File size: 287,866 Bytes
d82db76 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571 5572 5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605 5606 5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640 5641 5642 5643 5644 5645 5646 5647 5648 5649 5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 | #!/usr/bin/env python3
"""Audit a completed OCR eval output directory for review readiness."""
from __future__ import annotations
import argparse
import csv
import hashlib
import html
import importlib.util
import json
import random
import shlex
import sys
import unicodedata
from datetime import datetime
from pathlib import Path
from typing import Any
import regex
from rapidfuzz.distance import Levenshtein
from ocrtech.metrics import bbox_iou
from ocrtech.references import load_reference
from ocrtech.schemas import BBox, Document, TextLine
TOOLS_DIR = Path(__file__).resolve().parent
SUPPORTED_NORMALIZATION_PROFILE = "NFC_preserve_whitespace"
UNCERTAINTY_BOOTSTRAP_SAMPLES = 200
UNCERTAINTY_BOOTSTRAP_SEED = 1729
MIN_CLAIM_SCORING_UNITS = 30
RECOGNIZER_ARTIFACTS = (
"eval-run.json",
"metrics.json",
"metrics.md",
"predictions.tsv",
"unmatched.tsv",
"worst-lines.csv",
"sample-review.csv",
"sample-review.html",
"confusions.json",
"command.txt",
"claim-safety.md",
)
PAGE_ARTIFACTS = (
"page-eval-run.json",
"metrics.json",
"metrics.md",
"page-metrics.csv",
"reference-artifacts.tsv",
"text-pairs.tsv",
"missing.tsv",
"worst-pages.csv",
"confusions.json",
"sample-review.csv",
"sample-review.html",
"line-alignment.csv",
"command.txt",
"claim-safety.md",
)
RECIPE_ROOT_ARTIFACTS = (
"recipe-run.json",
"recipe.json",
"recipe-validation.json",
"audit.json",
"claim-safety.md",
"materialize/output-command-run.json",
"materialize/output-command-run.md",
"materialize/command-template.txt",
)
RECOGNIZER_REQUIRED_RUN_FIELDS = (
"created_at",
"model_name",
"checkpoint",
"config",
"dictionary",
"artifact_identity",
"label_file",
"label_file_sha256",
"prediction_file",
"prediction_file_sha256",
"data_dir",
"script_inventory_file",
"script_inventory_file_sha256",
"eval_pack_version",
"pack_content_sha256",
"frozen_eval_pack",
"unseen_for_selection",
"selected_checkpoint_on_test",
"normalization_profile",
"eval_mode",
"languages",
"scripts",
"domains",
"gt_statuses",
"splits",
"script_inventory",
"claim_safe",
"warnings",
"machine",
"command",
)
PAGE_REQUIRED_RUN_FIELDS = (
"created_at",
"system_name",
"model_name",
"checkpoint",
"config",
"dictionary",
"artifact_identity",
"eval_manifest",
"eval_manifest_sha256",
"train_manifests",
"documents_dir",
"ocr_output_kind",
"require_output_audit",
"output_audit_filename",
"output_audit_passed_field",
"page_output_audit",
"script_inventory_file",
"script_inventory_file_sha256",
"eval_pack_version",
"pack_content_sha256",
"frozen_eval_pack",
"unseen_for_selection",
"selected_checkpoint_on_test",
"line_iou_threshold",
"language_default",
"script_default",
"domain_default",
"gt_status_default",
"normalization_profile",
"languages",
"scripts",
"domains",
"gt_statuses",
"splits",
"eval_mode",
"script_inventory",
"claim_safe",
"warnings",
"machine",
"command",
)
RECIPE_REQUIRED_FIELDS = (
"created_at",
"recipe_id",
"recipe_file",
"recipe_file_sha256",
"root",
"outputs",
"train_manifests",
"recipe_validation",
"materialization",
"audit",
"score",
"machine",
"command",
)
CLAIM_SAFE_DOMAINS = {"real_print", "real_handwritten"}
CLAIM_SAFE_GT_STATUSES = {"human_verified", "machine_converted_spot_checked"}
CLAIM_EVAL_SPLITS = {"eval", "test", "heldout", "held_out", "validation"}
SOURCE_DOCUMENT_KEYS = ("source_document", "source_doc", "source_pdf", "source_image", "source_xml", "document", "doc_id")
PAGE_KEYS = ("page", "page_index", "page_number", "source_page", "page_id")
BBOX_KEYS = ("bbox", "bounding_box", "coords", "polygon")
CLAIM_ELIGIBLE_KEYS = ("claim_eligible", "claim_evidence_eligible", "paper_claim_eligible")
HUMAN_REVIEW_EVIDENCE_KEYS = (
"reviewed_by",
"reviewer",
"native_reviewer",
"gt_reviewer",
"review_artifact",
"gt_review_artifact",
"review_batch",
"double_reviewed_by",
"inter_annotator_agreement",
"correction_rate",
)
PROVENANCE_SIGNAL_KEYS = (
"label_provenance",
"gt_source",
"source_gt_status",
"source_label_status",
"reference_status",
"assisted_source",
"source_type",
"data_origin",
"evidence_type",
"dataset_type",
"slices",
)
def _read_json(path: Path) -> Any:
if not path.exists():
raise FileNotFoundError(f"JSON artifact does not exist: {path}")
try:
return json.loads(path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
raise ValueError(f"invalid JSON artifact: {path}: {exc}") from exc
def _load_tool(name: str) -> Any:
path = TOOLS_DIR / f"{name}.py"
spec = importlib.util.spec_from_file_location(name, path)
if spec is None or spec.loader is None:
raise RuntimeError(f"could not load tool module: {path}")
module = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = module
spec.loader.exec_module(module)
return module
def _score_dir(run_dir: Path) -> Path:
if (run_dir / "eval-run.json").exists() or (run_dir / "page-eval-run.json").exists():
return run_dir
if (run_dir / "score").is_dir():
return run_dir / "score"
raise FileNotFoundError(f"could not find eval-run.json, page-eval-run.json, or score/ in {run_dir}")
def _mode(score_dir: Path) -> str:
if (score_dir / "eval-run.json").exists():
return "recognizer_oracle_crop"
if (score_dir / "page-eval-run.json").exists():
return "page_level_ocr"
raise FileNotFoundError(f"score directory has no eval run JSON: {score_dir}")
def _is_number(value: Any) -> bool:
return isinstance(value, int | float) and not isinstance(value, bool)
def _close(left: float, right: float, epsilon: float = 1e-12) -> bool:
return abs(left - right) <= epsilon
def _percentile(values: list[float], q: float) -> float:
if not values:
return 0.0
ordered = sorted(values)
if len(ordered) == 1:
return ordered[0]
position = (len(ordered) - 1) * q
lower = int(position)
upper = min(lower + 1, len(ordered) - 1)
weight = position - lower
return ordered[lower] * (1.0 - weight) + ordered[upper] * weight
def _confidence_interval(values: list[float]) -> dict[str, float]:
return {
"low": _percentile(values, 0.025),
"high": _percentile(values, 0.975),
}
def _expected_uncertainty(
units: list[dict[str, int]],
*,
method: str,
metrics: tuple[str, ...],
) -> dict[str, Any]:
n = len(units)
samples: dict[str, list[float]] = {metric: [] for metric in metrics}
if n:
rng = random.Random(UNCERTAINTY_BOOTSTRAP_SEED)
for _ in range(UNCERTAINTY_BOOTSTRAP_SAMPLES):
totals = {
"codepoint_edits": 0,
"codepoint_ref_len": 0,
"grapheme_edits": 0,
"grapheme_ref_len": 0,
"exact": 0,
}
for _ in range(n):
unit = units[rng.randrange(n)]
for key in totals:
totals[key] += unit.get(key, 0)
if "cer_codepoint_micro" in samples:
samples["cer_codepoint_micro"].append(totals["codepoint_edits"] / max(totals["codepoint_ref_len"], 1))
if "cer_grapheme_micro" in samples:
samples["cer_grapheme_micro"].append(totals["grapheme_edits"] / max(totals["grapheme_ref_len"], 1))
if "exact_line_accuracy" in samples:
samples["exact_line_accuracy"].append(totals["exact"] / max(n, 1))
if "text_cer_codepoint_micro" in samples:
samples["text_cer_codepoint_micro"].append(totals["codepoint_edits"] / max(totals["codepoint_ref_len"], 1))
if "text_cer_grapheme_micro" in samples:
samples["text_cer_grapheme_micro"].append(totals["grapheme_edits"] / max(totals["grapheme_ref_len"], 1))
if "text_exact_accuracy" in samples:
samples["text_exact_accuracy"].append(totals["exact"] / max(totals.get("text_reference", n), 1))
return {
"method": method,
"confidence_level": 0.95,
"bootstrap_samples": UNCERTAINTY_BOOTSTRAP_SAMPLES,
"seed": UNCERTAINTY_BOOTSTRAP_SEED,
"scoring_units": n,
"effective_sample_size": n,
"min_claim_scoring_units": MIN_CLAIM_SCORING_UNITS,
"claim_sample_size_ok": n >= MIN_CLAIM_SCORING_UNITS,
"metrics": {metric: _confidence_interval(samples[metric]) for metric in metrics},
}
def _check_uncertainty(recorded: Any, expected: dict[str, Any], label: str) -> list[str]:
issues: list[str] = []
if not isinstance(recorded, dict):
return [f"{label} uncertainty is missing or not an object"]
for key in (
"method",
"confidence_level",
"bootstrap_samples",
"seed",
"scoring_units",
"effective_sample_size",
"min_claim_scoring_units",
"claim_sample_size_ok",
):
if recorded.get(key) != expected[key]:
issues.append(f"{label} uncertainty {key} mismatch: recorded={recorded.get(key)} recomputed={expected[key]}")
recorded_metrics = recorded.get("metrics")
expected_metrics = expected["metrics"]
if not isinstance(recorded_metrics, dict):
return issues + [f"{label} uncertainty metrics missing or not an object"]
missing = sorted(set(expected_metrics) - set(recorded_metrics))
extra = sorted(set(recorded_metrics) - set(expected_metrics))
if missing:
issues.append(f"{label} uncertainty metrics missing fields: {missing}")
if extra:
issues.append(f"{label} uncertainty metrics contain extra fields: {extra}")
for metric, expected_interval in expected_metrics.items():
recorded_interval = recorded_metrics.get(metric)
if not isinstance(recorded_interval, dict):
issues.append(f"{label} uncertainty {metric} interval missing or not an object")
continue
for bound in ("low", "high"):
value = recorded_interval.get(bound)
expected_value = expected_interval[bound]
if not _is_number(value) or not _close(float(value), expected_value):
issues.append(f"{label} uncertainty {metric}.{bound} mismatch: recorded={value} recomputed={expected_value}")
return issues
def _artifact_rows(path: Path) -> int:
if not path.exists():
return 0
with path.open("r", encoding="utf-8", newline="") as handle:
return max(sum(1 for _ in handle) - 1, 0)
def _csv_rows(path: Path) -> list[dict[str, str]]:
if not path.exists():
return []
with path.open("r", encoding="utf-8", newline="") as handle:
return list(csv.DictReader(handle))
def _tsv_rows(path: Path) -> list[dict[str, str]]:
if not path.exists():
return []
with path.open("r", encoding="utf-8", newline="") as handle:
return list(csv.DictReader(handle, delimiter="\t"))
def _short_rows(rows: list[Any], limit: int = 10) -> list[Any | str]:
visible: list[Any | str] = list(rows[:limit])
if len(rows) > limit:
visible.append(f"... {len(rows) - limit} more")
return visible
def _duplicate_values(rows: list[dict[str, str]], field: str) -> list[str]:
seen: set[str] = set()
duplicates: set[str] = set()
for row in rows:
value = row.get(field) or ""
if not value:
continue
if value in seen:
duplicates.add(value)
seen.add(value)
return sorted(duplicates)
def _duplicate_keys(rows: list[dict[str, str]], fields: tuple[str, ...]) -> list[tuple[str, ...]]:
seen: set[tuple[str, ...]] = set()
duplicates: set[tuple[str, ...]] = set()
for row in rows:
key = tuple(row.get(field) or "" for field in fields)
if not any(key):
continue
if key in seen:
duplicates.add(key)
seen.add(key)
return sorted(duplicates)
def _check_unique_field(label: str, rows: list[dict[str, str]], field: str) -> list[str]:
duplicates = _duplicate_values(rows, field)
if duplicates:
return [f"{label} duplicate {field} values: {duplicates[:10]}"]
return []
def _check_unique_key(label: str, rows: list[dict[str, str]], fields: tuple[str, ...]) -> list[str]:
duplicates = _duplicate_keys(rows, fields)
if duplicates:
return [f"{label} duplicate {','.join(fields)} values: {duplicates[:10]}"]
return []
def _sha256_file(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _check_hash_field(eval_run: dict[str, Any], path_field: str, hash_field: str, label: str) -> list[str]:
issues: list[str] = []
path_value = eval_run.get(path_field)
expected = eval_run.get(hash_field)
if path_value in (None, "") and expected in (None, ""):
return issues
if not isinstance(path_value, str) or not path_value:
issues.append(f"{label} hash is recorded but path is missing: {hash_field}")
return issues
if not isinstance(expected, str) or not expected:
issues.append(f"{label} path is recorded but hash is missing: {hash_field}")
return issues
path = Path(path_value)
if not path.exists():
issues.append(f"{label} path does not exist for hash verification: {path_value}")
return issues
actual = _sha256_file(path)
if actual != expected:
issues.append(f"{label} sha256 mismatch for {path_field}: recorded={expected} actual={actual}")
