-web" aria-labelledby="tab-web" tabindex="0" hidden>
<div class="panel-copy">
<h3>Web & full-stack applications</h3>
<p>From a design, a specification, or a legacy desktop app to a running web product with real persistence.</p>
<dl class="spec">
<div><dt>Model sees</dt><dd>Requirements, design references, legacy applications, or an existing codebase.</dd></div>
<div><dt>Model produces</dt><dd>Running frontends and full-stack services: routes, APIs, state, and storage.</dd></div>
<div><dt>We verify</dt><dd>Headless-browser tests across routes and viewports, API contracts, restart persistence, and rendered-page comparison against the design.</dd></div>
</dl>
</div>
<div class="panel-visual">
<div class="mock mock-browser">
<div class="mock-bar"><i></i><i></i><i></i><span class="url">localhost:8080/dispatch</span></div>
<div class="browser-body">
<div class="sk-head"><span class="sk-logo"></span><span class="sk-pill"></span><span class="sk-pill"></span><span class="sk-pill short"></span></div>
<div class="sk-grid">
<div class="sk-side"><span></span><span></span><span></span><span class="short"></span></div>
<div class="sk-main">
<div class="sk-row head"></div><div class="sk-row"></div><div class="sk-row"></div><div class="sk-row"></div><div class="sk-row short"></div>
</div>
</div>
<ul class="checklist">
<li>All routes render at mobile and desktop widths</li>
<li>API contract and validation rules hold</li>
<li>State survives a full restart</li>
<li>Rendered page matches the design</li>
</ul>
</div>
</div>
</div>
</div>
<div class="tab-panel" role="tabpanel" id="panel-visual" aria-labelledby="tab-visual" tabindex="0" hidden>
<div class="panel-copy">
<h3>Visual-to-code</h3>
<p>Charts, diagrams, vector graphics, and documents rebuilt as code that re-renders them, and keeps working when the data changes.</p>
<dl class="spec">
<div><dt>Model sees</dt><dd>A chart, diagram, SVG, or document page, optionally with the underlying data.</dd></div>
<div><dt>Model produces</dt><dd>Plotting code, SVG, or LaTeX that regenerates the image.</dd></div>
<div><dt>We verify</dt><dd>Re-render and compare structure and pixels, then re-run on hidden data variations so that copying one image never passes.</dd></div>
</dl>
</div>
<div class="panel-visual">
<div class="mock mock-split">
<div class="split-chart">
<svg viewBox="0 0 200 150" role="img" aria-label="Bar chart with a trend line">
<line x1="18" y1="132" x2="192" y2="132" class="axis"/>
<rect x="28" y="88" width="18" height="44" rx="3" class="bar-a"/>
<rect x="56" y="70" width="18" height="62" rx="3" class="bar-a"/>
<rect x="84" y="78" width="18" height="54" rx="3" class="bar-a"/>
<rect x="112" y="52" width="18" height="80" rx="3" class="bar-b"/>
<rect x="140" y="40" width="18" height="92" rx="3" class="bar-b"/>
<rect x="168" y="26" width="18" height="106" rx="3" class="bar-b"/>
<path d="M37 80 65 64 93 70 121 44 149 32 177 18" class="trend"/>
</svg>
</div>
<pre class="split-code"><code><span class="c-k">import</span> matplotlib.pyplot <span class="c-k">as</span> plt
fig, ax = plt.subplots()
ax.bar(q, revenue, color=c)
ax.plot(q, trend, marker=<span class="c-s">"o"</span>)
ax.set_ylim(<span class="c-n">0</span>, <span class="c-n">120</span>)
fig.savefig(out, dpi=<span class="c-n">200</span>)</code></pre>
<span class="mock-chip ok">✓ re-render match · ✓ hidden data variant</span>
</div>
</div>
</div>
<div class="tab-panel" role="tabpanel" id="panel-swe" aria-labelledby="tab-swe" tabindex="0" hidden>
<div class="panel-copy">
<h3>Multimodal software engineering</h3>
<p>Fix what users can see. Visual bug reports inside real repositories, resolved with patches that hold up in a browser.</p>
<dl class="spec">
<div><dt>Model sees</dt><dd>Bug reports with screenshots and screen recordings, plus the full repository.</dd></div>
<div><dt>Model produces</dt><dd>A patch that fixes the visible behavior without breaking anything else.</dd></div>
<div><dt>We verify</dt><dd>UI-level tests in a real browser, render diffs, and the project's regression suite.</dd></div>
</dl>
</div>
<div class="panel-visual">
<div class="mock mock-issue">
<div class="issue-head"><span class="issue-dot"></span><strong>Menu clips under the hero on narrow screens</strong></div>
<div class="issue-shot" aria-hidden="true"><span class="shot-nav"></span><span class="shot-menu"></span><span class="shot-hero"></span></div>
<pre class="diff"><code><span class="d-h">@@ components/NavMenu.vue</span>
<span class="d-del">- .menu { overflow: hidden; }</span>
<span class="d-add">+ .menu { overflow: visible; }</span>
<span class="d-add">+ .menu { z-index: 40; }</span></code></pre>
<span class="mock-chip ok">✓ UI tests pass · ✓ render diff clean</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section section-dark" id="pipeline" aria-labelledby="pipe-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">How it works</p>
<h2 id="pipe-title">Generate aggressively. Then try to break every task.</h2>
<p class="sub">Generation is the cheap part. The value is in the gates: a task reaches you only after it survives all of them.</p>
</div>
<ol class="pipeline reveal">
<li style="--d:0">
<span class="step-dot">1</span>
<h3>Seed</h3>
<p>Licensed or authored seed tasks and real software workflows. Never benchmark test sets.</p>
</li>
<li style="--d:1">
<span class="step-dot">2</span>
<h3>Extend</h3>
<p>Rewrite operators add tools, services, data, and failure modes. Solution, environment, verifier, and instruction change together.</p>
</li>
<li style="--d:2">
<span class="step-dot">3</span>
<h3>Verify</h3>
<p>Fresh-sandbox build. The reference must pass, a do-nothing agent must fail, and every verifier check must be stated in the instruction.</p>
</li>
<li style="--d:3">
<span class="step-dot">4</span>
<h3>Calibrate</h3>
<p>Measure pass rates with reference solvers or your checkpoint, and assign every task a difficulty tier.</p>
</li>
<li style="--d:4">
<span class="step-dot">5</span>
<h3>Deliver</h3>
<p>Environments, verified trajectories, and a calibration and integrity report for every batch.</p>
</li>
</ol>
<ul class="gate-row reveal" aria-label="Acceptance gates">
<li>Reference solution passes</li>
<li>No-op agent fails</li>
<li>Replay shortcuts fail</li>
<li>Instruction ↔ verifier consistency</li>
<li>Leakage & overlap audit</li>
<li>Sampled human review</li>
</ul>
</div>
</section>
<section class="section" id="trajectories" aria-labelledby="traj-title">
<div class="container traj-grid">
<div class="traj-copy reveal">
<p class="kicker">Trajectory scaling</p>
<h2 id="traj-title">More passing trajectories. One consistent harness.</h2>
<p class="sub">Hard tasks get solved in different ways under different harnesses, and raw trajectories carry each harness's scaffolding with them. We pool successes across harnesses, then rewrite each one into the single harness your model trains and runs under, so it learns the skill, not the scaffolding.</p>
<ol class="traj-steps">
<li><strong>Discover.</strong> Several harnesses attack the same tasks and solve ones no single harness can.</li>
<li><strong>Plan.</strong> Each success becomes a runbook: end state, milestones, checks, and recovery steps, never the finished answer.</li>
<li><strong>Screen.</strong> A critic rejects runbooks that leak the verifier or the solution and asks for revisions.</li>
<li><strong>Re-execute.</strong> The task is solved again in a fresh sandbox under one consistent harness. Only passing, leak-free runs are kept.</li>
</ol>
</div>
<div class="traj-visual reveal">
<svg class="traj-diagram" viewBox="0 0 640 300" role="img" aria-label="Successes from three harnesses are distilled into runbooks, screened by a critic, and re-executed into many verified trajectories under one harness">
<defs>
<marker id="tarrow" markerWidth="8" markerHeight="8" refX="6" refY="4" orient="auto" markerUnits="userSpaceOnUse"><path d="M0 0 8 4 0 8z" class="t-arrow"/></marker>
</defs>
<text x="0" y="22" class="t-stage">Discover</text>
<text x="232" y="22" class="t-stage">Rewrite</text>
<text x="560" y="22" class="t-stage">Train</text>
<g class="t-src">