return issues
def _read_summary_for_audit(eval_run: dict[str, Any], label: str, purpose: str) -> tuple[dict[str, Any], list[str]]:
summary_file = eval_run.get("summary_file")
if summary_file in (None, ""):
return {}, []
if not isinstance(summary_file, str):
return {}, [f"{label} summary_file is not a string"]
path = Path(summary_file)
if not path.exists():
return {}, [f"{label} summary file does not exist for {purpose}: {summary_file}"]
try:
payload = _read_json(path)
except (OSError, ValueError) as exc:
return {}, [f"{label} summary file could not be parsed for {purpose}: {exc}"]
if not isinstance(payload, dict):
return {}, [f"{label} summary file is not a JSON object for {purpose}: {summary_file}"]
return payload, []
def _check_summary_metadata_consistency(eval_run: dict[str, Any], label: str) -> list[str]:
summary, issues = _read_summary_for_audit(eval_run, label, "metadata audit")
if issues or not summary:
return issues
list_fields = (
("language", "languages"),
("script", "scripts"),
("domain", "domains"),
("gt_status", "gt_statuses"),
("split", "splits"),
)
for summary_field, run_field in list_fields:
summary_value = summary.get(summary_field)
if summary_value in (None, ""):
continue
run_values = eval_run.get(run_field)
if not isinstance(run_values, list):
continue
run_value_strings = {str(value) for value in run_values}
if str(summary_value) not in run_value_strings:
issues.append(f"{label} summary {summary_field} mismatch: summary={summary_value!r} run={run_values!r}")
summary_version = summary.get("eval_pack_version")
run_version = eval_run.get("eval_pack_version")
if summary_version not in (None, "") and run_version not in (None, "") and str(summary_version) != str(run_version):
issues.append(f"{label} summary eval_pack_version mismatch: summary={summary_version!r} run={run_version!r}")
for field in ("frozen_eval_pack", "unseen_for_selection", "claim_eligible_default"):
if field not in summary:
continue
summary_value = summary.get(field)
run_value = eval_run.get(field)
if not isinstance(summary_value, bool):
issues.append(f"{label} summary {field} is not boolean: {summary_value!r}")
continue
if isinstance(run_value, bool) and summary_value != run_value:
issues.append(f"{label} summary {field} mismatch: summary={summary_value!r} run={run_value!r}")
summary_eval_pack = summary.get("eval_pack")
if isinstance(summary_eval_pack, dict):
summary_task = summary_eval_pack.get("task")
run_task = eval_run.get("eval_mode")
if summary_task not in (None, "") and run_task not in (None, "") and str(summary_task) != str(run_task):
issues.append(f"{label} summary task mismatch: summary={summary_task!r} run={run_task!r}")
return issues
def _check_train_manifest_hashes(eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
manifests = eval_run.get("train_manifests")
if manifests in (None, []):
return issues
if not isinstance(manifests, list):
return ["train_manifests is not a list"]
for index, manifest in enumerate(manifests):
if not isinstance(manifest, dict):
issues.append(f"train_manifests[{index}] is not an object")
continue
path_value = manifest.get("path")
expected = manifest.get("sha256")
if not isinstance(path_value, str) or not path_value:
issues.append(f"train_manifests[{index}] missing path")
continue
if not isinstance(expected, str) or not expected:
issues.append(f"train_manifests[{index}] missing sha256")
continue
path = Path(path_value)
if not path.exists():
issues.append(f"train_manifests[{index}] path does not exist: {path_value}")
continue
actual = _sha256_file(path)
if actual != expected:
issues.append(f"train_manifests[{index}] sha256 mismatch: recorded={expected} actual={actual}")
return issues
def _audit_hash_text(value: str) -> str:
return hashlib.sha256(_nfc(value).encode("utf-8")).hexdigest()
def _audit_sample_hash(image_sha256: str, text_sha256: str) -> str:
return hashlib.sha256(f"{image_sha256}\0{text_sha256}".encode("utf-8")).hexdigest()
def _train_manifest_paths(eval_run: dict[str, Any]) -> tuple[list[Path], list[str]]:
manifests = eval_run.get("train_manifests")
if manifests in (None, []):
return [], []
if not isinstance(manifests, list):
return [], ["train_manifests is not a list for leakage recomputation"]
paths: list[Path] = []
issues: list[str] = []
for index, manifest in enumerate(manifests):
if not isinstance(manifest, dict):
issues.append(f"train_manifests[{index}] is not an object for leakage recomputation")
continue
path_value = manifest.get("path")
if not isinstance(path_value, str) or not path_value:
issues.append(f"train_manifests[{index}] missing path for leakage recomputation")
continue
paths.append(Path(path_value))
return paths, issues
def _audit_manifest_leakage_rows(path: Path, *, manifest_dir: Path | None = None, compute_missing_image_sha: bool) -> tuple[list[dict[str, str]], list[str]]:
rows: list[dict[str, str]] = []
issues: list[str] = []
if not path.exists():
return rows, [f"leakage manifest does not exist: {path}"]
base_dir = manifest_dir or path.parent
with path.open("r", encoding="utf-8") as handle:
for line_no, line in enumerate(handle, 1):
line = line.strip()
if not line:
continue
try:
payload = json.loads(line)
except json.JSONDecodeError as exc:
issues.append(f"leakage manifest {path}:{line_no} invalid JSON: {exc}")
continue
if not isinstance(payload, dict):
issues.append(f"leakage manifest {path}:{line_no} row is not an object")
continue
metadata = dict(payload.get("metadata") or {})
sample_id = str(payload.get("sample_id") or payload.get("id") or metadata.get("sample_id") or "")
image_path = str(payload.get("image_path") or payload.get("image") or metadata.get("image_path") or "")
image_sha256 = str(payload.get("sha256") or payload.get("image_sha256") or metadata.get("sha256") or "")
if compute_missing_image_sha and not image_sha256 and image_path:
resolved = _resolve_audit_manifest_path(image_path, base_dir)
if resolved.exists():
image_sha256 = _sha256_file(resolved)
text_value = payload.get("text") if payload.get("text") is not None else metadata.get("text")
text = str(text_value or "")
text_sha256 = str(payload.get("text_sha256") or metadata.get("text_sha256") or (_audit_hash_text(text) if text else ""))
sample_sha256 = str(payload.get("sample_sha256") or metadata.get("sample_sha256") or "")
if not sample_sha256 and image_sha256 and text_sha256:
sample_sha256 = _audit_sample_hash(image_sha256, text_sha256)
source_sample_id = str(payload.get("source_sample_id") or metadata.get("source_sample_id") or sample_id)
rows.append(
{
"sample_id": sample_id,
"source_sample_id": source_sample_id,
"image_path": image_path,
"image_sha256": image_sha256,
"text_sha256": text_sha256,
"sample_sha256": sample_sha256,
}
)
return rows, issues
def _leakage_overlap(left: list[dict[str, str]], right: list[dict[str, str]], field: str) -> list[str]:
left_values = {row.get(field) or "" for row in left if row.get(field)}
right_values = {row.get(field) or "" for row in right if row.get(field)}
return sorted(left_values & right_values)
def _leakage_overlaps(train_rows: list[dict[str, str]], eval_rows: list[dict[str, str]]) -> dict[str, list[str]]:
return {
"sample_id": _leakage_overlap(train_rows, eval_rows, "sample_id"),
"source_sample_id": _leakage_overlap(train_rows, eval_rows, "source_sample_id"),
"image_sha256": _leakage_overlap(train_rows, eval_rows, "image_sha256"),
"sample_sha256": _leakage_overlap(train_rows, eval_rows, "sample_sha256"),
"image_path": _leakage_overlap(train_rows, eval_rows, "image_path"),
"text_sha256": _leakage_overlap(train_rows, eval_rows, "text_sha256"),
}
def _compare_leakage_report(label: str, recorded: Any, expected: dict[str, Any]) -> list[str]:
if not isinstance(recorded, dict):
return [f"{label} leakage report is missing or not an object"]
issues: list[str] = []
for key in ("train_manifests", "checked", "train_rows", "eval_rows"):
if recorded.get(key) != expected[key]:
issues.append(f"{label} leakage {key} mismatch: recorded={recorded.get(key)} recomputed={expected[key]}")
recorded_overlaps = recorded.get("overlaps")
if not isinstance(recorded_overlaps, dict):
issues.append(f"{label} leakage overlaps missing or not an object")
return issues
for key, expected_values in expected["overlaps"].items():
recorded_values = recorded_overlaps.get(key)
if recorded_values != expected_values:
issues.append(f"{label} leakage overlaps.{key} mismatch: recorded={recorded_values} recomputed={expected_values}")
return issues
def _spotcheck_calibration_summary(path: Path) -> tuple[dict[str, Any] | None, list[str]]:
if not path.exists():
return None, [f"spotcheck calibration file does not exist for summary audit: {path}"]
try:
payload = json.loads(path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return None, [f"spotcheck calibration file is invalid JSON for summary audit: {path}: {exc}"]
if not isinstance(payload, dict):
return None, [f"spotcheck calibration file must contain an object for summary audit: {path}"]
return (
{
"claim_evidence_eligible": payload.get("claim_evidence_eligible"),
"recommended_pack_gt_status": payload.get("recommended_pack_gt_status"),
"language": payload.get("language"),
"script": payload.get("script"),
"domain": payload.get("domain"),
"source_gt_status": payload.get("source_gt_status"),
"counts": payload.get("counts", {}),
"machine_label_metrics": payload.get("machine_label_metrics", {}),
"eligibility_reasons": payload.get("eligibility_reasons", []),
},
[],
)
def _check_spotcheck_calibration_summary(eval_run: dict[str, Any], metrics: dict[str, Any], label: str) -> list[str]:
path_value = eval_run.get("spotcheck_calibration_file")
if path_value in (None, ""):
if eval_run.get("spotcheck_calibration") not in (None, {}):
return [f"{label} spotcheck_calibration is recorded without spotcheck_calibration_file"]
if metrics.get("spotcheck_calibration") not in (None, {}):
return [f"{label} metrics spotcheck_calibration is recorded without spotcheck_calibration_file"]
return []
if not isinstance(path_value, str):
return [f"{label} spotcheck_calibration_file is not a string"]
expected, issues = _spotcheck_calibration_summary(Path(path_value))
if issues:
return issues
assert expected is not None
recorded_run = eval_run.get("spotcheck_calibration")
recorded_metrics = metrics.get("spotcheck_calibration")
if recorded_run != expected:
issues.append(f"{label} eval-run spotcheck_calibration does not match calibration file: recorded={recorded_run} recomputed={expected}")
if recorded_metrics != expected:
issues.append(f"{label} metrics spotcheck_calibration does not match calibration file: recorded={recorded_metrics} recomputed={expected}")
return issues
def _check_artifact_identity(eval_run: dict[str, Any], required: tuple[str, ...], label: str) -> list[str]:
issues: list[str] = []
artifacts = eval_run.get("artifact_identity")
if not isinstance(artifacts, dict):
return [f"{label} artifact_identity is missing or not an object"]
for name in required:
identity = artifacts.get(name)
if not isinstance(identity, dict):
issues.append(f"{label} artifact_identity.{name} is missing or not an object")
continue
if not identity.get("verified"):
issues.append(f"{label} artifact_identity.{name} is not hash-verifiable: {identity.get('reason') or 'unverified'}")
continue
recorded_value = eval_run.get(name)
identity_value = identity.get("value")
if not isinstance(recorded_value, str) or not recorded_value:
issues.append(f"{label} {name} field is missing for artifact_identity consistency")
elif identity_value != recorded_value:
issues.append(f"{label} artifact_identity.{name}.value does not match eval-run {name}: recorded={identity_value} eval_run={recorded_value}")
path_value = identity.get("path") or identity.get("value")
expected = identity.get("sha256")
if not isinstance(path_value, str) or not path_value:
issues.append(f"{label} artifact_identity.{name} missing path")
continue
if not isinstance(expected, str) or not expected:
issues.append(f"{label} artifact_identity.{name} missing sha256")
continue
path = Path(path_value)
if not path.exists():
issues.append(f"{label} artifact_identity.{name} path does not exist: {path_value}")
continue
actual = _sha256_file(path)
if actual != expected:
issues.append(f"{label} artifact_identity.{name} sha256 mismatch: recorded={expected} actual={actual}")
return issues
def _check_required_artifacts(score_dir: Path, names: tuple[str, ...]) -> list[str]:
issues: list[str] = []
for name in names:
path = score_dir / name
if not path.exists():
issues.append(f"missing required artifact: {name}")
elif path.is_file() and path.stat().st_size == 0:
issues.append(f"empty required artifact: {name}")
return issues
def _fmt_report_metric(value: Any) -> str:
if _is_number(value):
return f"{float(value):.6f}"
return str(value)
def _expect_report_line(text: str, issues: list[str], line: str) -> None:
if line not in text.splitlines():
issues.append(f"claim-safety.md missing or stale line: {line}")
def _check_claim_safety_report_consistency(score_dir: Path, eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
issues: list[str] = []
path = score_dir / "claim-safety.md"
if not path.exists():
return issues
text = path.read_text(encoding="utf-8")
machine = eval_run.get("machine") if isinstance(eval_run.get("machine"), dict) else {}
expected_lines = [
f"- eval mode: {eval_run.get('eval_mode')}",
f"- claim safe: {eval_run.get('claim_safe')}",
f"- created at: {eval_run.get('created_at')}",
f"- machine platform: {machine.get('platform')}",
f"- machine python: {machine.get('python')}",
f"- command: {eval_run.get('command')}",
f"- eval pack version: {eval_run.get('eval_pack_version') or '(missing)'}",
f"- frozen eval pack: {eval_run.get('frozen_eval_pack')}",
f"- unseen for selection: {eval_run.get('unseen_for_selection')}",
f"- selected checkpoint on test: {eval_run.get('selected_checkpoint_on_test')}",
f"- languages: {eval_run.get('languages')}",
f"- scripts: {eval_run.get('scripts')}",
f"- domains: {eval_run.get('domains')}",
f"- GT statuses: {eval_run.get('gt_statuses')}",
f"- splits: {eval_run.get('splits')}",
f"- metadata contradictions: {metrics.get('metadata_contradictions', [])}",
]
mode = eval_run.get("eval_mode")
if mode == "recognizer_oracle_crop":
uncertainty = metrics.get("uncertainty") if isinstance(metrics.get("uncertainty"), dict) else {}
intervals = uncertainty.get("metrics") if isinstance(uncertainty.get("metrics"), dict) else {}
graph_interval = intervals.get("cer_grapheme_micro") if isinstance(intervals.get("cer_grapheme_micro"), dict) else {}
inventory = metrics.get("script_inventory") if isinstance(metrics.get("script_inventory"), dict) else {}
normalization = metrics.get("normalization_summary") if isinstance(metrics.get("normalization_summary"), dict) else {}
dictionary_coverage = metrics.get("dictionary_coverage") if isinstance(metrics.get("dictionary_coverage"), dict) else {}
expected_lines.extend(
[
f"- model: {eval_run.get('model_name')}",
f"- checkpoint: {eval_run.get('checkpoint') or '(not recorded)'}",
f"- config: {eval_run.get('config') or '(not recorded)'}",
f"- dictionary: {eval_run.get('dictionary') or '(not recorded)'}",
f"- artifact identity: {json.dumps(eval_run.get('artifact_identity', {}), ensure_ascii=False, sort_keys=True)}",
f"- label file: {eval_run.get('label_file')}",
f"- prediction file: {eval_run.get('prediction_file')}",
f"- scored lines: {metrics.get('n_lines')}",
f"- unmatched labels: {metrics.get('unmatched_labels')}",
f"- unmatched predictions: {metrics.get('unmatched_predictions')}",
f"- script inventory file: {inventory.get('inventory_file')}",
f"- observed codepoints/graphemes: {inventory.get('observed_codepoint_count')}/{inventory.get('observed_grapheme_count')}",
f"- missing required codepoints: {inventory.get('missing_required_codepoints')}",
f"- missing required graphemes: {inventory.get('missing_required_graphemes')}",
f"- normalization profile: {normalization.get('profile')}",
f"- normalization changed GT/prediction/any: {normalization.get('gt_changed')}/{normalization.get('prediction_changed')}/{normalization.get('any_changed')} of {normalization.get('total')}",
f"- dictionary coverage checked: {dictionary_coverage.get('checked')}",
f"- dictionary GT codepoints/dictionary codepoints: {dictionary_coverage.get('gt_codepoint_count')}/{dictionary_coverage.get('dictionary_codepoint_count')}",
f"- dictionary missing GT codepoints: {dictionary_coverage.get('missing_gt_codepoints')}",
f"- codepoint CER micro: {_fmt_report_metric(metrics.get('cer_codepoint_micro'))}",
f"- grapheme CER micro: {_fmt_report_metric(metrics.get('cer_grapheme_micro'))}",
f"- exact-line accuracy: {_fmt_report_metric(metrics.get('exact_line_accuracy'))}",
f"- prediction match rate: {_fmt_report_metric(metrics.get('prediction_match_rate'))}",
f"- codepoint edits/ref: {metrics.get('codepoint_edits')}/{metrics.get('codepoint_ref_len')}",
f"- grapheme edits/ref: {metrics.get('grapheme_edits')}/{metrics.get('grapheme_ref_len')}",
f"- zero-length GT lines: {metrics.get('zero_length_gt_lines')}",
f"- uncertainty method: {uncertainty.get('method')}",
f"- uncertainty scoring units: {uncertainty.get('effective_sample_size')}/{uncertainty.get('min_claim_scoring_units')}",
f"- grapheme CER 95% CI: {_fmt_report_metric(graph_interval.get('low'))}..{_fmt_report_metric(graph_interval.get('high'))}",
"- command.txt",
"- predictions.tsv",
"- unmatched.tsv",
"- worst-lines.csv",
"- sample-review.csv",
"- sample-review.html",
"- confusions.json",
]
)
slices = metrics.get("slices") if isinstance(metrics.get("slices"), list) else []
for row in slices:
if not isinstance(row, dict):
continue
expected_lines.append(
f"- {row.get('language')} / {row.get('script')} / {row.get('domain')} / {row.get('gt_status')} / {row.get('split')}: "
f"n={row.get('n_lines')} graph_CER={_fmt_report_metric(row.get('cer_grapheme_micro'))} "
f"cp_CER={_fmt_report_metric(row.get('cer_codepoint_micro'))} exact={_fmt_report_metric(row.get('exact_line_accuracy'))}"
)
elif mode == "page_level_ocr":
metric_means = metrics.get("metric_means") if isinstance(metrics.get("metric_means"), dict) else {}
uncertainty = metrics.get("uncertainty") if isinstance(metrics.get("uncertainty"), dict) else {}
intervals = uncertainty.get("metrics") if isinstance(uncertainty.get("metrics"), dict) else {}
graph_interval = intervals.get("text_cer_grapheme_micro") if isinstance(intervals.get("text_cer_grapheme_micro"), dict) else {}
inventory = metrics.get("script_inventory") if isinstance(metrics.get("script_inventory"), dict) else {}
normalization = metrics.get("normalization_summary") if isinstance(metrics.get("normalization_summary"), dict) else {}
dictionary_coverage = metrics.get("dictionary_coverage") if isinstance(metrics.get("dictionary_coverage"), dict) else {}
expected_lines.extend(
[
f"- system: {eval_run.get('system_name')}",
f"- model: {eval_run.get('model_name')}",
f"- checkpoint: {eval_run.get('checkpoint') or '(not recorded)'}",
f"- config: {eval_run.get('config') or '(not recorded)'}",
f"- dictionary: {eval_run.get('dictionary') or '(not recorded)'}",
f"- artifact identity: {json.dumps(eval_run.get('artifact_identity', {}), ensure_ascii=False, sort_keys=True)}",
f"- normalization profile: {eval_run.get('normalization_profile')}",
f"- eval manifest: {eval_run.get('eval_manifest')}",
f"- documents dir: {eval_run.get('documents_dir')}",
f"- manifest samples: {metrics.get('manifest_samples')}",
f"- scored pages: {metrics.get('scored_pages')}",
f"- missing documents: {metrics.get('missing_documents')}",
f"- missing references: {metrics.get('missing_references')}",
f"- script inventory file: {inventory.get('inventory_file')}",
f"- observed codepoints/graphemes: {inventory.get('observed_codepoint_count')}/{inventory.get('observed_grapheme_count')}",
f"- missing required codepoints: {inventory.get('missing_required_codepoints')}",
f"- missing required graphemes: {inventory.get('missing_required_graphemes')}",
f"- normalization profile: {normalization.get('profile')}",
f"- normalization changed reference/prediction/any: {normalization.get('reference_changed')}/{normalization.get('prediction_changed')}/{normalization.get('any_changed')} of {normalization.get('total')}",
f"- dictionary coverage checked: {dictionary_coverage.get('checked')}",
f"- dictionary reference codepoints/dictionary codepoints: {dictionary_coverage.get('gt_codepoint_count')}/{dictionary_coverage.get('dictionary_codepoint_count')}",
f"- dictionary missing reference codepoints: {dictionary_coverage.get('missing_gt_codepoints')}",
f"- leakage checked: {metrics.get('leakage', {}).get('checked') if isinstance(metrics.get('leakage'), dict) else None}",
f"- text codepoint CER micro: {_fmt_report_metric(metrics.get('text_cer_codepoint_micro'))}",
f"- text grapheme CER micro: {_fmt_report_metric(metrics.get('text_cer_grapheme_micro'))}",
f"- text exact accuracy: {_fmt_report_metric(metrics.get('text_exact_accuracy'))}",
f"- line detection F1 mean: {_fmt_report_metric(metric_means.get('line_detection_f1'))}",
f"- line mean IoU: {_fmt_report_metric(metric_means.get('line_mean_iou'))}",
f"- reading-order pair accuracy: {_fmt_report_metric(metric_means.get('reading_order_pair_accuracy'))}",
f"- text codepoint edits/ref: {metrics.get('text_codepoint_edits')}/{metrics.get('text_codepoint_ref_len')}",
f"- text grapheme edits/ref: {metrics.get('text_grapheme_edits')}/{metrics.get('text_grapheme_ref_len')}",
f"- text exact pages/ref pages: {metrics.get('text_exact_pages')}/{metrics.get('text_reference_pages')}",
f"- zero-length text references: {metrics.get('zero_length_text_references')}",
f"- uncertainty method: {uncertainty.get('method')}",
f"- uncertainty scoring units: {uncertainty.get('effective_sample_size')}/{uncertainty.get('min_claim_scoring_units')}",
f"- text grapheme CER 95% CI: {_fmt_report_metric(graph_interval.get('low'))}..{_fmt_report_metric(graph_interval.get('high'))}",
"- command.txt",
"- page-metrics.csv",
"- reference-artifacts.tsv",
"- text-pairs.tsv",
"- missing.tsv",
"- worst-pages.csv",
"- sample-review.csv",
"- sample-review.html",
"- confusions.json",
"- line-alignment.csv",
]
)
slices = metrics.get("slices") if isinstance(metrics.get("slices"), list) else []
for row in slices:
if not isinstance(row, dict):
continue
row_means = row.get("metric_means") if isinstance(row.get("metric_means"), dict) else {}
expected_lines.append(
f"- {row.get('language')} / {row.get('script')} / {row.get('domain')} / {row.get('gt_status')} / {row.get('split')}: "
f"pages={row.get('scored_pages')} text_graph_CER={_fmt_report_metric(row.get('text_cer_grapheme_micro'))} "
f"text_cp_CER={_fmt_report_metric(row.get('text_cer_codepoint_micro'))} "
f"exact={_fmt_report_metric(row.get('text_exact_accuracy'))} line_F1={_fmt_report_metric(row_means.get('line_detection_f1', 0.0))}"
)
for line in expected_lines:
_expect_report_line(text, issues, line)
warnings = eval_run.get("warnings") if isinstance(eval_run.get("warnings"), list) else []
if warnings:
for warning in warnings:
_expect_report_line(text, issues, f"- {warning}")
else:
_expect_report_line(text, issues, "- none")
return issues
def _expect_metrics_line(text: str, issues: list[str], line: str) -> None:
if line not in text.splitlines():
issues.append(f"metrics.md missing or stale line: {line}")
def _check_metrics_md_consistency(score_dir: Path, eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
issues: list[str] = []
path = score_dir / "metrics.md"
if not path.exists():
return issues
text = path.read_text(encoding="utf-8")
expected_lines: list[str] = []
mode = eval_run.get("eval_mode")
if mode == "recognizer_oracle_crop":
uncertainty = metrics.get("uncertainty") if isinstance(metrics.get("uncertainty"), dict) else {}
intervals = uncertainty.get("metrics") if isinstance(uncertainty.get("metrics"), dict) else {}
graph_interval = intervals.get("cer_grapheme_micro") if isinstance(intervals.get("cer_grapheme_micro"), dict) else {}
inventory = metrics.get("script_inventory") if isinstance(metrics.get("script_inventory"), dict) else {}
normalization = metrics.get("normalization_summary") if isinstance(metrics.get("normalization_summary"), dict) else {}
dictionary_coverage = metrics.get("dictionary_coverage") if isinstance(metrics.get("dictionary_coverage"), dict) else {}
expected_lines.extend(
[
"# Recognizer Eval Metrics",
f"- lines: {metrics.get('n_lines')}",
f"- label lines: {metrics.get('label_lines')}",
f"- prediction match rate: {_fmt_report_metric(metrics.get('prediction_match_rate'))}",