<circle cx="138" cy="70" r="5" class="src-a"/><text x="126" y="74" text-anchor="end">General loop</text>
<circle cx="138" cy="150" r="5" class="src-b"/><text x="126" y="154" text-anchor="end">Stateful workflow</text>
<circle cx="138" cy="230" r="5" class="src-c"/><text x="126" y="234" text-anchor="end">Self-reflect & retry</text>
</g>
<path class="t-lane lane-a" d="M144 70C186 70 198 150 230 150"/>
<path class="t-lane lane-b" d="M144 150H230"/>
<path class="t-lane lane-c" d="M144 230C186 230 198 150 230 150"/>
<g class="t-node"><rect x="232" y="126" width="88" height="48" rx="12"/><text x="276" y="147" text-anchor="middle">Runbook</text><text x="276" y="163" text-anchor="middle" class="t-sub">planner</text></g>
<g class="t-node"><rect x="340" y="126" width="88" height="48" rx="12"/><text x="384" y="147" text-anchor="middle">Critic</text><text x="384" y="163" text-anchor="middle" class="t-sub">leak screen</text></g>
<g class="t-node t-node-strong"><rect x="448" y="126" width="96" height="48" rx="12"/><text x="496" y="147" text-anchor="middle">Re-execute</text><text x="496" y="163" text-anchor="middle" class="t-sub">fresh sandbox</text></g>
<path class="t-link" d="M320 150H336" marker-end="url(#tarrow)"/>
<path class="t-link" d="M428 150H444" marker-end="url(#tarrow)"/>
<path class="t-revise" d="M384 124C384 92 276 92 276 122" marker-end="url(#tarrow)"/>
<text x="330" y="88" text-anchor="middle" class="t-sub">revise</text>
<path class="t-out" d="M544 150C576 150 584 86 616 86"/>
<path class="t-out" d="M544 150C576 150 584 118 616 118"/>
<path class="t-out" d="M544 150H616"/>
<path class="t-out" d="M544 150C576 150 584 182 616 182"/>
<path class="t-out" d="M544 150C576 150 584 214 616 214"/>
<g class="t-ok"><circle cx="622" cy="86" r="5"/><circle cx="622" cy="118" r="5"/><circle cx="622" cy="150" r="5"/><circle cx="622" cy="182" r="5"/><circle cx="622" cy="214" r="5"/></g>
<text x="582" y="252" text-anchor="middle" class="t-sub">verified runs,</text>
<text x="582" y="268" text-anchor="middle" class="t-sub">one harness</text>
</svg>
<dl class="traj-metrics">
<div><dt>+34%</dt><dd>more tasks solved by pooling harnesses than by the best single harness</dd></div>
<div><dt>5.5×</dt><dd>verified training trajectories per successful source run</dd></div>
<div><dt>57% → 74%</dt><dd>Terminal-Bench 2 pass@3 after training on rewritten trajectories</dd></div>
<div><dt>+20.8 pts</dt><dd>over fine-tuning directly on the raw successes (Terminal-Bench 2)</dd></div>
</dl>
</div>
</div>
</section>
<section class="section section-tint" id="difficulty" aria-labelledby="diff-title">
<div class="container diff-grid">
<div class="diff-copy reveal">
<p class="kicker">Calibrated difficulty</p>
<h2 id="diff-title">A curriculum, not a pile.</h2>
<p class="sub">Each synthesis round makes tasks measurably harder: longer solutions, more tools, more checks. In our runs, a fixed strong solver went from solving 90% of first-round tasks to 2.5% by round fifteen, with no ceiling in sight.</p>
<p class="sub">We measure your checkpoint on every tier and ship the band where it succeeds sometimes, but not yet reliably. As your model improves, the band moves up with it.</p>
<div class="slider-card">
<label for="ladder-strength">Where is your model today?</label>
<input id="ladder-strength" type="range" min="-3" max="4" step="0.1" value="0" aria-describedby="ladder-readout">
<div class="slider-ends"><span>Weaker</span><span id="ladder-strength-label">Reference solver</span><span>Stronger</span></div>
<p class="readout" id="ladder-readout" aria-live="polite"></p>
</div>
</div>
<figure class="chart-card reveal">
<svg id="ladder-chart" role="img" aria-label="Success rate by synthesis round, with the learning zone between 10 and 50 percent highlighted"></svg>
<figcaption>
<span class="legend-item"><i class="lg-zone"></i>Learning zone (10–50% success)</span>
<span class="legend-item"><i class="lg-bar"></i>Your model (modeled)</span>
<span class="legend-item"><i class="lg-line"></i>Reference solver</span>
<span class="legend-item"><i class="lg-dot"></i>Measured reference results (pass@4)</span>
<span class="note">Illustrative: the curve interpolates measured reference-solver results. Moving the slider shows how the recommended band shifts.</span>
</figcaption>
</figure>
</div>
<div class="container">
<div class="growth reveal" id="growth">
<div class="growth-head">
<div class="growth-copy">
<p class="kicker">Watch one task grow</p>
<h3>Same goal. More to get right at every stage.</h3>
<p>Successive stages from one task family: <strong>build a Git commit using only low-level plumbing commands</strong>. Each stage keeps the goal and adds requirements, and the reference solution and verifier grow with it. New requirements are highlighted.</p>
</div>
<div class="growth-steps" id="growth-steps" role="group" aria-label="Synthesis stages"></div>
</div>
<div class="growth-body">
<div class="growth-instruction" id="growth-instruction" aria-live="polite"></div>
<div class="growth-stats">
<div class="gs-stat">
<span>Reference solution</span>
<strong id="growth-lines">–</strong>
<div class="gs-bar"><i id="growth-lines-bar"></i></div>
</div>
<div class="gs-stat">
<span>Instruction</span>
<strong id="growth-words">–</strong>
<div class="gs-bar gs-bar-alt"><i id="growth-words-bar"></i></div>
</div>
<figure class="gs-spark">
<svg id="growth-spark" viewBox="0 0 240 72" role="img" aria-label="Reference solution length across stages"></svg>
<figcaption>Reference solution length by stage</figcaption>
</figure>
<p class="gs-note">Illustrative family: stages are related tasks from successive synthesis rounds, shown in order of depth.</p>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="deliverables" aria-labelledby="deliv-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Deliverables</p>
<h2 id="deliv-title">Everything you need to train on it tomorrow.</h2>
<p class="sub">Delivered in a reproducible, container-first format that drops into SFT and RL pipelines, with adapters for your stack on request.</p>
</div>
<div class="deliv-grid">
<article class="deliv-card reveal">
<h3>RL environments</h3>
<p>Self-contained Docker environments with instruction, executable verifier, reference solution, and metadata, in Harbor task format.</p>
</article>
<article class="deliv-card reveal">
<h3>Verified trajectories</h3>
<p>Agent runs kept only when they pass the hidden checks, rewritten into one consistent harness, and exported as chat-format JSONL for fine-tuning.</p>
</article>
<article class="deliv-card reveal">
<h3>Calibration report</h3>
<p>Tier and measured pass rate for every task, against reference solvers or your own checkpoint.</p>
</article>
<article class="deliv-card reveal">
<h3>Integrity report & data card</h3>
<p>Overlap audit against public benchmark test sets, shortcut controls, and sampled human review, plus the domains, tools, licenses, and known limitations of every batch.</p>
</article>
<article class="deliv-card tree-card reveal" aria-label="Example environment bundle">
<pre class="tree"><code><span class="tr-dir">task/</span>
├── instruction.md <span class="tr-c"># what the agent sees</span>
├── task.toml <span class="tr-c"># tier, tags, timeouts</span>
├── environment/
│ └── Dockerfile <span class="tr-c"># fresh, pinned sandbox</span>
├── solution/
│ └── solve.sh <span class="tr-c"># reference solution</span>
└── tests/
└── test_state.py <span class="tr-c"># executable verifier</span>
<span class="tr-dir">trajectories.jsonl</span> <span class="tr-c"># verified agent runs</span>
<span class="tr-dir">report.json</span> <span class="tr-c"># calibration + audit</span></code></pre>
</article>
</div>
</div>
</section>
<section class="section section-tint" id="benchmarks" aria-labelledby="bench-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Alignment</p>
<h2 id="bench-title">Built for the skills frontier evaluations measure.</h2>
<p class="sub">We train skills, not test items. No benchmark test data is used, copied, or derived, and every batch ships with an overlap audit.</p>
</div>
<div class="bench-grid">
<div class="bench-group reveal">
<h3>Terminal & long-horizon agents</h3>
<ul class="chip-row"><li>Terminal-Bench 2</li><li>Terminal-Bench 4</li><li>Terminal-Bench Hard</li><li>Long-Horizon Terminal-Bench</li></ul>
</div>
<div class="bench-group reveal">
<h3>Multimodal SWE & web</h3>
<ul class="chip-row"><li>SWE-bench Multimodal</li><li>Vision2Web</li><li>Interaction2Code</li><li>Design2Code</li></ul>
</div>
<div class="bench-group reveal">
<h3>CAD & visual-to-code</h3>
<ul class="chip-row"><li>BenchCAD</li><li>Chart2Code</li><li>Image2Struct</li><li>MMSVGBench</li></ul>
</div>
</div>
</div>
</section>
<section class="section" id="research" aria-labelledby="research-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Research</p>
<h2 id="research-title">Proven in our research.</h2>
<p class="sub">Each method is measured across multiple seeds and against matched-budget baselines.</p>
</div>
<div class="research-grid">
<article class="paper-card reveal">
<p class="paper-tag">Scientific software</p>
<h3>Software-in-the-loop reconstruction</h3>
<p>Existing scientific software becomes both the reference solution and the source of hidden test cases, so verification scales with the software rather than with authoring effort.</p>
<dl class="metrics">
<div><dt>+5.6 pts</dt><dd>Terminal-Bench 2 after fine-tuning (47.9% → 53.6%, three seeds)</dd></div>
<div><dt>#1 of 5</dt><dd>training corpora at a matched token budget, on all four evaluations</dd></div>
<div><dt>98%</dt><dd>verifier agreement with blinded human review</dd></div>
</dl>
</article>
<article class="paper-card reveal">
<p class="paper-tag">Task difficulty</p>
<h3>Recursive verified synthesis</h3>
<p>Accepted tasks seed the next round. Each round extends the solution, verifier, and instruction together and re-validates the whole task in a fresh sandbox, with no human authoring in the loop.</p>
<dl class="metrics">
<div><dt>+10 pts</dt><dd>best fine-tuning gain across Terminal-Bench 2, Terminal-Bench Hard, and LHTB</dd></div>
<div><dt>+41%</dt><dd>relative gain on Terminal-Bench Hard with agentic PPO</dd></div>
<div><dt>36×</dt><dd>drop in fixed-solver success from round 1 to 15 (90% → 2.5%), no ceiling observed</dd></div>
</dl>
</article>
<article class="paper-card reveal">
<p class="paper-tag">Trajectory scaling</p>
<h3>Recursive self-rewrite</h3>
<p>Successes discovered under diverse harnesses are rewritten by the same base model into one general harness, with leakage screening and fresh-sandbox re-execution.</p>
<dl class="metrics">
<div><dt>74.2%</dt><dd>Terminal-Bench 2 pass@3, up from 57.0%</dd></div>
<div><dt>6×</dt><dd>Terminal-Bench 4 pass@3, from 1.5% to 9.1%</dd></div>
<div><dt>+20.8 pts</dt><dd>over direct fine-tuning on the same successes</dd></div>
</dl>
</article>
</div>
</div>
</section>
<section class="section section-tint" id="engage" aria-labelledby="engage-title">
<div class="container">
<div class="section-head reveal">
<p class="kicker">Working together</p>
<h2 id="engage-title">Start small. Scale every month.</h2>
</div>
<div class="engage-grid">
<article class="engage-card reveal">
<span class="engage-num">01</span>
<h3>Pilot</h3>
<p>A calibrated sample batch in your format, measured on your checkpoint, for the task families that matter most to you.</p>
</article>
<article class="engage-card reveal">
<span class="engage-num">02</span>
<h3>Monthly refresh</h3>
<p>Fresh, never-published environments every month at the tiers you choose. Difficulty moves up as your model improves.</p>
</article>
<article class="engage-card reveal">
<span class="engage-num">03</span>
<h3>Custom & exclusive</h3>
<p>New domains, software stacks, and modalities built for your team alone, on your timeline.</p>
</article>
</div>
</div>
</section>
<section class="cta" id="contact" aria-labelledby="cta-title">
<div class="container cta-inner reveal">
<h2 id="cta-title">Tell us what your next checkpoint needs to learn.</h2>
<p>We'll come back with a pilot plan: task families, difficulty tiers, delivery format, and a calibration run on your model.</p>
<div class="hero-actions center">
<a class="btn btn-light" data-contact data-subject="Pilot inquiry" href="#contact">Email us
<svg viewBox="0 0 20 20" aria-hidden="true"><path d="M4 10h11m-4-4 4 4-4 4" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"/></svg>
</a>
<a class="btn btn-outline-light" href="#capabilities">Explore capabilities</a>
</div>
</div>
</section>
</main>
<footer class="footer">
<div class="container footer-inner">
<a class="brand" href="#top" aria-label="IntelligenceLab home">
<svg class="brand-mark" viewBox="0 0 32 32" aria-hidden="true">
<rect width="32" height="32" rx="8" fill="url(#brand-grad)"/>
<path d="M7 23h5v-5h5v-5h5" fill="none" stroke="#fff" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"/>
<path d="M19.6 9.4l2.2 2.2 4.2-4.4" fill="none" stroke="#fff" stroke-width="2.6" stroke-linecap="round" stroke-linejoin="round"/>
</svg>
<span>IntelligenceLab</span>
</a>
<nav class="footer-links" aria-label="Footer">
<a href="#examples">Examples</a>
<a href="#trajectories">Trajectories</a>
<a href="#research">Research</a>
<a data-contact data-subject="Hello" href="#contact">Contact</a>
</nav>
<p class="footer-note">© <span data-year>2026</span> IntelligenceLab. Benchmark names belong to their respective authors and are listed for skill alignment only.</p>
</div>
</footer>
<div class="ex-modal" id="ex-modal" hidden>
<div class="ex-backdrop" data-ex-close></div>
<div class="ex-sheet" role="dialog" aria-modal="true" aria-labelledby="exm-title">
<header class="ex-head">
<div>
<p class="ex-domain" id="exm-domain"></p>
<h3 id="exm-title"></h3>
<ul class="ex-tracks" id="exm-tracks"></ul>
</div>
<div class="ex-nav">
<button class="icon-btn" type="button" data-ex-prev aria-label="Previous example">←</button>
<button class="icon-btn" type="button" data-ex-next aria-label="Next example">→</button>
<button class="icon-btn" type="button" data-ex-close aria-label="Close">✕</button>
</div>
</header>
<div class="ex-body">
<figure class="ex-figure">
<div class="ex-live-hero" id="exm-live" aria-hidden="true" hidden></div>
<figcaption class="ex-evidence-label">Evidence from the task's public runs</figcaption>
<img id="exm-img" alt="" width="1600" height="939">
</figure>
<p class="ex-desc" id="exm-desc"></p>
<div class="ex-facts">
<div><h4>Software</h4><ul id="exm-software"></ul></div>
<div><h4>Agent observes</h4><ul id="exm-artifacts"></ul></div>
<div><h4>Hidden interventions</h4><ul id="exm-axes"></ul></div>
<div><h4>Graded on</h4><ul id="exm-outputs"></ul></div>
</div>
<p class="ex-how">The agent sees public runs of the real software and writes a program that reproduces it. That program is then run on hidden configurations of these parameters and must match the software's outputs, field by field.</p>
<a class="btn btn-dark" id="exm-cta" data-contact data-subject="Sample request" href="#contact">Request the full sample
<svg viewBox="0 0 20 20" aria-hidden="true"><path d="M4 10h11m-4-4 4 4-4 4" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"/></svg>
</a>
</div>
</div>
</div>
<div class="ex-modal" id="tx-modal" hidden>
<div class="ex-backdrop" data-tx-close></div>
<div class="ex-sheet" role="dialog" aria-modal="true" aria-labelledby="txm-title">
<header class="ex-head">
<div>
<p class="ex-domain" id="txm-area"></p>
<h3 id="txm-title"></h3>
<ul class="ex-tracks" id="txm-tags"></ul>
</div>
<div class="ex-nav">
<button class="icon-btn" type="button" data-tx-prev aria-label="Previous task">←</button>
<button class="icon-btn" type="button" data-tx-next aria-label="Next task">→</button>
<button class="icon-btn" type="button" data-tx-close aria-label="Close">✕</button>
</div>
</header>
<div class="ex-body">
<div class="txm-grid">
<div class="txm-col">
<h4 class="txm-h">Instruction the agent receives</h4>
<div class="txm-instruction" id="txm-instruction"></div>
<dl class="txm-facts">
<div><dt>Environment</dt><dd id="txm-base"></dd></div>
<div><dt>Reference solution</dt><dd id="txm-lines"></dd></div>