f"- unmatched labels: {metrics.get('unmatched_labels')}",
f"- unmatched predictions: {metrics.get('unmatched_predictions')}",
f"- codepoint CER: {_fmt_report_metric(metrics.get('cer_codepoint_micro'))}",
f"- grapheme CER: {_fmt_report_metric(metrics.get('cer_grapheme_micro'))}",
f"- exact-line accuracy: {_fmt_report_metric(metrics.get('exact_line_accuracy'))}",
f"- codepoint edits/ref: {metrics.get('codepoint_edits')}/{metrics.get('codepoint_ref_len')}",
f"- grapheme edits/ref: {metrics.get('grapheme_edits')}/{metrics.get('grapheme_ref_len')}",
f"- zero-length GT lines: {metrics.get('zero_length_gt_lines')}",
f"- normalization profile: {normalization.get('profile')}",
f"- normalization changed GT/prediction/any: {normalization.get('gt_changed')}/{normalization.get('prediction_changed')}/{normalization.get('any_changed')} of {normalization.get('total')}",
f"- dictionary coverage checked: {dictionary_coverage.get('checked')}",
f"- dictionary GT codepoints/dictionary codepoints: {dictionary_coverage.get('gt_codepoint_count')}/{dictionary_coverage.get('dictionary_codepoint_count')}",
f"- dictionary missing GT codepoints: {dictionary_coverage.get('missing_gt_codepoints')}",
f"- uncertainty method: {uncertainty.get('method')}",
f"- uncertainty scoring units: {uncertainty.get('effective_sample_size')}/{uncertainty.get('min_claim_scoring_units')}",
f"- grapheme CER 95% CI: {_fmt_report_metric(graph_interval.get('low'))}..{_fmt_report_metric(graph_interval.get('high'))}",
f"- script inventory file: {inventory.get('inventory_file')}",
f"- observed codepoints/graphemes: {inventory.get('observed_codepoint_count')}/{inventory.get('observed_grapheme_count')}",
f"- missing required codepoints: {inventory.get('missing_required_codepoints')}",
f"- missing required graphemes: {inventory.get('missing_required_graphemes')}",
]
)
slices = metrics.get("slices") if isinstance(metrics.get("slices"), list) else []
for row in slices:
if not isinstance(row, dict):
continue
expected_lines.append(
f"- {row.get('language')} / {row.get('script')} / {row.get('domain')} / {row.get('gt_status')} / {row.get('split')}: "
f"n={row.get('n_lines')} graph_CER={_fmt_report_metric(row.get('cer_grapheme_micro'))} "
f"cp_CER={_fmt_report_metric(row.get('cer_codepoint_micro'))} exact={_fmt_report_metric(row.get('exact_line_accuracy'))}"
)
elif mode == "page_level_ocr":
metric_means = metrics.get("metric_means") if isinstance(metrics.get("metric_means"), dict) else {}
uncertainty = metrics.get("uncertainty") if isinstance(metrics.get("uncertainty"), dict) else {}
intervals = uncertainty.get("metrics") if isinstance(uncertainty.get("metrics"), dict) else {}
graph_interval = intervals.get("text_cer_grapheme_micro") if isinstance(intervals.get("text_cer_grapheme_micro"), dict) else {}
inventory = metrics.get("script_inventory") if isinstance(metrics.get("script_inventory"), dict) else {}
normalization = metrics.get("normalization_summary") if isinstance(metrics.get("normalization_summary"), dict) else {}
dictionary_coverage = metrics.get("dictionary_coverage") if isinstance(metrics.get("dictionary_coverage"), dict) else {}
expected_lines.extend(
[
"# Page OCR Eval Metrics",
f"- manifest samples: {metrics.get('manifest_samples')}",
f"- scored pages: {metrics.get('scored_pages')}",
f"- missing documents: {metrics.get('missing_documents')}",
f"- missing references: {metrics.get('missing_references')}",
f"- text codepoint CER micro: {_fmt_report_metric(metrics.get('text_cer_codepoint_micro'))}",
f"- text grapheme CER micro: {_fmt_report_metric(metrics.get('text_cer_grapheme_micro'))}",
f"- text exact accuracy: {_fmt_report_metric(metrics.get('text_exact_accuracy'))}",
f"- text codepoint edits/ref: {metrics.get('text_codepoint_edits')}/{metrics.get('text_codepoint_ref_len')}",
f"- text grapheme edits/ref: {metrics.get('text_grapheme_edits')}/{metrics.get('text_grapheme_ref_len')}",
f"- text exact pages/ref pages: {metrics.get('text_exact_pages')}/{metrics.get('text_reference_pages')}",
f"- zero-length text references: {metrics.get('zero_length_text_references')}",
f"- normalization profile: {normalization.get('profile')}",
f"- normalization changed reference/prediction/any: {normalization.get('reference_changed')}/{normalization.get('prediction_changed')}/{normalization.get('any_changed')} of {normalization.get('total')}",
f"- dictionary coverage checked: {dictionary_coverage.get('checked')}",
f"- dictionary reference codepoints/dictionary codepoints: {dictionary_coverage.get('gt_codepoint_count')}/{dictionary_coverage.get('dictionary_codepoint_count')}",
f"- dictionary missing reference codepoints: {dictionary_coverage.get('missing_gt_codepoints')}",
f"- line detection F1 mean: {_fmt_report_metric(metric_means.get('line_detection_f1', 0.0))}",
f"- line mean IoU: {_fmt_report_metric(metric_means.get('line_mean_iou', 0.0))}",
f"- uncertainty method: {uncertainty.get('method')}",
f"- uncertainty scoring units: {uncertainty.get('effective_sample_size')}/{uncertainty.get('min_claim_scoring_units')}",
f"- text grapheme CER 95% CI: {_fmt_report_metric(graph_interval.get('low'))}..{_fmt_report_metric(graph_interval.get('high'))}",
f"- script inventory file: {inventory.get('inventory_file')}",
f"- observed codepoints/graphemes: {inventory.get('observed_codepoint_count')}/{inventory.get('observed_grapheme_count')}",
f"- missing required codepoints: {inventory.get('missing_required_codepoints')}",
f"- missing required graphemes: {inventory.get('missing_required_graphemes')}",
]
)
for key, value in metric_means.items():
expected_lines.append(f"- {key}: {_fmt_report_metric(value)}")
slices = metrics.get("slices") if isinstance(metrics.get("slices"), list) else []
for row in slices:
if not isinstance(row, dict):
continue
row_means = row.get("metric_means") if isinstance(row.get("metric_means"), dict) else {}
expected_lines.append(
f"- {row.get('language')} / {row.get('script')} / {row.get('domain')} / {row.get('gt_status')} / {row.get('split')}: "
f"pages={row.get('scored_pages')} text_graph_CER={_fmt_report_metric(row.get('text_cer_grapheme_micro'))} "
f"text_cp_CER={_fmt_report_metric(row.get('text_cer_codepoint_micro'))} "
f"line_F1={_fmt_report_metric(row_means.get('line_detection_f1', 0.0))}"
)
coverage = metrics.get("reference_coverage") if isinstance(metrics.get("reference_coverage"), dict) else {}
for key, value in coverage.items():
expected_lines.append(f"- {key}: {value}")
for line in expected_lines:
_expect_metrics_line(text, issues, line)
warnings = eval_run.get("warnings") if isinstance(eval_run.get("warnings"), list) else []
if warnings:
for warning in warnings:
_expect_metrics_line(text, issues, f"- {warning}")
else:
_expect_metrics_line(text, issues, "- none")
return issues
def _expect_root_report_line(text: str, issues: list[str], line: str) -> None:
if line not in text.splitlines():
issues.append(f"root claim-safety.md missing or stale line: {line}")
def _check_root_claim_safety_report_consistency(
run_dir: Path,
recipe_run: dict[str, Any],
eval_run: dict[str, Any],
metrics: dict[str, Any],
) -> list[str]:
issues: list[str] = []
path = run_dir / "claim-safety.md"
if not path.exists():
return issues
text = path.read_text(encoding="utf-8")
recipe: dict[str, Any] = {}
if (run_dir / "recipe.json").exists():
try:
raw_recipe = _read_json(run_dir / "recipe.json")
except (FileNotFoundError, ValueError) as exc:
issues.append(str(exc))
raw_recipe = {}
if isinstance(raw_recipe, dict):
recipe = raw_recipe
else:
issues.append("recipe.json must contain a JSON object")
recipe_audit = recipe_run.get("audit") if isinstance(recipe_run.get("audit"), dict) else {}
machine = eval_run.get("machine") if isinstance(eval_run.get("machine"), dict) else {}
leakage = metrics.get("leakage") if isinstance(metrics.get("leakage"), dict) else {}
uncertainty = metrics.get("uncertainty") if isinstance(metrics.get("uncertainty"), dict) else {}
intervals = uncertainty.get("metrics") if isinstance(uncertainty.get("metrics"), dict) else {}
script_inventory = metrics.get("script_inventory") if isinstance(metrics.get("script_inventory"), dict) else {}
expected_lines = [
f"- recipe: {recipe_run.get('recipe_id')}",
f"- task: {recipe.get('task')}",
f"- audit passed: {recipe_audit.get('audit_passed')}",
f"- claim safe: {eval_run.get('claim_safe')}",
f"- eval mode: {eval_run.get('eval_mode')}",
f"- score created at: {eval_run.get('created_at')}",
f"- score machine platform: {machine.get('platform')}",
f"- score machine python: {machine.get('python')}",
f"- score command: {eval_run.get('command')}",
]
mode = eval_run.get("eval_mode")
if mode == "page_level_ocr":
expected_lines.extend(
[
f"- system: {eval_run.get('system_name')}",
f"- model: {eval_run.get('model_name')}",
f"- checkpoint: {eval_run.get('checkpoint') or '(not recorded)'}",
f"- config: {eval_run.get('config') or '(not recorded)'}",
f"- dictionary: {eval_run.get('dictionary') or '(not recorded)'}",
f"- artifact identity: {json.dumps(eval_run.get('artifact_identity', {}), ensure_ascii=False, sort_keys=True)}",
f"- normalization profile: {eval_run.get('normalization_profile')}",
f"- eval manifest: {eval_run.get('eval_manifest')}",
f"- documents dir: {eval_run.get('documents_dir')}",
f"- manifest samples: {metrics.get('manifest_samples')}",
f"- scored pages: {metrics.get('scored_pages')}",
f"- missing documents: {metrics.get('missing_documents')}",
f"- missing references: {metrics.get('missing_references')}",
]
)
else:
expected_lines.extend(
[
f"- model: {eval_run.get('model_name')}",
f"- checkpoint: {eval_run.get('checkpoint') or '(not recorded)'}",
f"- config: {eval_run.get('config') or '(not recorded)'}",
f"- dictionary: {eval_run.get('dictionary') or '(not recorded)'}",
f"- artifact identity: {json.dumps(eval_run.get('artifact_identity', {}), ensure_ascii=False, sort_keys=True)}",
f"- normalization profile: {eval_run.get('normalization_profile')}",
f"- label file: {eval_run.get('label_file')}",
f"- prediction file: {eval_run.get('prediction_file')}",
f"- scored lines: {metrics.get('n_lines')}",
f"- unmatched labels: {metrics.get('unmatched_labels')}",
f"- unmatched predictions: {metrics.get('unmatched_predictions')}",
]
)
expected_lines.extend(
[
f"- eval pack version: {eval_run.get('eval_pack_version') or '(missing)'}",
f"- frozen eval pack: {eval_run.get('frozen_eval_pack')}",
f"- unseen for selection: {eval_run.get('unseen_for_selection')}",
f"- selected checkpoint on test: {eval_run.get('selected_checkpoint_on_test')}",
f"- languages: {eval_run.get('languages')}",
f"- scripts: {eval_run.get('scripts')}",
f"- domains: {eval_run.get('domains')}",
f"- GT statuses: {eval_run.get('gt_statuses')}",
f"- splits: {eval_run.get('splits')}",
f"- script inventory file: {script_inventory.get('inventory_file')}",
f"- observed codepoints/graphemes: {script_inventory.get('observed_codepoint_count')}/{script_inventory.get('observed_grapheme_count')}",
f"- missing required codepoints: {script_inventory.get('missing_required_codepoints')}",
f"- missing required graphemes: {script_inventory.get('missing_required_graphemes')}",
f"- metadata contradictions: {metrics.get('metadata_contradictions', [])}",
]
)
if mode == "page_level_ocr":
metric_means = metrics.get("metric_means") if isinstance(metrics.get("metric_means"), dict) else {}
normalization = metrics.get("normalization_summary") if isinstance(metrics.get("normalization_summary"), dict) else {}
dictionary_coverage = metrics.get("dictionary_coverage") if isinstance(metrics.get("dictionary_coverage"), dict) else {}
expected_lines.extend(
[
f"- metadata coverage: {metrics.get('metadata_coverage')}",
f"- reference coverage: {metrics.get('reference_coverage')}",
f"- normalization changed reference/prediction/any: {normalization.get('reference_changed')}/{normalization.get('prediction_changed')}/{normalization.get('any_changed')} of {normalization.get('total')}",
f"- dictionary coverage checked: {dictionary_coverage.get('checked')}",
f"- dictionary reference codepoints/dictionary codepoints: {dictionary_coverage.get('gt_codepoint_count')}/{dictionary_coverage.get('dictionary_codepoint_count')}",
f"- dictionary missing reference codepoints: {dictionary_coverage.get('missing_gt_codepoints')}",
f"- leakage checked: {leakage.get('checked')}",