<div><dt>Verification</dt><dd>The reference solution passes every check in a fresh sandbox, and each check is stated in the instruction.</dd></div>
</dl>
</div>
<div class="txm-col">
<h4 class="txm-h">Replay of the reference solution</h4>
<div class="tx-term tx-term-lg" aria-hidden="true">
<div class="tx-bar"><i></i><i></i><i></i><em id="txm-term-base"></em></div>
<pre class="tx-screen" id="txm-replay"><code></code></pre>
</div>
<h4 class="txm-h">Excerpt of <code>solve.sh</code></h4>
<pre class="txm-code" id="txm-code"></pre>
</div>
</div>
<a class="btn btn-dark" id="txm-cta" data-contact data-subject="Terminal task sample request" href="#contact">Request the full task
<svg viewBox="0 0 20 20" aria-hidden="true"><path d="M4 10h11m-4-4 4 4-4 4" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"/></svg>
</a>
</div>
</div>
</div>
<script type="application/json" id="examples-data">[{"kind": "stl", "title": "OpenSCAD / BOSL2 parametric fan duct", "domain": "Spatial, Visual & 3D", "color": "#7c3aed", "description": "Regenerate a swept fan-duct mount STL whose inlet size, outlet size, bend angle, and rib count change the geometry.", "software": ["OpenSCAD", "BOSL2"], "artifacts": ["STL mesh", "geometry metrics JSON"], "axes": ["inlet size", "outlet size", "bend angle", "rib count"], "outputs": ["bounding box", "volume", "surface area", "connected components", "sampled surface distance"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "3D mesh", "img": "assets/examples/stl-detail.webp", "alt": "Rendered fan-duct mesh from a public OpenSCAD run"}, {"kind": "foam", "title": "OpenFOAM 2-D dam-break CFD", "domain": "Physical & Engineering Simulation", "color": "#0891b2", "description": "Reconstruct a 2-D dam-break CFD workflow across water height, mesh, gravity, and simulation time.", "software": ["OpenFOAM"], "artifacts": ["blockMesh log", "polyMesh", "initial field files", "case dictionaries"], "axes": ["water height", "mesh", "gravity", "time"], "outputs": ["mesh statistics", "field summaries", "residual/probe behavior"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "CFD mesh", "img": "assets/examples/foam-detail.webp", "alt": "OpenFOAM dam-break mesh drawn from the public polyMesh"}, {"kind": "drc", "title": "KiCad LED-matrix driver PCB", "domain": "Spatial, Visual & 3D", "color": "#7c3aed", "description": "Reconstruct LED-matrix driver board outputs (DRC report and project files) under row, column, driver, and routing interventions.", "software": ["KiCad"], "artifacts": ["DRC report", "project file list"], "axes": ["rows", "cols", "driver", "routing"], "outputs": ["DRC report summary", "project file inventory"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "PCB layout", "img": "assets/examples/drc-detail.webp", "alt": "KiCad board with design-rule findings plotted at their positions"}, {"kind": "container", "title": "Container SBOM and supply-chain attestation", "domain": "Software & Formal Systems", "color": "#2563eb", "description": "Infer coupled container base-image, package-set, label, entrypoint, and SBOM-scope behavior.", "software": ["Hadolint", "Syft", "Skopeo"], "artifacts": ["Dockerfile", "in-toto attestation", "SPDX SBOM"], "axes": ["base image", "entrypoint", "labels", "package set", "SBOM scope"], "outputs": ["attestation", "base", "layer count", "merkle", "vuln budget"], "tracks": ["Terminal", "Coding"], "modality": "", "img": "assets/examples/container-detail.webp", "alt": "Container image layers and attestation from a public run"}, {"kind": "blender", "title": "Blender cloth flag scene", "domain": "Spatial, Visual & 3D", "color": "#7c3aed", "description": "Reconstruct cloth-scene behavior under wind, pin-group, frame, and subdivision interventions.", "software": ["Blender"], "artifacts": ["scene graph JSON"], "axes": ["wind strength", "pin group", "frame", "cloth subdiv"], "outputs": ["evaluated objects", "frame", "materials", "node groups", "object counts", "rotation"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "3D scene", "img": "assets/examples/blender-detail.webp", "alt": "Blender cloth scene rotation dial across public scenarios"}, {"kind": "robot", "title": "MuJoCo / URDF mass-inertia repair", "domain": "Physical & Engineering Simulation", "color": "#0891b2", "description": "Infer robot state and trajectory behavior under actuator, collision, pose, joint-limit, and target interventions.", "software": ["MuJoCo", "URDF/MJCF tooling"], "artifacts": ["URDF model", "trajectory CSV", "control summary"], "axes": ["actuator gain", "collision margin", "initial pose", "joint limit", "trajectory target"], "outputs": ["gain", "nq", "nv", "qpos", "qvel"], "tracks": ["Terminal", "Coding"], "modality": "", "img": "assets/examples/robot-detail.webp", "alt": "MuJoCo joint-position trajectories across public scenarios"}, {"kind": "micro", "title": "Microscopy cell segmentation and counting", "domain": "Life & Molecular Sciences", "color": "#16a34a", "description": "Reconstruct cell segmentation and measurement behavior across channel, illumination, threshold, object size, and measurement feature.", "software": ["CellProfiler", "scikit-image", "Bio-Formats"], "artifacts": ["label-mask image", "object-measurements CSV"], "axes": ["channel selection", "illumination correction", "measurement feature", "object size range", "threshold method"], "outputs": ["object count", "area sum", "eccentricity sum", "threshold", "minimum size"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "Microscopy images", "img": "assets/examples/micro-detail.webp", "alt": "Microscopy label masks from public runs"}, {"kind": "astro", "title": "Astronomy FITS/WCS cutout and photometry", "domain": "Earth & Space Sciences", "color": "#db2777", "description": "Reconstruct FITS/WCS cutout and photometry behavior under aperture, background, clipping, threshold, and coordinate-frame interventions.", "software": ["Astropy", "Photutils"], "artifacts": ["FITS image", "photometry catalog CSV"], "axes": ["aperture radius", "background model", "sigma clip", "source threshold", "WCS frame"], "outputs": ["flux", "coordinate frame", "median", "shape", "standard deviation"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "FITS images", "img": "assets/examples/astro-detail.webp", "alt": "Astronomy FITS cutouts from public runs"}, {"kind": "ocr", "title": "Tesseract table-image OCR", "domain": "Data & Document Workflows", "color": "#d97706", "description": "Reconstruct how DPI, font, language, metadata, and page range affect OCR behavior.", "software": ["Pandoc", "Poppler", "Tesseract", "QPDF"], "artifacts": ["OCR summary JSON", "page-preview image"], "axes": ["DPI", "font", "language", "metadata", "page range"], "outputs": ["ink", "line count", "page size", "text digest", "text length"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "Document images", "img": "assets/examples/ocr-detail.webp", "alt": "OCR page preview from a public run"}, {"kind": "color", "title": "Color-space and tone-curve pipeline", "domain": "Spatial, Visual & 3D", "color": "#7c3aed", "description": "Infer a color pipeline across color space, white point, tone curve, demosaicing, and compression quality.", "software": ["colour-science", "ImageMagick", "scikit-image"], "artifacts": ["color measurements JSON", "PPM preview"], "axes": ["color space", "white point", "tone curve", "demosaic method", "compression quality"], "outputs": ["mean color", "image shape", "XYZ tristimulus values"], "tracks": ["Terminal", "Coding", "Multimodal"], "modality": "Images", "img": "assets/examples/color-detail.webp", "alt": "Color-pipeline previews from public runs"}]</script>