f"- text codepoint CER micro: {_fmt_report_metric(metrics.get('text_cer_codepoint_micro'))}",
f"- text grapheme CER micro: {_fmt_report_metric(metrics.get('text_cer_grapheme_micro'))}",
f"- text exact accuracy: {_fmt_report_metric(metrics.get('text_exact_accuracy'))}",
f"- line detection F1 mean: {_fmt_report_metric(metric_means.get('line_detection_f1'))}",
f"- line mean IoU: {_fmt_report_metric(metric_means.get('line_mean_iou'))}",
f"- reading-order pair accuracy: {_fmt_report_metric(metric_means.get('reading_order_pair_accuracy'))}",
f"- text codepoint edits/ref: {metrics.get('text_codepoint_edits')}/{metrics.get('text_codepoint_ref_len')}",
f"- text grapheme edits/ref: {metrics.get('text_grapheme_edits')}/{metrics.get('text_grapheme_ref_len')}",
f"- text exact pages/ref pages: {metrics.get('text_exact_pages')}/{metrics.get('text_reference_pages')}",
f"- zero-length text references: {metrics.get('zero_length_text_references')}",
f"- uncertainty method: {uncertainty.get('method')}",
f"- uncertainty scoring units: {uncertainty.get('effective_sample_size')}/{uncertainty.get('min_claim_scoring_units')}",
f"- text grapheme CER 95% CI: {_fmt_report_metric((intervals.get('text_cer_grapheme_micro') or {}).get('low'))}..{_fmt_report_metric((intervals.get('text_cer_grapheme_micro') or {}).get('high'))}",
]
)
slices = metrics.get("slices") if isinstance(metrics.get("slices"), list) else []
for row in slices:
if not isinstance(row, dict):
continue
row_means = row.get("metric_means") if isinstance(row.get("metric_means"), dict) else {}
expected_lines.append(
f"- {row.get('language')} / {row.get('script')} / {row.get('domain')} / {row.get('gt_status')} / {row.get('split')}: "
f"pages={row.get('scored_pages')} text_graph_CER={_fmt_report_metric(row.get('text_cer_grapheme_micro'))} "
f"text_cp_CER={_fmt_report_metric(row.get('text_cer_codepoint_micro'))} "
f"exact={_fmt_report_metric(row.get('text_exact_accuracy'))} line_F1={_fmt_report_metric(row_means.get('line_detection_f1', 0.0))}"
)
expected_lines.extend(
[
"- audit JSON: audit.json",
"- eval run JSON: score/page-eval-run.json",
"- metrics JSON: score/metrics.json",
"- metrics Markdown: score/metrics.md",
"- command text: score/command.txt",
"- page metrics CSV: score/page-metrics.csv",
"- reference artifacts TSV: score/reference-artifacts.tsv",
"- text pairs TSV: score/text-pairs.tsv",
"- missing TSV: score/missing.tsv",
"- worst pages: score/worst-pages.csv",
"- line alignment CSV: score/line-alignment.csv",
"- review CSV: score/sample-review.csv",
"- review HTML: score/sample-review.html",
"- confusions JSON: score/confusions.json",
]
)
else:
normalization = metrics.get("normalization_summary") if isinstance(metrics.get("normalization_summary"), dict) else {}
dictionary_coverage = metrics.get("dictionary_coverage") if isinstance(metrics.get("dictionary_coverage"), dict) else {}
expected_lines.extend(
[
f"- provenance coverage: {metrics.get('provenance_coverage')}",
f"- normalizer counts: {metrics.get('normalizer_counts', {})}",
f"- normalization changed GT/prediction/any: {normalization.get('gt_changed')}/{normalization.get('prediction_changed')}/{normalization.get('any_changed')} of {normalization.get('total')}",
f"- dictionary coverage checked: {dictionary_coverage.get('checked')}",
f"- dictionary GT codepoints/dictionary codepoints: {dictionary_coverage.get('gt_codepoint_count')}/{dictionary_coverage.get('dictionary_codepoint_count')}",
f"- dictionary missing GT codepoints: {dictionary_coverage.get('missing_gt_codepoints')}",
f"- leakage checked: {leakage.get('checked')}",
f"- codepoint CER micro: {_fmt_report_metric(metrics.get('cer_codepoint_micro'))}",
f"- grapheme CER micro: {_fmt_report_metric(metrics.get('cer_grapheme_micro'))}",
f"- exact-line accuracy: {_fmt_report_metric(metrics.get('exact_line_accuracy'))}",
f"- codepoint edits/ref: {metrics.get('codepoint_edits')}/{metrics.get('codepoint_ref_len')}",
f"- grapheme edits/ref: {metrics.get('grapheme_edits')}/{metrics.get('grapheme_ref_len')}",
f"- zero-length GT lines: {metrics.get('zero_length_gt_lines')}",
f"- uncertainty method: {uncertainty.get('method')}",
f"- uncertainty scoring units: {uncertainty.get('effective_sample_size')}/{uncertainty.get('min_claim_scoring_units')}",
f"- grapheme CER 95% CI: {_fmt_report_metric((intervals.get('cer_grapheme_micro') or {}).get('low'))}..{_fmt_report_metric((intervals.get('cer_grapheme_micro') or {}).get('high'))}",
]
)
slices = metrics.get("slices") if isinstance(metrics.get("slices"), list) else []
for row in slices:
if not isinstance(row, dict):
continue
expected_lines.append(
f"- {row.get('language')} / {row.get('script')} / {row.get('domain')} / {row.get('gt_status')} / {row.get('split')}: "
f"n={row.get('n_lines')} graph_CER={_fmt_report_metric(row.get('cer_grapheme_micro'))} "
f"cp_CER={_fmt_report_metric(row.get('cer_codepoint_micro'))} exact={_fmt_report_metric(row.get('exact_line_accuracy'))}"
)
expected_lines.extend(
[
"- audit JSON: audit.json",
"- eval run JSON: score/eval-run.json",
"- metrics JSON: score/metrics.json",
"- metrics Markdown: score/metrics.md",
"- command text: score/command.txt",
"- predictions TSV: score/predictions.tsv",
"- unmatched TSV: score/unmatched.tsv",
"- worst lines: score/worst-lines.csv",
"- review CSV: score/sample-review.csv",
"- review HTML: score/sample-review.html",
"- confusions JSON: score/confusions.json",
]
)
for line in expected_lines:
_expect_root_report_line(text, issues, line)
warnings = eval_run.get("warnings") if isinstance(eval_run.get("warnings"), list) else []
if warnings:
for warning in warnings:
_expect_root_report_line(text, issues, f"- {warning}")
else:
_expect_root_report_line(text, issues, "- none")
return issues
def _check_timestamp(payload: dict[str, Any], field: str, label: str) -> list[str]:
value = payload.get(field)
if not isinstance(value, str) or not value:
return [f"{label} {field} timestamp is missing"]
try:
datetime.fromisoformat(value.replace("Z", "+00:00"))
except ValueError:
return [f"{label} {field} timestamp is invalid: {value}"]
return []
def _check_required_fields(payload: dict[str, Any], fields: tuple[str, ...], label: str) -> list[str]:
issues: list[str] = []
for field in fields:
if field not in payload:
issues.append(f"{label} missing required field: {field}")
elif payload[field] in (None, "") and field not in {"checkpoint", "config", "dictionary"}:
issues.append(f"{label} has empty required field: {field}")
return issues
def _check_normalization_profile(eval_run: dict[str, Any]) -> list[str]:
profile = eval_run.get("normalization_profile")
if profile != SUPPORTED_NORMALIZATION_PROFILE:
return [
"eval run normalization_profile does not match supported metric recomputation profile: "
f"recorded={profile!r} expected={SUPPORTED_NORMALIZATION_PROFILE!r}"
]
return []
def _check_normalization_summary(recorded: Any, expected: dict[str, Any], label: str) -> list[str]:
if not isinstance(recorded, dict):
return [f"{label} normalization_summary missing or not an object"]
issues: list[str] = []
for key, expected_value in expected.items():
if recorded.get(key) != expected_value:
issues.append(f"{label} normalization_summary {key} mismatch: recorded={recorded.get(key)} recomputed={expected_value}")
return issues
def _expected_recognizer_normalization_summary(score_dir: Path) -> dict[str, Any]:
rows = _tsv_rows(score_dir / "predictions.tsv")
gt_changed = sum(1 for row in rows if (row.get("gt_text") or "") != _nfc(row.get("gt_text") or ""))
prediction_changed = sum(1 for row in rows if (row.get("prediction") or "") != _nfc(row.get("prediction") or ""))
return {
"profile": SUPPORTED_NORMALIZATION_PROFILE,
"unit": "line",
"total": len(rows),
"gt_changed": gt_changed,
"prediction_changed": prediction_changed,
"any_changed": sum(
1
for row in rows
if (row.get("gt_text") or "") != _nfc(row.get("gt_text") or "")
or (row.get("prediction") or "") != _nfc(row.get("prediction") or "")
),
}
def _expected_page_normalization_summary(score_dir: Path) -> dict[str, Any]:
rows = _tsv_rows(score_dir / "text-pairs.tsv")
prediction_changed = sum(1 for row in rows if (row.get("prediction_text") or "") != _nfc(row.get("prediction_text") or ""))
reference_changed = sum(1 for row in rows if (row.get("reference_text") or "") != _nfc(row.get("reference_text") or ""))
return {
"profile": SUPPORTED_NORMALIZATION_PROFILE,
"unit": "page",
"total": len(rows),
"prediction_changed": prediction_changed,
"reference_changed": reference_changed,
"any_changed": sum(
1
for row in rows
if (row.get("prediction_text") or "") != _nfc(row.get("prediction_text") or "")
or (row.get("reference_text") or "") != _nfc(row.get("reference_text") or "")
),
}
def _check_machine_metadata(payload: dict[str, Any], label: str) -> list[str]:
machine = payload.get("machine")
if not isinstance(machine, dict):
return [f"{label} machine metadata is missing or not an object"]
issues: list[str] = []
for field in ("platform", "python"):
value = machine.get(field)
if not isinstance(value, str) or not value:
issues.append(f"{label} machine.{field} is missing")
return issues
def _check_common(score_dir: Path, eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
issues: list[str] = []
if eval_run.get("claim_safe") is True and eval_run.get("warnings"):
issues.append("claim_safe is true but warnings are present")
if eval_run.get("claim_safe") is False and not eval_run.get("warnings"):
issues.append("claim_safe is false but warnings are empty")
issues.extend(_check_machine_metadata(eval_run, "eval run"))
if not isinstance(eval_run.get("command"), str) or not eval_run["command"]:
issues.append("eval run command is missing")
else:
command_file = score_dir / "command.txt"
if command_file.exists() and command_file.read_text(encoding="utf-8").strip() != eval_run["command"].strip():
issues.append("command.txt does not match eval run command")
issues.extend(_check_normalization_profile(eval_run))
issues.extend(_check_timestamp(eval_run, "created_at", "eval run"))
metrics_md = score_dir / "metrics.md"
if metrics_md.exists() and not metrics_md.read_text(encoding="utf-8").strip():
issues.append("metrics.md is blank")
elif metrics_md.exists():
issues.extend(_check_metrics_md_consistency(score_dir, eval_run, metrics))
claim_report = score_dir / "claim-safety.md"
if claim_report.exists():
text = claim_report.read_text(encoding="utf-8")
if not text.strip():
issues.append("claim-safety.md is blank")
for marker in ("## What Was Scored", "## Ground Truth And Pack", "## Metrics", "## Warnings"):
if marker not in text:
issues.append(f"claim-safety.md missing section: {marker}")
issues.extend(_check_claim_safety_report_consistency(score_dir, eval_run, metrics))
if not isinstance(metrics, dict):
issues.append("metrics artifact is not a JSON object")
return issues
def _warning_texts(eval_run: dict[str, Any]) -> list[str]:
warnings = eval_run.get("warnings")
if not isinstance(warnings, list):
return []
return [str(warning) for warning in warnings]
def _has_warning(eval_run: dict[str, Any], prefix: str) -> bool:
return any(warning.startswith(prefix) for warning in _warning_texts(eval_run))
def _claim_blocker(issues: list[str], eval_run: dict[str, Any], mode: str, reason: str, warning_prefix: str) -> None:
if eval_run.get("claim_safe") is True:
issues.append(f"{mode} claim_safe is true despite {reason}")
if not _has_warning(eval_run, warning_prefix):
issues.append(f"{mode} missing claim-safety warning for {reason}: {warning_prefix}")
def _string_set(value: Any) -> set[str]:
if not isinstance(value, list):
return set()
return {str(item) for item in value}
def _check_eval_run_identity_lists(
eval_run: dict[str, Any],
rows: list[dict[str, str]],
*,
run_label: str,
row_label: str,
) -> list[str]:
issues: list[str] = []
if not rows:
return issues
mappings = (
("languages", "language"),
("scripts", "script"),
("domains", "domain"),
("gt_statuses", "gt_status"),
("splits", "split"),
)
for run_field, row_field in mappings:
recorded = eval_run.get(run_field)
expected = sorted({row.get(row_field) or "" for row in rows})
if not isinstance(recorded, list):
issues.append(f"{run_label} {run_field} is not a list for {row_label} identity audit")
continue
recorded_values = [str(item) for item in recorded]