<script type="application/json" id="terminal-data">{"tasks": [{"title": "Repair and cross-build an AArch64 assembly program", "area": "Software engineering", "summary": "Fix the assembly source at /app/horas_seg.s so it prompts for integer hours and correctly prints the equivalent seconds (e.g., 2 → 7200 segundos.).", "instruction": "<p>Fix the assembly source at /app/horas_seg.s so it prompts for integer hours and correctly prints the equivalent seconds (e.g., 2 → 7200 segundos.).<\/p><p>Compile the repaired source into /app/horas_seg.<\/p><p>Write a self-validating build script at /app/build.sh and make it executable.<\/p><p>The script must source /app/toolchain.cfg to obtain CROSS_COMPILE, use the prefix to locate the cross-assembler and linker (no hardcoded paths), assemble and link the program, then validate the resulting binary by running it with input 2 and checking that the output contains "7200 segundos.".<\/p><p>On success the script must print "BUILD SUCCESS" and exit 0; on failure it must print "VALIDATION FAILED" and exit non-zero.<\/p><p>The sysroot at /usr/aarch64-linux-gnu is available for cross-execution via QEMU if needed.<\/p><p>Generate a build manifest at /app/build-info.txt using the format:<\/p><p>as_version: <assembler version string><\/p><p>ld_version: <linker version string><\/p><p>sha256: <64-character lowercase hex hash of /app/horas_seg><\/p><p>Create a symlink /usr/local/bin/horas_seg → /app/horas_seg.<\/p>", "base": "ubuntu:22.04", "lines": 239, "tags": ["assembly", "arm64", "aarch64", "compilation"], "replay": [["$", "source \"$CFG\""], ["$", "\"$AS\" --version"], ["$", "cat > /app/horas_seg.s << 'SOURCEEOF'"], ["$", "TEST_OUT=$(run_program \"2\" 2>&1)"], ["$", "chmod +x /app/build.sh"], ["$", "ln -sf /app/horas_seg /usr/local/bin/horas_seg"]], "code": "set -e\nCFG=\"/app/toolchain.cfg\"\nif [ ! -f \"$CFG\" ]; then\n echo \"ERROR: Toolchain configuration file $CFG not found\"\n exit 1\nfi\nsource \"$CFG\"\nif [ -z \"$CROSS_COMPILE\" ]; then\n echo \"ERROR: CROSS_COMPILE not defined in $CFG\"\n exit 1\nfi\nAS=$(command -v \"${CROSS_COMPILE}as\")\nLD=$(command -v \"${CROSS_COMPILE}ld\")\nif [ -z \"$AS\" ]; then\n echo \"ERROR: ${CROSS_COMPILE}as not found in PATH\"\n exit 1\nfi\nif [ -z \"$LD\" ]; then\n echo \"ERROR: ${CROSS_COMPILE}ld not found in PATH\"\n exit 1\nfi\n\"$AS\" --version >/dev/null 2>&1 || { echo \"ERROR: $AS does not work\"; exit 1; }\n\"$LD\" --version >/dev/null 2>&1 || { echo \"ERROR: $LD does not work\"; exit 1; }\nrun_program() {\n local input_str=\"$1\"\n if /app/horas_seg <<< \"$input_str\" 2>/dev/null; then\n return 0\n fi\n if command -v qemu-aarch64 >/dev/null 2>&1; then\n local qemu_args=(\"qemu-aarch64\")"}, {"title": "Wire a network namespace with a veth pair", "area": "Networking", "summary": "Set up a Linux network namespace `t1` with a veth pair.", "instruction": "<p>Set up a Linux network namespace <code>t1<\/code> with a veth pair.<\/p><p>Inside <code>t1<\/code> assign <code>10.0.0.1/24<\/code>; on the host assign <code>10.0.0.2/24<\/code>.<\/p><p>Bring all interfaces up and enable loopback inside the namespace.<\/p><p>Derive the MTU from <code>/app/veth_mtu.cfg<\/code> and the host kernel version; apply that MTU to both veth endpoints and write the same value to <code>/etc/veth_mtu_policy.conf<\/code>.<\/p><p>Confirm bidirectional pings between the two IPs.<\/p><p>Produce a network state report at <code>/app/network_state_report.txt<\/code> following <code>/app/REPORT_SPEC.md<\/code>.<\/p><p>Record the derivation steps and chosen MTU in <code>/tmp/veth_debug.log<\/code>.<\/p><p>Run the validation script <code>/app/validate_report.sh<\/code>.<\/p><p>If it fails, interpret the failure, correct the underlying problem, and re‑run the script until it succeeds.<\/p>", "base": "ubuntu:22.04", "lines": 213, "tags": ["network-namespace", "veth", "iproute2", "linux-networking"], "replay": [["$", "ip netns add t1"], ["$", "ip link add veth0 mtu \"$CHOSEN\" type veth peer name veth1 mtu \"$CHOSEN\""], ["$", "ip link set veth0 netns t1"], ["$", "ip netns exec t1 ip addr add 10.0.0.1/24 dev veth0"], ["$", "ip addr add 10.0.0.2/24 dev veth1"], ["$", "ip netns exec t1 ip link set veth0 up"], ["$", "ip link set veth1 up"]], "code": "set -e\nCFG=\"/app/veth_mtu.cfg\"\nREPORT=\"/app/network_state_report.txt\"\nPOLICY=\"/etc/veth_mtu_policy.conf\"\nDEBUG_LOG=\"/tmp/veth_debug.log\"\nVALIDATE_SCRIPT=\"/app/validate_report.sh\"\nSPEC=\"/app/REPORT_SPEC.md\"\necho \"=== Inspect configuration and specification ===\"\nif [ ! -f \"$CFG\" ]; then\n echo \"Error: $CFG not found\" >&2\n exit 1\nfi\nif [ ! -f \"$SPEC\" ]; then\n echo \"Error: $SPEC not found\" >&2\n exit 1\nfi\nDEFAULT_MTU=$(grep -oP '^mtu=\\K[0-9]+' \"$CFG\" || true)\nif [ -z \"$DEFAULT_MTU\" ]; then\n echo \"Error: no default mtu= line in config\" >&2\n exit 1\nfi\nCOND=$(grep -oP '^if_kernel_gte=\\K[^:]+:[0-9]+' \"$CFG\" || true)\nif [ -n \"$COND\" ]; then\n THRESHOLD=$(echo \"$COND\" | cut -d: -f1)\n COND_MTU=$(echo \"$COND\" | cut -d: -f2)\nelse\n THRESHOLD=\"\"\n COND_MTU=\"\"\nfi\nKV=$(uname -r | cut -d- -f1)"}, {"title": "Align a container build with its Kustomize manifest", "area": "Containers & cloud", "summary": "Align the Dockerfile at /environment/Dockerfile with the files referenced by the kustomization in /app (kustomization.yaml), notably config.yaml.", "instruction": "<p>Align the Dockerfile at /environment/Dockerfile with the files referenced by the kustomization in /app (kustomization.yaml), notably config.yaml.<\/p><p>Generate all required deliverables: result.txt (containing 'apiVersion=v1'), integrity checksums, a summary, a build log, and an artifact inventory.<\/p><p>Use the inventory generator script at /opt/sources/templates/inventory_gen.sh — follow /environment/GENERATOR.md for its working directory and invocation.<\/p><p>Inspect /environment/REQUIREMENTS.md and /environment/EVIDENCE_GUIDE.md for additional requirements, evidence bundle specifications, and archive packaging rules.<\/p><p>Assemble an evidence bundle in /app/evidence that includes the deliverables and a manifest, then create a compressed tar archive named evidence.tar.gz inside that directory that contains the bundle contents.<\/p><p>Before final verification, update config.yaml according to /environment/CONFIG_SCHEMA.md and validate it with /opt/tools/validate_config.sh.<\/p><p>Finally, run /opt/tools/verify.sh to complete verification (it now also checks the evidence archive).<\/p>", "base": "ubuntu:22.04", "lines": 311, "tags": ["kubernetes", "kustomize", "secrets"], "replay": [["$", "cp -f \"$SOURCES_DIR/config.yaml\" \"$APP_DIR/config.yaml\""], ["$", "sed -i \"/^CMD /i COPY config.yaml /app/config.yaml\" \"$DOCKERFILE\""], ["$", "sha256sum \"$APP_DIR/$f\" >> \"$INTEGRITY\""], ["$", "kustomize build \"$APP_DIR\" > \"$RESULT\""], ["$", "sed -i 's/^apiVersion: v1$/apiVersion=v1/' \"$RESULT\""], ["$", "tar czf evidence.tar.gz *"]], "code": "set -euo pipefail\nAPP_DIR=\"/app\"\nSOURCES_DIR=\"/opt/sources\"\nDOCKERFILE=\"/environment/Dockerfile\"\nREQUIREMENTS=\"/environment/REQUIREMENTS.md\"\nRESULT=\"$APP_DIR/result.txt\"\nSUMMARY=\"$APP_DIR/summary.txt\"\nINTEGRITY=\"$APP_DIR/integrity.txt\"\nINVENTORY=\"$APP_DIR/artifact_inventory.txt\"\nBUILD_LOG=\"$APP_DIR/build.log\"\nVERIFY_SCRIPT=\"/opt/tools/verify.sh\"\nVALIDATE_CONFIG=\"/opt/tools/validate_config.sh\"\nCONFIG_SCHEMA=\"/environment/CONFIG_SCHEMA.md\"\nGENERATOR_DOC=\"/environment/GENERATOR.md\"\nEVIDENCE_GUIDE=\"/environment/EVIDENCE_GUIDE.md\"\nEVIDENCE_DIR=\"/app/evidence\"\nBUNDLE_MANIFEST=\"$EVIDENCE_DIR/manifest.txt\"\nBUNDLE_SUMMARY=\"$EVIDENCE_DIR/bundle_summary.txt\"\nMAX_ATTEMPTS=5\necho \"=== Stage: inspect ===\"\nfor f in \"$VERIFY_SCRIPT\" \"$DOCKERFILE\" \"$APP_DIR/kustomization.yaml\" \"$SOURCES_DIR/config.yaml\" \"$REQUIREMENTS\" \"$GENERATOR_DOC\" \"$CONFIG_SCHEMA\" \"$VALIDATE_CONFIG\" \"$EVIDENCE_GUIDE\"; do\n if [ ! -f \"$f\" ]; then\n echo \"ERROR: $f not found.\" >&2\n exit 1\n fi\ndone\nif [ ! -x \"$VALIDATE_CONFIG\" ]; then\n echo \"ERROR: $VALIDATE_CONFIG is not executable.\" >&2\n exit 1\nfi"}, {"title": "Stand up InfluxDB 2 against a compliance policy", "area": "Databases", "summary": "Set up InfluxDB 2 using the configuration at /etc/influxdb/init.json.", "instruction": "<p>Set up InfluxDB 2 using the configuration at /etc/influxdb/init.json.<\/p><p>A policy file at /etc/influxdb/schema_requirements.json defines required rules—including organization, bucket, retention, and bucket description—that your setup must match.<\/p><p>Use the compliance check script in /app/ to verify your configuration against these rules, then run the validation script and iterate until all checks pass.