if recorded_values != expected:
issues.append(
f"{run_label} {run_field} does not match {row_label} {row_field} values: "
f"recorded={recorded_values} recomputed={expected}"
)
return issues
def _check_shared_claim_blockers(mode: str, eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
issues: list[str] = []
if not isinstance(eval_run.get("model_name"), str) or not eval_run["model_name"].strip():
_claim_blocker(issues, eval_run, mode, "missing model identity", "model name not recorded")
if mode == "page" and (not isinstance(eval_run.get("system_name"), str) or not eval_run["system_name"].strip()):
_claim_blocker(issues, eval_run, mode, "missing system identity", "system name not recorded")
if not eval_run.get("eval_pack_version"):
_claim_blocker(issues, eval_run, mode, "missing eval-pack version", "eval pack version not recorded")
artifact_identity = eval_run.get("artifact_identity")
if isinstance(artifact_identity, dict):
for name, identity in sorted(artifact_identity.items()):
if isinstance(identity, dict) and not identity.get("verified"):
_claim_blocker(issues, eval_run, mode, f"unverified {name} artifact", f"{name} artifact is not hash-verifiable:")
if eval_run.get("frozen_eval_pack") is not True:
_claim_blocker(issues, eval_run, mode, "non-frozen eval pack", "eval pack is not declared frozen")
if eval_run.get("unseen_for_selection") is not True:
_claim_blocker(issues, eval_run, mode, "eval pack not declared unseen for selection", "eval pack is not declared unseen for model/checkpoint selection")
if eval_run.get("selected_checkpoint_on_test") is True:
_claim_blocker(issues, eval_run, mode, "checkpoint selection on test data", "checkpoint/model selection on this test pack is declared")
if not eval_run.get("summary_file"):
_claim_blocker(issues, eval_run, mode, "missing eval pack summary/datasheet", "eval pack summary/datasheet missing")
if metrics.get("metadata_contradictions"):
_claim_blocker(issues, eval_run, mode, "metadata contradictions", "metadata contradictions detected:")
if eval_run.get("known_leakage"):
_claim_blocker(issues, eval_run, mode, "known leakage acknowledgement", "known leakage acknowledged:")
languages = _string_set(eval_run.get("languages"))
scripts = _string_set(eval_run.get("scripts"))
domains = _string_set(eval_run.get("domains"))
gt_statuses = _string_set(eval_run.get("gt_statuses"))
splits = _string_set(eval_run.get("splits"))
if len(languages) > 1:
_claim_blocker(issues, eval_run, mode, "mixed languages", "mixed languages:")
if len(scripts) > 1:
_claim_blocker(issues, eval_run, mode, "mixed scripts", "mixed scripts:")
if len(domains) > 1:
_claim_blocker(issues, eval_run, mode, "mixed eval domains", "mixed eval domains:")
if len(gt_statuses) > 1:
_claim_blocker(issues, eval_run, mode, "mixed GT statuses", "mixed GT statuses:")
if len(splits) > 1:
_claim_blocker(issues, eval_run, mode, "mixed split statuses", "mixed split statuses:")
if splits and not splits <= CLAIM_EVAL_SPLITS:
_claim_blocker(issues, eval_run, mode, "unexpected split status", "unexpected split status for claim evaluation:")
if domains and not domains <= CLAIM_SAFE_DOMAINS:
_claim_blocker(issues, eval_run, mode, "unsafe eval domain", "not claim-safe real domain:")
if gt_statuses and not gt_statuses <= CLAIM_SAFE_GT_STATUSES:
_claim_blocker(issues, eval_run, mode, "untrusted GT status", "not human-verified/spot-checked GT:")
uncertainty = metrics.get("uncertainty")
if not isinstance(uncertainty, dict):
_claim_blocker(issues, eval_run, mode, "missing claim uncertainty evidence", "claim-facing uncertainty missing")
elif uncertainty.get("claim_sample_size_ok") is not True:
_claim_blocker(issues, eval_run, mode, "claim-facing sample size too small", "claim-facing sample size too small:")
slices = metrics.get("slices")
if isinstance(slices, list):
for row in slices:
if not isinstance(row, dict):
continue
slice_uncertainty = row.get("uncertainty")
if isinstance(slice_uncertainty, dict):
count = int(slice_uncertainty.get("effective_sample_size", 0) or 0)
minimum = int(slice_uncertainty.get("min_claim_scoring_units", MIN_CLAIM_SCORING_UNITS) or MIN_CLAIM_SCORING_UNITS)
ok = slice_uncertainty.get("claim_sample_size_ok") is True
else:
count = int(row.get("n_lines", row.get("scored_pages", 0)) or 0)
minimum = MIN_CLAIM_SCORING_UNITS
ok = count >= minimum
if not ok or count < minimum:
_claim_blocker(issues, eval_run, mode, "slice sample size too small", "slice sample size too small:")
script_inventory = metrics.get("script_inventory")
if not isinstance(script_inventory, dict) or not script_inventory.get("inventory_file"):
_claim_blocker(issues, eval_run, mode, "missing script inventory evidence", "script inventory file missing")
elif script_inventory.get("missing_required_codepoints") or script_inventory.get("missing_required_graphemes"):
if script_inventory.get("missing_required_codepoints"):
_claim_blocker(issues, eval_run, mode, "missing required script codepoints", "missing required script codepoints:")
if script_inventory.get("missing_required_graphemes"):
_claim_blocker(issues, eval_run, mode, "missing required script graphemes", "missing required script graphemes:")
if "machine_converted_spot_checked" in gt_statuses:
calibration = eval_run.get("spotcheck_calibration")
if not isinstance(calibration, dict):
_claim_blocker(
issues,
eval_run,
mode,
"spot-checked machine GT without calibration artifact",
"machine_converted_spot_checked GT declared without native spot-check calibration artifact",
)
elif not calibration.get("claim_evidence_eligible"):
_claim_blocker(
issues,
eval_run,
mode,
"ineligible native spot-check calibration",
"native spot-check calibration artifact is not claim evidence eligible",
)
leakage = metrics.get("leakage")
if not isinstance(leakage, dict):
_claim_blocker(issues, eval_run, mode, "unrun leakage check", "train/dev leakage check not run")
return issues
if not leakage.get("checked"):
train_manifests = eval_run.get("train_manifests")
warning_prefix = (
"train/dev leakage manifests were provided but no train rows were checked"
if isinstance(train_manifests, list) and train_manifests
else "train/dev leakage check not run"
)
_claim_blocker(issues, eval_run, mode, "unrun leakage check", warning_prefix)
overlaps = leakage.get("overlaps")
if not isinstance(overlaps, dict):
return issues
hard_leakage = {
key: values
for key, values in overlaps.items()
if key != "text_sha256" and isinstance(values, list) and values
}
if hard_leakage:
_claim_blocker(issues, eval_run, mode, "train/dev/eval leakage", "train/dev/eval leakage detected:")
text_overlap = overlaps.get("text_sha256")
if isinstance(text_overlap, list) and text_overlap:
_claim_blocker(issues, eval_run, mode, "train/dev/eval text overlap", "train/dev/eval text overlap detected:")
return issues
def _check_recognizer_claim_blockers(eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
issues = _check_shared_claim_blockers("recognizer", eval_run, metrics)
matched_lines = metrics.get("matched_lines")
if matched_lines == 0:
_claim_blocker(issues, eval_run, "recognizer", "no matched rows scored", "no matched rows scored")
unmatched_count = int(metrics.get("unmatched_labels", 0) or 0) + int(metrics.get("unmatched_predictions", 0) or 0)
if unmatched_count:
_claim_blocker(issues, eval_run, "recognizer", "unmatched prediction/label rows", "unmatched rows present:")
match_rate = metrics.get("prediction_match_rate")
min_match_rate = eval_run.get("min_match_rate")
if _is_number(match_rate) and _is_number(min_match_rate) and float(match_rate) < float(min_match_rate):
_claim_blocker(issues, eval_run, "recognizer", "low prediction match rate", "low prediction match rate:")
if not eval_run.get("manifest_file"):
_claim_blocker(issues, eval_run, "recognizer", "missing eval manifest", "eval manifest missing")
if eval_run.get("claim_eligible_default") is not True:
_claim_blocker(issues, eval_run, "recognizer", "rows not guaranteed claim eligible", "not all rows are claim eligible:")
coverage = metrics.get("provenance_coverage")
if isinstance(coverage, dict):
for key, count in sorted(coverage.items()):
if count:
_claim_blocker(issues, eval_run, "recognizer", f"missing provenance field {key}", f"{key}:")
return issues
def _check_recognizer_row_claim_eligibility_blockers(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
rows = _tsv_rows(score_dir / "predictions.tsv")
if not rows:
return []
ineligible = [
row.get("sample_id") or row.get("image_path") or f"row-{index}"
for index, row in enumerate(rows, 1)
if row.get("claim_eligible") != "True"
]
if not ineligible:
return []
issues: list[str] = []
_claim_blocker(issues, eval_run, "recognizer", "row-level claim ineligible values", "not all rows are claim eligible:")
return issues
def _read_page_output_audit_for_replay(
sample_id: str,
output_dir: Path,
*,
filename: str,
passed_field: str,
) -> dict[str, Any]:
audit_path = output_dir / filename
row: dict[str, Any] = {
"sample_id": sample_id,
"output_dir": str(output_dir),
"audit_path": str(audit_path),
"exists": audit_path.is_file(),
"passed": False,
"sha256": None,
"size_bytes": None,
"issue": "",
}
if not audit_path.is_file():
row["issue"] = f"missing page output audit: {audit_path}"
return row
row["sha256"] = _sha256_file(audit_path)
row["size_bytes"] = audit_path.stat().st_size
try:
payload = json.loads(audit_path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
row["issue"] = f"page output audit is invalid JSON: {audit_path}: {exc}"
return row
if not isinstance(payload, dict):
row["issue"] = f"page output audit must contain a JSON object: {audit_path}"
return row
passed = payload.get(passed_field)
row["passed"] = passed is True
if passed is not True:
row["issue"] = f"page output audit did not pass: {audit_path} {passed_field}={passed!r}"
return row
def _expected_page_output_audit(score_dir: Path, eval_run: dict[str, Any]) -> dict[str, Any]:
required = bool(eval_run.get("require_output_audit"))
filename = str(eval_run.get("output_audit_filename") or "limbu-output-audit.json")
passed_field = str(eval_run.get("output_audit_passed_field") or "passed")
output_audit_rows: list[dict[str, Any]] = []
page_rows = _csv_rows(score_dir / "page-metrics.csv")
if required:
for row in page_rows:
sample_id = str(row.get("sample_id") or "")
document_path_value = row.get("document_path") or ""
if not document_path_value:
output_audit_rows.append(
{
"sample_id": sample_id,
"output_dir": "",
"audit_path": filename,
"exists": False,
"passed": False,
"sha256": None,
"size_bytes": None,
"issue": f"page-metrics.csv row has no document_path for output audit replay: {sample_id}",
}
)
continue
output_audit_rows.append(
_read_page_output_audit_for_replay(
sample_id,
Path(document_path_value).parent,
filename=filename,
passed_field=passed_field,
)
)
return {
"required": required,
"filename": filename,
"passed_field": passed_field,
"expected_count": len(page_rows) if required else 0,
"checked_count": len(output_audit_rows),
"passed_count": sum(1 for row in output_audit_rows if row.get("passed") is True),
"missing_count": sum(1 for row in output_audit_rows if row.get("exists") is not True),
"failed_count": sum(1 for row in output_audit_rows if row.get("exists") is True and row.get("passed") is not True),
"issues": [str(row.get("issue")) for row in output_audit_rows if row.get("issue")],
"artifacts": output_audit_rows,
}
def _check_page_output_audit(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
expected = _expected_page_output_audit(score_dir, eval_run)
recorded_metrics = metrics.get("page_output_audit")
if not isinstance(recorded_metrics, dict):
issues.append("metrics.json missing page_output_audit object")
elif recorded_metrics != expected:
issues.append("metrics.json page_output_audit does not match current page output audit artifacts")
recorded_eval_run = eval_run.get("page_output_audit")
if not isinstance(recorded_eval_run, dict):
issues.append("page-eval-run.json missing page_output_audit object")
elif recorded_eval_run != expected:
issues.append("page-eval-run.json page_output_audit does not match current page output audit artifacts")
if expected["required"] and expected["passed_count"] != expected["expected_count"]:
issues.append(
"replayed page output audits are not all passing: "