<\/p><p>Both compliance and validation must succeed before finalizing.<\/p><p>Produce a report at /app/result.txt containing the active connection configuration, listings of all organizations and buckets, a reconciliation section that shows MATCH for every parameter, and the final output of the compliance and validation steps.<\/p><p>Avoid destructive actions if the database already holds configuration.<\/p>", "base": "debian:bookworm", "lines": 329, "tags": ["influxdb", "database", "system-administration"], "replay": [["$", "jq empty \"$f\""], ["$", "influx setup --username \"$USERNAME\" --password \"$PASSWORD\" --org \"$ORG\" --bucket \"$BUCKET\" --retention…"], ["$", "influx bucket update --id \"$BID\" --retention \"$RETENTION\""], ["$", "influx bucket update --id \"$BID\" --description \"$REQUIRED_BUCKET_DESC\""], ["$", "chmod +x /app/validate.sh"]], "code": "set -e\nfor i in {1..30}; do\n if curl -s http://localhost:8086/ping >/dev/null 2>&1; then\n break\n fi\n sleep 1\ndone\nif ! command -v jq >/dev/null 2>&1; then\n apt-get update -qq && apt-get install -y -qq jq\nfi\nfor f in /etc/influxdb/init.json /etc/influxdb/schema_requirements.json; do\n if [ ! -f \"$f\" ]; then\n echo \"Error: missing evidence file $f\" >&2\n exit 1\n fi\n jq empty \"$f\" || {\n echo \"Error: invalid JSON in $f\" >&2\n exit 1\n }\ndone\nREQUIRED_ORG=$(jq -r '.org // empty' /etc/influxdb/schema_requirements.json)\nREQUIRED_BUCKET=$(jq -r '.bucket // empty' /etc/influxdb/schema_requirements.json)\nREQUIRED_RETENTION=$(jq -r '.retention // \"0s\"' /etc/influxdb/schema_requirements.json)\nREQUIRED_BUCKET_DESC=$(jq -r '.bucket_description // empty' /etc/influxdb/schema_requirements.json)\nUSERNAME=$(jq -r '.username' /etc/influxdb/init.json)\nPASSWORD=$(jq -r '.password' /etc/influxdb/init.json)\nORG=$(jq -r '.org' /etc/influxdb/init.json)\nBUCKET=$(jq -r '.bucket' /etc/influxdb/init.json)\nRETENTION=$(jq -r '.retention' /etc/influxdb/init.json)\nCHANGED=0"}, {"title": "Version a dataset with DVC and reconcile its history", "area": "Version control", "summary": "Set up a DVC-tracked dataset under /app/dataset/ with two items: dataset/foo and dataset/test/0.", "instruction": "<p>Set up a DVC-tracked dataset under /app/dataset/ with two items: dataset/foo and dataset/test/0.<\/p><p>Track both with DVC, remove the original data, and commit the resulting .dvc files.<\/p><p>Then produce the following artifacts under /app: dvc_inventory.json, dvc_reconciliation.md, dvc_changes.diff, dvc_status.txt, and a directory /app/patches/ containing a patch file for each of dataset/foo.dvc and dataset/test/0.dvc that captures the changes introduced when that file was first committed.<\/p><p>The reconciliation report must document the absence of the data, the git commits that introduced each .dvc file (SHA and message), and the baseline tag (name and commit SHA).<\/p><p>The diff should be taken from that tag to HEAD.<\/p><p>The repository already has git and DVC initialized, and a git tag marks the initial DVC initialization; discover and use that tag as the baseline.<\/p><p>Refer to /app/task_files/ for the required formats of the inventory and the reconciliation report.<\/p>", "base": "ubuntu:22.04", "lines": 313, "tags": ["dvc", "data-versioning", "version-control"], "replay": [["$", "dvc add dataset/foo"], ["$", "dvc add dataset/test/0"], ["$", "git add dataset/foo.dvc dataset/test/0.dvc"], ["$", "git commit -m \"DVC track dataset items\""], ["$", "git diff \"${PARENT}..${ADD_COMMIT}\" -- \"$dvc_file\" > \"$patch_path\""]], "code": "set -euo pipefail\nif [ -s /app/dvc_changes.diff ] && \\\n [ -s /app/dvc_reconciliation.md ] && \\\n [ -s /app/dvc_inventory.json ] && \\\n [ -s /app/dvc_status.txt ] && \\\n grep -q 'Commit message:' /app/dvc_reconciliation.md && \\\n grep -q 'Baseline tag:' /app/dvc_reconciliation.md && \\\n [ -d /app/patches ] && \\\n [ \"$(ls -1 /app/patches/*.patch 2>/dev/null | wc -l)\" -ge 2 ]; then\n echo \"Required artifacts with commit messages, baseline tag, and patches already present. Nothing to do.\"\n exit 0\nfi\ncd /app\nrm -f dataset/foo.dvc dataset/test/0.dvc\nrm -rf dataset/foo dataset/test/0\nrm -f /app/dvc_reconciliation.md \\\n /app/dvc_inventory.json \\\n /app/dvc_changes.diff \\\n /app/dvc_status.txt\nrm -rf /app/patches\nINIT_COMMIT=$(git rev-list --max-parents=0 HEAD)\nif [ -z \"$INIT_COMMIT\" ]; then\n echo \"ERROR: could not determine initial commit\"\n exit 1\nfi\necho \"Initial commit: $INIT_COMMIT\"\nreadarray -t TAGS < <(git tag --points-at \"$INIT_COMMIT\")\nif [ \"${#TAGS[@]}\" -eq 0 ]; then\n echo \"ERROR: no tag found pointing to the initial commit\"\n exit 1"}, {"title": "Install an HPC scheduler's Python bindings, verified", "area": "Data & scientific computing", "summary": "Install the Flux Python package from the tarball at /app/flux-python-0.48.0rc6.tar.gz after verifying its integrity against the provided checksum file.", "instruction": "<p>Install the Flux Python package from the tarball at /app/flux-python-0.48.0rc6.tar.gz after verifying its integrity against the provided checksum file.<\/p><p>The expected manifest file may be outdated; if the installed layout differs, update the manifest and note the discrepancy in the installation report.<\/p><p>The installation report at /app/installation_report.txt must include the verified tarball hash, describe any manifest changes, and also contain additional package metadata specified by the configuration at /app/config/metadata.json — follow that config's instructions for intermediate reporting.<\/p><p>Finally, use the Flux Python bindings to produce /app/result.txt containing exactly the string found in /app/config/expected.txt.<\/p>", "base": "fluxrm/flux-core:focal-v0.48.0", "lines": 171, "tags": ["flux", "hpc", "job-scheduling", "python-bindings"], "replay": [["$", "ACTUAL_HASH=$(sha256sum \"$TARBALL\" | awk '{print $1}')"], ["$", "pip install --no-compile \"$TARBALL\""], ["$", "find \"$FLUX_ROOT\" -type f ! -name '*.pyc' -exec sha256sum {} \\; | sort -k2 > \"$INSTALLED_MANIFEST\""], ["$", "cp \"$INSTALLED_MANIFEST\" \"$EXPECTED_MANIFEST\""], ["$", "flux start --test-size=4 bash -c '…'"]], "code": "set -euo pipefail\nTARBALL=\"/app/flux-python-0.48.0rc6.tar.gz\"\nCHECKSUM_FILE=\"/app/checksum.sha256\"\nVALIDATION_FILE=\"/app/validation.txt\"\nCONFIG_FILE=\"/app/config/expected.txt\"\nRESULT_FILE=\"/app/result.txt\"\nEXPECTED_MANIFEST=\"/app/expected_manifest.txt\"\nINSTALLED_MANIFEST=\"/app/installed_manifest.txt\"\nINSTALLATION_REPORT=\"/app/installation_report.txt\"\nMETADATA_CONFIG=\"/app/config/metadata.json\"\necho \"=== INSPECT: verify required inputs exist ===\"\nfor f in \"$TARBALL\" \"$CHECKSUM_FILE\" \"$EXPECTED_MANIFEST\" \"$CONFIG_FILE\" \"$METADATA_CONFIG\"; do\n if [ ! -f \"$f\" ]; then\n echo \"ERROR: Required file $f not found\" >&2\n exit 1\n fi\ndone\necho \"=== INSPECT: read metadata configuration ===\"\nMETADATA_FILE=$(python3 -c \"import json; cfg=json.load(open('$METADATA_CONFIG')); print(cfg['metadata_file'])\")\nFIELDS=$(python3 -c \"import json; cfg=json.load(open('$METADATA_CONFIG')); print(' '.join(cfg['required_metadata']))\")\necho \"Metadata will be gathered for fields: $FIELDS\"\necho \"Intermediate metadata file: /app/$METADATA_FILE\"\necho \"=== DERIVE: verify tarball hash ===\"\nACTUAL_HASH=$(sha256sum \"$TARBALL\" | awk '{print $1}')\nEXPECTED_HASH=$(awk '{print $1}' \"$CHECKSUM_FILE\")\nif [ \"$ACTUAL_HASH\" != \"$EXPECTED_HASH\" ]; then\n echo \"ERROR: Tarball checksum mismatch\" >&2\n exit 1\nfi\necho \"=== VALIDATE checksum OK, writing validation report ===\""}, {"title": "Rebuild a GRUB boot sequence from a broken system", "area": "System administration", "summary": "Assemble a three-line GRUB boot sequence (linux, initrd, boot) in /app/result.txt for the system rooted at /mnt/target.", "instruction": "<p>Assemble a three-line GRUB boot sequence (linux, initrd, boot) in /app/result.txt for the system rooted at /mnt/target.<\/p><p>Use /app/boot.cfg for console and baud (the linux line must contain <code>console=ttyS0,115200<\/code>).