f"{expected['passed_count']}/{expected['expected_count']} filename={expected['filename']!r}"
)
return issues
def _check_page_claim_blockers(eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
issues = _check_shared_claim_blockers("page", eval_run, metrics)
if int(metrics.get("missing_documents", 0) or 0):
_claim_blocker(issues, eval_run, "page", "missing document outputs", "missing document outputs:")
if int(metrics.get("missing_references", 0) or 0):
_claim_blocker(issues, eval_run, "page", "missing references", "missing references:")
scored_pages = metrics.get("scored_pages")
manifest_samples = metrics.get("manifest_samples")
if _is_number(scored_pages) and _is_number(manifest_samples) and int(scored_pages) < int(manifest_samples):
_claim_blocker(issues, eval_run, "page", "incomplete page scoring coverage", "incomplete page scoring coverage:")
if metrics.get("document_source_path_mismatches"):
_claim_blocker(issues, eval_run, "page", "document source_path mismatches", "document source_path mismatches detected:")
if eval_run.get("claim_eligible_default") is not True:
_claim_blocker(issues, eval_run, "page", "eval pack not declared claim eligible", "eval pack is not declared claim eligible")
claim_eligible_rows = metrics.get("claim_eligible_rows")
if _is_number(claim_eligible_rows) and _is_number(manifest_samples) and int(claim_eligible_rows) != int(manifest_samples):
_claim_blocker(issues, eval_run, "page", "not all manifest rows are claim eligible", "not all manifest rows are claim eligible:")
metadata_coverage = metrics.get("metadata_coverage")
if isinstance(metadata_coverage, dict):
for key, count in sorted(metadata_coverage.items()):
if count:
_claim_blocker(issues, eval_run, "page", f"missing row metadata {key}", f"{key}:")
reference_coverage = metrics.get("reference_coverage")
if isinstance(reference_coverage, dict):
if not reference_coverage.get("text"):
_claim_blocker(issues, eval_run, "page", "no text references scored", "no text references scored")
if not reference_coverage.get("reading_order"):
_claim_blocker(issues, eval_run, "page", "no reading-order references scored", "no reading-order references scored")
if not reference_coverage.get("line_alignment_scored"):
_claim_blocker(issues, eval_run, "page", "no line-bbox alignment references scored", "no line-bbox alignment references scored")
if not (reference_coverage.get("tables") or reference_coverage.get("figures")):
_claim_blocker(issues, eval_run, "page", "no layout detection references scored", "no layout detection references scored")
return issues
def _check_ratio(metrics: dict[str, Any], numerator: str, denominator: str, ratio: str) -> list[str]:
issues: list[str] = []
for key in (numerator, denominator, ratio):
if key not in metrics or not _is_number(metrics[key]):
issues.append(f"metrics missing numeric field: {key}")
return issues
expected = metrics[numerator] / max(metrics[denominator], 1)
if not _close(float(metrics[ratio]), float(expected)):
issues.append(f"metrics {ratio} does not equal {numerator}/{denominator}")
return issues
def _nfc(value: str) -> str:
return unicodedata.normalize("NFC", value or "")
def _graphemes(value: str) -> list[str]:
return regex.findall(r"\X", _nfc(value))
def _codepoint_id(ch: str) -> str:
return f"U+{ord(ch):04X}"
def _hash_existing_file(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _read_script_inventory(path: Path | None) -> tuple[dict[str, Any], list[str]]:
if path is None:
return {}, []
if not path.exists():
return {}, [f"script inventory file does not exist: {path}"]
try:
payload = json.loads(path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return {}, [f"script inventory file is invalid JSON: {path}: {exc}"]
if not isinstance(payload, dict):
return {}, [f"script inventory file must contain a JSON object: {path}"]
return payload, []
def _required_codepoints(payload: dict[str, Any]) -> tuple[list[str], list[str]]:
raw = payload.get("required_codepoints", [])
if not isinstance(raw, list):
return [], ["script inventory required_codepoints must be a list"]
required: set[str] = set()
issues: list[str] = []
for item in raw:
if not isinstance(item, str):
issues.append(f"script inventory codepoint entry must be a string: {item!r}")
continue
value = item.strip()
if not value:
continue
if value.upper().startswith("U+"):
try:
required.add(f"U+{int(value[2:], 16):04X}")
except ValueError:
issues.append(f"script inventory codepoint entry is not valid hex: {item!r}")
continue
normalized = _nfc(value)
if len(normalized) != 1:
issues.append(f"script inventory codepoint entry must be one character or U+XXXX: {item!r}")
continue
required.add(_codepoint_id(normalized))
return sorted(required), issues
def _required_graphemes(payload: dict[str, Any]) -> tuple[list[str], list[str]]:
raw = payload.get("required_graphemes", [])
if not isinstance(raw, list):
return [], ["script inventory required_graphemes must be a list"]
required: set[str] = set()
issues: list[str] = []
for item in raw:
if not isinstance(item, str):
issues.append(f"script inventory grapheme entry must be a string: {item!r}")
continue
value = _nfc(item)
if value.strip():
required.add(value)
return sorted(required), issues
def _expected_script_inventory(texts: list[str], inventory_path: Path | None) -> tuple[dict[str, Any], list[str]]:
payload, issues = _read_script_inventory(inventory_path)
required_cps, cp_issues = _required_codepoints(payload)
required_graphs, graph_issues = _required_graphemes(payload)
issues.extend(cp_issues)
issues.extend(graph_issues)
text = _nfc("\n".join(texts))
codepoints = sorted({_codepoint_id(ch) for ch in text if not ch.isspace()})
graphemes = sorted({cluster for cluster in _graphemes(text) if cluster.strip()})
categories = [unicodedata.category(ch) for ch in text if not ch.isspace()]
return {
"inventory_file": str(inventory_path) if inventory_path else None,
"inventory_file_sha256": _hash_existing_file(inventory_path) if inventory_path and inventory_path.exists() else None,
"observed_codepoint_count": len(codepoints),
"observed_grapheme_count": len(graphemes),
"observed_codepoints": codepoints,
"observed_graphemes": graphemes,
"required_codepoints": required_cps,
"required_graphemes": required_graphs,
"missing_required_codepoints": sorted(set(required_cps) - set(codepoints)),
"missing_required_graphemes": sorted(set(required_graphs) - set(graphemes)),
"digit_codepoint_count": sum(1 for ch in text if not ch.isspace() and unicodedata.category(ch) == "Nd"),
"punctuation_codepoint_count": sum(1 for category in categories if category.startswith("P")),
"combining_mark_codepoint_count": sum(1 for category in categories if category.startswith("M")),
}, issues
def _check_script_inventory(recorded: Any, expected: dict[str, Any], label: str) -> list[str]:
if not isinstance(recorded, dict):
return [f"{label} script_inventory missing or not an object"]
issues: list[str] = []
for key, expected_value in expected.items():
if recorded.get(key) != expected_value:
issues.append(f"{label} script_inventory {key} mismatch: recorded={recorded.get(key)} recomputed={expected_value}")
return issues
def _dictionary_codepoints(dictionary: str) -> set[str] | None:
path = Path(str(dictionary or ""))
if not path.exists() or not path.is_file():
return None
codepoints: set[str] = set()
for line in path.read_text(encoding="utf-8").splitlines():
token = _nfc(line.strip())
if not token:
continue
for ch in token:
if not ch.isspace():
codepoints.add(_codepoint_id(ch))
return codepoints
def _expected_dictionary_coverage(texts: list[str], dictionary: Any) -> dict[str, Any]:
dictionary_text = str(dictionary or "")
dictionary_path = Path(dictionary_text)
gt_codepoints = sorted({_codepoint_id(ch) for text in texts for ch in _nfc(text) if not ch.isspace()})
dictionary_codepoints = _dictionary_codepoints(dictionary_text)
if dictionary_codepoints is None:
return {
"dictionary": dictionary_text,
"dictionary_file_sha256": None,
"checked": False,
"gt_codepoint_count": len(gt_codepoints),
"missing_gt_codepoints": [],
"missing_gt_codepoint_count": 0,
"reason": "dictionary file is not readable",
}
missing = sorted(set(gt_codepoints) - dictionary_codepoints)
return {
"dictionary": dictionary_text,
"dictionary_file_sha256": _hash_existing_file(dictionary_path),
"checked": True,
"dictionary_codepoint_count": len(dictionary_codepoints),
"gt_codepoint_count": len(gt_codepoints),
"missing_gt_codepoints": missing,
"missing_gt_codepoint_count": len(missing),
}
def _check_dictionary_coverage(recorded: Any, expected: dict[str, Any], label: str) -> list[str]:
if not isinstance(recorded, dict):
return [f"{label} dictionary_coverage missing or not an object"]
issues: list[str] = []
for key, expected_value in expected.items():
if recorded.get(key) != expected_value:
issues.append(f"{label} dictionary_coverage {key} mismatch: recorded={recorded.get(key)} recomputed={expected_value}")
return issues
def _bool_text(value: str) -> bool | None:
lowered = str(value).strip().lower()
if lowered in {"true", "1", "yes", "y"}:
return True
if lowered in {"false", "0", "no", "n"}:
return False
return None
def _metadata_bool_value(metadata: dict[str, Any], keys: tuple[str, ...], fallback: Any) -> bool:
for key in keys:
if key not in metadata:
continue
value = metadata[key]
if isinstance(value, bool):
return value
if isinstance(value, str):
parsed = _bool_text(value)
if parsed is not None:
return parsed
return bool(value)
return fallback is True
def _metadata_signal_text(metadata: dict[str, Any]) -> str:
values: list[str] = []
for key in PROVENANCE_SIGNAL_KEYS:
value = metadata.get(key)
if value in (None, "", [], {}):
continue
if isinstance(value, list):
values.extend(str(item) for item in value)
elif isinstance(value, dict):
values.extend(f"{item_key}:{item_value}" for item_key, item_value in value.items())
else:
values.append(str(value))
return " ".join(values).lower().replace("_", " ").replace("-", " ")
def _metadata_contradictions_for(sample_id: str, metadata: dict[str, Any], domain: str, gt_status: str) -> list[str]:
signal_text = _metadata_signal_text(metadata)
if not signal_text:
return []
contradictions: list[str] = []
if domain in CLAIM_SAFE_DOMAINS and any(term in signal_text for term in ("synthetic", "transliterated", "transliteration")):
contradictions.append(f"{sample_id}: synthetic/transliterated provenance is declared as real domain {domain}")
if gt_status == "human_verified" and any(
term in signal_text
for term in ("machine converted", "machine", "synthetic", "transliterated", "transliteration")
):
contradictions.append(f"{sample_id}: machine/synthetic provenance is declared as human_verified GT")
return contradictions
def _metadata_has_any(metadata: dict[str, Any], keys: tuple[str, ...]) -> bool:
for key in keys:
if metadata.get(key) not in (None, "", [], {}):
return True
return False
def _metadata_value_for_keys(metadata: dict[str, Any], keys: tuple[str, ...]) -> str:
for key in keys:
value = metadata.get(key)
if value not in (None, "", [], {}):
if isinstance(value, (list, dict)):
return json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
return str(value)
return ""
def _bbox_text(bbox: BBox | None) -> str:
if bbox is None:
return ""
return json.dumps(bbox.to_list(), ensure_ascii=False, separators=(",", ":"))
def _int_text(value: str, field: str, row_index: int, issues: list[str]) -> int | None:
try:
return int(value)
except (TypeError, ValueError):
issues.append(f"predictions.tsv row {row_index} has non-integer {field}: {value!r}")
return None
def _float_text(value: str, field: str, row_index: int, issues: list[str]) -> float | None:
try:
return float(value)
except (TypeError, ValueError):
issues.append(f"CSV row {row_index} has non-numeric {field}: {value!r}")
return None
def _check_recognizer_predictions_recompute(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
rows = _tsv_rows(score_dir / "predictions.tsv")
if not rows:
return ["predictions.tsv has no scored rows"]
cp_edits_total = 0
cp_len_total = 0
gr_edits_total = 0
gr_len_total = 0
exact_total = 0
zero_length_gt = 0
slice_totals: dict[tuple[str, str, str, str, str], dict[str, int]] = {}
required = {
"sample_id",
"image_path",
"image_sha256",
"language",
"script",
"domain",
"gt_status",