<\/p><p>The target's /mnt/target/etc/device-map provides the GRUB device prefix, and /mnt/target/boot/RULES.txt gives the kernel selection rule (pick the highest version).<\/p><p>Some configuration files may be missing; use documentation under /mnt/target/etc/ to reconstruct them.<\/p><p>After building the sequence, run /app/check_boot.sh to validate.<\/p><p>Its log in /var/log/boot_check.log will report missing boot-params.<\/p><p>Inspect the log, fix any issues, re-run the check until it returns no errors, then write a root-cause diagnosis report to /app/diagnosis.txt citing the logged evidence.<\/p>", "base": "ubuntu:22.04", "lines": 214, "tags": ["grub", "bootloader", "linux", "system-administration"], "replay": [["$", "\"$CHECK_SCRIPT\""], ["$", "ERROR_LINE=$(grep -i \"ERROR\" \"$LOG\" | grep -i \"boot-params\" | head -1 || true)"], ["$", "printf '%s\\n' \"${param_lines[@]}\" > \"$BOOT_PARAMS_FILE\""], ["$", "linux_cmd=\"linux ${grub_prefix}/${actual_kernel_path} console=${console},${baud}\""], ["$", "initrd_cmd=\"initrd ${grub_prefix}/${initrd_path}\""], ["$", "cat > \"$RESULT\" <<EOF"]], "code": "set -euo pipefail\nmkdir -p /var/log\nLOG=\"/var/log/boot_check.log\"\nCHECK_SCRIPT=\"/app/check_boot.sh\"\nRESULT=\"/app/result.txt\"\nDIAGNOSIS=\"/app/diagnosis.txt\"\nTARGET=\"/mnt/target\"\nBOOT_PARAMS_FILE=\"${TARGET}/etc/boot-params\"\nSPEC_FILE=\"${TARGET}/etc/params-spec.txt\"\nDEVICE_MAP_FILE=\"${TARGET}/etc/device-map\"\nRULES_FILE=\"${TARGET}/boot/RULES.txt\"\nCONFIG=\"/app/boot.cfg\"\nif [[ ! -x \"$CHECK_SCRIPT\" ]]; then\n echo \"Error: $CHECK_SCRIPT not found or not executable\" >&2\n exit 1\nfi\necho \"Running initial boot check...\" >&2\nset +e # initial run is expected to fail because boot-params missing\n\"$CHECK_SCRIPT\"\ninitial_exit=$?\nset -e\nif [[ ! -f \"$LOG\" ]]; then\n echo \"Error: Log file $LOG was not created by $CHECK_SCRIPT\" >&2\n exit 1\nfi\necho \"Inspecting $LOG...\" >&2\nERROR_LINE=$(grep -i \"ERROR\" \"$LOG\" | grep -i \"boot-params\" | head -1 || true)\nif [[ -z \"$ERROR_LINE\" ]]; then\n echo \"Warning: No ERROR about boot-params found in $LOG. Proceeding without repair.\" >&2\n ROOT_CAUSE=\"No error detected; boot-params may already exist.\""}, {"title": "Diagnose and fix a Python dependency bug", "area": "Debugging & performance", "summary": "In /app, diagnose and fix the environs URL parsing bug.", "instruction": "<p>In /app, diagnose and fix the environs URL parsing bug.<\/p><p>Run bug.py to capture the original output and use validate_output.py to diagnose the failure.<\/p><p>Consult changelog.md to determine the required environs version, upgrade the library, and produce a corrected output in fixed.txt.<\/p><p>Write a test suite test_bug.py covering both mongodb and mongodb+srv URL schemes.<\/p><p>The suite must include a test that runs validate_output.py against bug.py’s output after the fix and asserts the validator exits successfully.<\/p><p>Run the tests and save the report as test_report.txt.<\/p><p>Finally, create validation_passed.txt and a debug summary debug_summary.txt describing the error and the fix.<\/p>", "base": "python:3.11-slim", "lines": 202, "tags": ["python", "library-bug", "url-parsing", "environs"], "replay": [["$", "python /app/bug.py > /app/result.txt"], ["$", "/app/validate_output.py /app/result.txt > /app/validation_stdout.txt 2> /app/validation_stderr.txt"], ["$", "pip install \"environs==${REQUIRED_VERSION}\" 'marshmallow<4.0.0'"], ["$", "python /app/bug.py > /app/fixed.txt"]], "code": "set -euo pipefail\necho \"Ensuring pytest is available\"\npython -m pytest --version >/dev/null 2>&1 || pip install pytest\necho \"Stage 1: Create original bug.py (known buggy)\"\ncat > /app/bug.py << 'EOF'\nfrom environs import Env\nenv = Env()\nMONGODB_URL = env.url(\n \"MONGODB_URL\",\n default=\"mongodb://user:pass@mongo.example.local/db?authSource=admin\",\n schemes={\"mongodb\", \"mongodb+srv\"},\n require_tld=False,\n)\nprint(MONGODB_URL)\nEOF\necho \"Stage 2: Run bug.py and capture original output\"\npython /app/bug.py > /app/result.txt\necho \"Stage 3: Validate original output and interpret failure\"\nset +e\n/app/validate_output.py /app/result.txt > /app/validation_stdout.txt 2> /app/validation_stderr.txt\nVALIDATION_EXIT=$?\nset -e\nif [ $VALIDATION_EXIT -eq 0 ]; then\n echo \"ERROR: validate_output.py unexpectedly succeeded on buggy output\" >&2\n exit 1\nfi\nif ! grep -q \"ParseResult\" /app/validation_stderr.txt; then\n echo \"ERROR: Expected 'ParseResult' indication in validator failure output\" >&2\n exit 1\nfi"}, {"title": "Count attack telemetry with the EQL shell", "area": "Security analytics", "summary": "Working in /app, count each event_type in the dataset '/app/normalized-atomic-red-team.json.gz' using the EQL shell and save results to counts.json.", "instruction": "<p>Working in /app, count each event_type in the dataset '/app/normalized-atomic-red-team.json.gz' using the EQL shell and save results to counts.json.<\/p><p>Generate a Markdown summary table in counts_summary.md.<\/p><p>Compare the actual counts to the expected values in config/expected_counts.json.<\/p><p>Before repairing any mismatches, run the local validation script ./check.sh and capture its exact failure output to /app/initial_failure.txt.<\/p><p>The initial failure output includes a network event type count mismatch—ensure this output is saved before applying fixes.<\/p><p>Then, using the validation script and the failure information, fix the mismatches: update config/expected_counts.json (keeping its metadata), record changes in repair_note.txt, and iterate until check.sh passes.<\/p><p>Finally, produce validation_report.json with status 'pass' and empty mismatches, and create /app/analysis_metadata.json as described in README.md.<\/p><p>See README.md for the full list of deliverables.<\/p>", "base": "python:3.9-slim", "lines": 242, "tags": ["eql", "event-query-language", "data-analysis", "json"], "replay": [["$", "eql shell << 'EQL_IN'"], ["eql>", "input normalized-atomic-red-team.json.gz"], ["eql>", "search | count event_type"], ["eql>", "output counts.json"], ["$", "/app/check.sh > /app/initial_failure.txt 2>&1"], ["$", "python3 << 'PYREPAIR'"]], "code": "set -euo pipefail\ncd /app\necho \"[1/6] Running EQL analysis...\"\neql shell << 'EQL_IN'\ninput normalized-atomic-red-team.json.gz\nsearch | count event_type\noutput counts.json\nEQL_IN\nif [ ! -s counts.json ]; then\n echo \"ERROR: counts.json is missing or empty after EQL run.\" >&2\n exit 2\nfi\necho \"[1/6] counts.json generated.\"\necho \"[2/6] Generating markdown summary...\"\npython3 << 'PYMDS'\nimport json\ndef load_counts(path):\n with open(path) as f:\n raw = f.read().strip()\n try:\n data = json.loads(raw)\n except json.JSONDecodeError:\n data = [json.loads(line) for line in raw.splitlines() if line.strip()]\n if isinstance(data, dict):\n return {k: v for k, v in data.items() if isinstance(v, int)}\n counts = {}\n for item in data:\n if not isinstance(item, dict):\n continue\n key = item.get(\"event_type\", item.get(\"key\"))"}], "growth": {"goal": "Build a Git commit using only low-level plumbing commands", "stages": [{"lines": 21, "words": 97, "segments": [{"html": "Create a Git repository at <code>/tmp/test<\/code> and build a single commit using only low-level plumbing commands—no <code>git add<\/code>, <code>git commit<\/code>, or other porcelain.", "new": false}, {"html": "Before creating the commit, source the environment file at <code>/env/git-env.sh<\/code> to provide the author and committer identity.", "new": false}, {"html": "The commit must contain a single file <code>a.txt<\/code> with the content <code>test content<\/code> (followed by a newline, as <code>echo<\/code> produces), have the commit message <code>commit 1<\/code>, and be reachable from the <code>master<\/code> branch.", "new": false}, {"html": "When finished, <code>git log master<\/code> should show exactly that one commit and <code>git cat-file -p<\/code> on the commit’s tree object should list only <code>a.txt<\/code>.", "new": false}]}, {"lines": 95, "words": 107, "segments": [{"html": "Create a Git repository at <code>/tmp/test<\/code> and construct a complete commit using only Git's low-level plumbing commands (no <code>git add<\/code>, <code>git commit<\/code>, etc.).", "new": false}, {"html": "The commit must contain a single file <code>a.txt<\/code> with the exact content <code>test content<\/code> followed by a newline, have the commit message <code>commit 1<\/code>, and be reachable from the <code>master<\/code> branch (<code>git log master<\/code> should show it as the sole commit).", "new": false}, {"html": "After the commit is ready, produce a JSON snapshot of the repository's commit structure at <code>/tmp/test/repo_snapshot.json<\/code>.", "new": true}, {"html": "Validate it with <code>/usr/local/bin/validate_snapshot.sh<\/code>.", "new": true}, {"html": "Then produce a YAML snapshot at <code>/tmp/test/repo_snapshot.yaml<\/code> that captures the same semantic information.", "new": true}, {"html": "Validate it with <code>/usr/local/bin/validate_snapshot_yaml.sh<\/code>.", "new": true}, {"html": "Both validation scripts must exit successfully.", "new": true}]}, {"lines": 114, "words": 76, "segments": [{"html": "Create a Git repository at <code>/tmp/test<\/code> using only plumbing commands (no <code>git add<\/code> or <code>git commit<\/code>).", "new": false}, {"html": "The repository must contain a single commit on branch <code>master<\/code> with file <code>a.txt<\/code> (content <code>test content<\/code>) and a file <code>b.txt<\/code>—determine the correct content for <code>b.txt<\/code> from available environment clues.", "new": true}, {"html": "After the commit, produce an object manifest at <code>/tmp/test/manifest.txt<\/code> following the format described in an environment file named <code>manifest_schema.txt<\/code> (locate and refer to it).", "new": true}, {"html": "Then generate a reconciliation report at <code>/tmp/test/report.txt<\/code>.", "new": true}]}, {"lines": 157, "words": 89, "segments": [{"html": "Create a Git repository at <code>/tmp/test<\/code> using only plumbing commands (no <code>git add<\/code> or <code>git commit<\/code>).", "new": false}, {"html": "The repository must contain a single commit on branch <code>master<\/code> with message <code>first commit<\/code>, containing file <code>a.txt<\/code> (content <code>test content<\/code>) and a file <code>b.txt<\/code>—determine the correct content for <code>b.txt<\/code> from available environment clues.", "new": false}, {"html": "After the commit, produce an object manifest at <code>/tmp/test/manifest.txt<\/code> following the format described in an environment file named <code>manifest_schema.txt<\/code> (locate and refer to it).", "new": false}, {"html": "Then generate a reconciliation report at <code>/tmp/test/report.txt<\/code> that verifies object types and sizes against the manifest.", "new": true}]}, {"lines": 187, "words": 98, "segments": [{"html": "Create a Git repository at <code>/tmp/test<\/code> using only plumbing commands (no <code>git add<\/code>, <code>git commit<\/code>).", "new": false}, {"html": "The repository must contain a single commit on branch <code>master<\/code> with message <code>first commit<\/code>, including <code>a.txt<\/code> (content <code>test content<\/code>) and <code>b.txt<\/code>.", "new": false}, {"html": "Obtain the content of <code>b.txt<\/code> from the commit tagged <code>release/v1.0<\/code> in the forensic repository at <code>/app/forensics<\/code>.", "new": true}, {"html": "After the commit, produce an object manifest at <code>/tmp/test/manifest.txt<\/code> following the format specified in the <code>manifest_schema.txt<\/code> file (locate and refer to it).", "new": false}, {"html": "The manifest must include repository-relative paths for blobs.", "new": true}, {"html": "Finally, generate a reconciliation report at <code>/tmp/test/report.txt<\/code> that verifies object types and sizes against the manifest.", "new": false}]}, {"lines": 217, "words": 119, "segments": [{"html": "Create a Git repository at <code>/tmp/test<\/code> using only plumbing commands (no <code>git add<\/code>, <code>git commit<\/code>).", "new": false}, {"html": "The repository must contain a single commit on branch <code>master<\/code> with message <code>first commit<\/code>, including <code>a.txt<\/code> (content <code>test content<\/code>) and <code>b.txt<\/code>.", "new": false}, {"html": "Obtain the content of <code>b.txt<\/code> from the commit tagged <code>release/v1.0<\/code> in the forensic repository at <code>/app/forensics<\/code>.", "new": false}, {"html": "After the commit, produce an object manifest at <code>/tmp/test/manifest.txt<\/code> following the format specified in the <code>manifest_schema.txt<\/code> file (locate and refer to it).", "new": false}, {"html": "The manifest must include repository-relative paths for blobs.", "new": false}, {"html": "Generate a forensic diff report at <code>/tmp/test/forensic_diff.txt<\/code> showing the diff of <code>b.txt<\/code> between the parent of <code>release/v1.0<\/code> and the tagged commit.", "new": true}, {"html": "Finally, generate a reconciliation report at <code>/tmp/test/report.txt<\/code> that verifies object types and sizes against the manifest.", "new": false}]}, {"lines": 236, "words": 139, "segments": [{"html": "Create a Git repository at /tmp/test using only plumbing commands (no <code>git add<\/code>, <code>git commit<\/code>).", "new": false}, {"html": "The repository must contain a single commit on branch <code>master<\/code> with message <code>first commit<\/code>.", "new": false}, {"html": "Include <code>a.txt<\/code> with content <code>test content<\/code> and <code>b.txt<\/code>.", "new": true}, {"html": "Obtain the content of <code>b.txt<\/code> from the commit tagged <code>release/v1.0<\/code> in the forensic repository at /app/forensics.", "new": false}, {"html": "After the commit, produce an object manifest at /tmp/test/manifest.txt following the format described in the manifest schema file.", "new": false}, {"html": "Use the workspace README at /app/README.md to locate that schema file.", "new": true}, {"html": "The manifest must list repository-relative paths for blobs.", "new": false}, {"html": "Generate a forensic diff report at /tmp/test/forensic_diff.txt showing the diff of <code>b.txt<\/code> between the parent of <code>release/v1.0<\/code> and the tagged commit.", "new": false}, {"html": "The report must be in diff format (including the 'diff --git' header).", "new": true}, {"html": "Finally, generate a reconciliation report at /tmp/test/report.txt that verifies object types and sizes against the manifest.", "new": false}]}, {"lines": 398, "words": 141, "segments": [{"html": "Create a Git repository at /tmp/test using only plumbing commands (no git add or git commit).", "new": false}, {"html": "The repository must have a single commit on branch master with message 'first commit', containing files a.txt (content 'test content') and b.txt.", "new": false}, {"html": "Discover the correct content for b.txt by inspecting the tagged commit release/v1.0 in the forensic repository at /app/forensics.", "new": false}, {"html": "After the commit, produce an object manifest and a reconciliation report inside the repository following the schema files in /app (manifest_schema.txt and report_schema.txt).", "new": true}, {"html": "Also create a forensic object manifest at /tmp/test/forensic-manifest.txt, listing all loose objects in /app/forensics in the same format as the manifest.", "new": true}, {"html": "The report must include lines for forensic-diff, forensic-diff-stat, forensic-message, and forensic-integrity as described in the report schema.", "new": true}, {"html": "The forensic-diff-stat line must be derived from the forensic repository's diff of b.txt between the parent of the tagged commit and the tag release/v1.0.", "new": true}]}]}}</script>
<script src="app.js" defer></script>
<script src="live.js" defer></script>
<script src="terminal.js" defer></script>
</body>
</html>
|