"split",
"claim_eligible",
"source_document",
"source_page",
"source_bbox",
"gt_text",
"prediction",
"codepoint_edits",
"grapheme_edits",
"exact",
}
for row_index, row in enumerate(rows, 1):
missing = sorted(required - set(row))
if missing:
issues.append(f"predictions.tsv row {row_index} missing columns: {missing}")
continue
gt_text = row.get("gt_text") or ""
prediction = row.get("prediction") or ""
prediction_nfc = _nfc(prediction)
gt_nfc = _nfc(gt_text)
cp_edits = Levenshtein.distance(prediction_nfc, gt_nfc)
cp_len = len(gt_nfc)
if cp_len == 0:
zero_length_gt += 1
pred_graph = _graphemes(prediction_nfc)
gt_graph = _graphemes(gt_nfc)
gr_edits = Levenshtein.distance(pred_graph, gt_graph)
gr_len = len(gt_graph)
exact = prediction_nfc == gt_nfc
recorded_cp = _int_text(row.get("codepoint_edits", ""), "codepoint_edits", row_index, issues)
recorded_gr = _int_text(row.get("grapheme_edits", ""), "grapheme_edits", row_index, issues)
recorded_exact = _bool_text(row.get("exact", ""))
sample_id = row.get("sample_id") or f"row {row_index}"
if recorded_cp is not None and recorded_cp != cp_edits:
issues.append(f"predictions.tsv {sample_id} codepoint_edits mismatch: recorded={recorded_cp} recomputed={cp_edits}")
if recorded_gr is not None and recorded_gr != gr_edits:
issues.append(f"predictions.tsv {sample_id} grapheme_edits mismatch: recorded={recorded_gr} recomputed={gr_edits}")
if recorded_exact is None:
issues.append(f"predictions.tsv {sample_id} exact is not boolean-like: {row.get('exact')!r}")
elif recorded_exact != exact:
issues.append(f"predictions.tsv {sample_id} exact mismatch: recorded={recorded_exact} recomputed={exact}")
cp_edits_total += cp_edits
cp_len_total += cp_len
gr_edits_total += gr_edits
gr_len_total += gr_len
exact_total += int(exact)
slice_key = (
row.get("language") or "",
row.get("script") or "",
row.get("domain") or "",
row.get("gt_status") or "",
row.get("split") or "",
)
bucket = slice_totals.setdefault(
slice_key,
{
"n_lines": 0,
"matched_lines": 0,
"codepoint_edits": 0,
"codepoint_ref_len": 0,
"grapheme_edits": 0,
"grapheme_ref_len": 0,
"exact": 0,
},
)
bucket["n_lines"] += 1
bucket["matched_lines"] += 1
bucket["codepoint_edits"] += cp_edits
bucket["codepoint_ref_len"] += cp_len
bucket["grapheme_edits"] += gr_edits
bucket["grapheme_ref_len"] += gr_len
bucket["exact"] += int(exact)
expected_values = {
"n_lines": len(rows),
"matched_lines": len(rows),
"zero_length_gt_lines": zero_length_gt,
"codepoint_edits": cp_edits_total,
"codepoint_ref_len": cp_len_total,
"grapheme_edits": gr_edits_total,
"grapheme_ref_len": gr_len_total,
}
for key, expected in expected_values.items():
if metrics.get(key) != expected:
issues.append(f"metrics {key} does not match recomputed predictions.tsv value: recorded={metrics.get(key)} recomputed={expected}")
if zero_length_gt:
_claim_blocker(issues, eval_run, "recognizer", "zero-length GT rows", "zero-length GT rows present:")
expected_exact_accuracy = exact_total / max(len(rows), 1)
if not _is_number(metrics.get("exact_line_accuracy")) or not _close(float(metrics["exact_line_accuracy"]), expected_exact_accuracy):
issues.append(
"metrics exact_line_accuracy does not match recomputed predictions.tsv value: "
f"recorded={metrics.get('exact_line_accuracy')} recomputed={expected_exact_accuracy}"
)
expected_cp_cer = cp_edits_total / max(cp_len_total, 1)
if not _is_number(metrics.get("cer_codepoint_micro")) or not _close(float(metrics["cer_codepoint_micro"]), expected_cp_cer):
issues.append(
"metrics cer_codepoint_micro does not match recomputed predictions.tsv value: "
f"recorded={metrics.get('cer_codepoint_micro')} recomputed={expected_cp_cer}"
)
expected_gr_cer = gr_edits_total / max(gr_len_total, 1)
if not _is_number(metrics.get("cer_grapheme_micro")) or not _close(float(metrics["cer_grapheme_micro"]), expected_gr_cer):
issues.append(
"metrics cer_grapheme_micro does not match recomputed predictions.tsv value: "
f"recorded={metrics.get('cer_grapheme_micro')} recomputed={expected_gr_cer}"
)
issues.extend(_check_recognizer_slice_metrics(slice_totals, metrics.get("slices")))
return issues
def _check_recognizer_uncertainty(score_dir: Path, metrics: dict[str, Any]) -> list[str]:
units: list[dict[str, int]] = []
issues: list[str] = []
for row_index, row in enumerate(_tsv_rows(score_dir / "predictions.tsv"), 1):
prediction_nfc = _nfc(row.get("prediction") or "")
gt_nfc = _nfc(row.get("gt_text") or "")
pred_graph = _graphemes(prediction_nfc)
gt_graph = _graphemes(gt_nfc)
units.append(
{
"codepoint_edits": Levenshtein.distance(prediction_nfc, gt_nfc),
"codepoint_ref_len": len(gt_nfc),
"grapheme_edits": Levenshtein.distance(pred_graph, gt_graph),
"grapheme_ref_len": len(gt_graph),
"exact": int(row.get("exact") == "True"),
}
)
expected = _expected_uncertainty(
units,
method="deterministic_row_bootstrap",
metrics=("cer_codepoint_micro", "cer_grapheme_micro", "exact_line_accuracy"),
)
issues.extend(_check_uncertainty(metrics.get("uncertainty"), expected, "recognizer"))
for index, row in enumerate(metrics.get("slices") if isinstance(metrics.get("slices"), list) else [], 1):
if not isinstance(row, dict):
continue
slice_key = (
str(row.get("language") or ""),
str(row.get("script") or ""),
str(row.get("domain") or ""),
str(row.get("gt_status") or ""),
str(row.get("split") or ""),
)
slice_units: list[dict[str, int]] = []
for source_row in _tsv_rows(score_dir / "predictions.tsv"):
source_key = (
str(source_row.get("language") or ""),
str(source_row.get("script") or ""),
str(source_row.get("domain") or ""),
str(source_row.get("gt_status") or ""),
str(source_row.get("split") or ""),
)
if source_key != slice_key:
continue
prediction_nfc = _nfc(source_row.get("prediction") or "")
gt_nfc = _nfc(source_row.get("gt_text") or "")
pred_graph = _graphemes(prediction_nfc)
gt_graph = _graphemes(gt_nfc)
slice_units.append(
{
"codepoint_edits": Levenshtein.distance(prediction_nfc, gt_nfc),
"codepoint_ref_len": len(gt_nfc),
"grapheme_edits": Levenshtein.distance(pred_graph, gt_graph),
"grapheme_ref_len": len(gt_graph),
"exact": int(source_row.get("exact") == "True"),
}
)
expected_slice = _expected_uncertainty(
slice_units,
method="deterministic_row_bootstrap",
metrics=("cer_codepoint_micro", "cer_grapheme_micro", "exact_line_accuracy"),
)
issues.extend(_check_uncertainty(row.get("uncertainty"), expected_slice, f"recognizer slices[{index}]"))
return issues
def _check_recognizer_script_inventory(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
inventory_path = Path(eval_run["script_inventory_file"]) if isinstance(eval_run.get("script_inventory_file"), str) and eval_run.get("script_inventory_file") else None
expected, issues = _expected_script_inventory([row.get("gt_text") or "" for row in _tsv_rows(score_dir / "predictions.tsv")], inventory_path)
issues.extend(_check_script_inventory(metrics.get("script_inventory"), expected, "recognizer metrics"))
issues.extend(_check_script_inventory(eval_run.get("script_inventory"), expected, "recognizer eval-run"))
return issues
def _check_recognizer_dictionary_coverage(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
expected = _expected_dictionary_coverage(
[row.get("gt_text") or "" for row in _tsv_rows(score_dir / "predictions.tsv")],
eval_run.get("dictionary"),
)
issues = _check_dictionary_coverage(metrics.get("dictionary_coverage"), expected, "recognizer metrics")
if expected.get("checked") and expected.get("missing_gt_codepoints"):
_claim_blocker(issues, eval_run, "recognizer", "GT codepoints missing from dictionary", "GT codepoints missing from dictionary:")
return issues
def _recognizer_confusions_from_predictions(score_dir: Path, limit: int = 200) -> dict[str, int]:
counts: dict[str, int] = {}
for row in _tsv_rows(score_dir / "predictions.tsv"):
gt = _nfc(row.get("gt_text") or "")
pred = _nfc(row.get("prediction") or "")
for opcode in Levenshtein.opcodes(pred, gt):
tag, i1, i2, j1, j2 = opcode
if tag == "equal":
continue
left = pred[i1:i2] or "<eps>"
right = gt[j1:j2] or "<eps>"
key = f"{left}->{right}"
counts[key] = counts.get(key, 0) + 1
return dict(sorted(counts.items(), key=lambda item: item[1], reverse=True)[:limit])
def _check_recognizer_confusions(score_dir: Path) -> list[str]:
path = score_dir / "confusions.json"
if not path.exists():
return []
try:
payload = _read_json(path)
except (FileNotFoundError, ValueError) as exc:
return [str(exc)]
if not isinstance(payload, dict):
return ["confusions.json must contain a JSON object"]
issues: list[str] = []
recorded: dict[str, int] = {}
for key, value in payload.items():
if isinstance(value, bool) or not isinstance(value, int):
issues.append(f"confusions.json {key!r} count is not an integer: {value!r}")
continue
recorded[str(key)] = value
expected = _recognizer_confusions_from_predictions(score_dir)
if recorded != expected:
issues.append(f"confusions.json does not match recomputed predictions.tsv confusions: recorded={recorded} recomputed={expected}")
return issues
def _check_recognizer_slice_metrics(slice_totals: dict[tuple[str, str, str, str, str], dict[str, int]], slices: Any) -> list[str]:
issues: list[str] = []
if not isinstance(slices, list) or not slices:
return ["recognizer slice metrics are missing"]
recorded_by_key: dict[tuple[str, str, str, str, str], dict[str, Any]] = {}
for index, row in enumerate(slices, 1):
if not isinstance(row, dict):
issues.append(f"recognizer slices[{index}] is not an object")
continue
key = (
str(row.get("language") or ""),
str(row.get("script") or ""),
str(row.get("domain") or ""),
str(row.get("gt_status") or ""),
str(row.get("split") or ""),
)
if key in recorded_by_key:
issues.append(f"duplicate recognizer slice metric for {key}")
recorded_by_key[key] = row
expected_keys = set(slice_totals)
recorded_keys = set(recorded_by_key)
missing = sorted(expected_keys - recorded_keys)
extra = sorted(recorded_keys - expected_keys)
if missing:
issues.append(f"recognizer slice metrics missing slices: {missing[:10]}")
if extra:
issues.append(f"recognizer slice metrics contain extra slices: {extra[:10]}")
for key in sorted(expected_keys & recorded_keys):
totals = slice_totals[key]
recorded = recorded_by_key[key]
expected_numbers: dict[str, float | int] = {
"n_lines": totals["n_lines"],
"matched_lines": totals["matched_lines"],
"codepoint_edits": totals["codepoint_edits"],
"codepoint_ref_len": totals["codepoint_ref_len"],
"cer_codepoint_micro": totals["codepoint_edits"] / max(totals["codepoint_ref_len"], 1),
"grapheme_edits": totals["grapheme_edits"],
"grapheme_ref_len": totals["grapheme_ref_len"],
"cer_grapheme_micro": totals["grapheme_edits"] / max(totals["grapheme_ref_len"], 1),
"exact_line_accuracy": totals["exact"] / max(totals["n_lines"], 1),
}
for field, expected in expected_numbers.items():
value = recorded.get(field)
if isinstance(expected, int):
if value != expected:
issues.append(f"recognizer slice {key} {field} mismatch: recorded={value} recomputed={expected}")
elif not _is_number(value) or not _close(float(value), expected):
issues.append(f"recognizer slice {key} {field} mismatch: recorded={value} recomputed={expected}")
return issues
def _expected_recognizer_worst_rows(score_dir: Path) -> tuple[list[dict[str, Any]], list[str]]:
issues: list[str] = []
expected_rows: list[dict[str, Any]] = []
for row_index, row in enumerate(_tsv_rows(score_dir / "predictions.tsv"), 1):
sample_id = row.get("sample_id") or f"row {row_index}"
gt_text = row.get("gt_text") or ""
prediction = row.get("prediction") or ""
prediction_nfc = _nfc(prediction)
gt_nfc = _nfc(gt_text)
cp_edits = Levenshtein.distance(prediction_nfc, gt_nfc)
cp_len = len(gt_nfc)
pred_graph = _graphemes(prediction_nfc)
gt_graph = _graphemes(gt_nfc)
gr_edits = Levenshtein.distance(pred_graph, gt_graph)
gr_len = len(gt_graph)
expected_rows.append(
{
"sample_id": sample_id,
"image_path": row.get("image_path") or "",
"image_sha256": row.get("image_sha256") or "",
"language": row.get("language") or "",
"script": row.get("script") or "",
"domain": row.get("domain") or "",
"gt_status": row.get("gt_status") or "",
"split": row.get("split") or "",
"claim_eligible": row.get("claim_eligible") or "",
"source_document": row.get("source_document") or "",
"source_page": row.get("source_page") or "",
"source_bbox": row.get("source_bbox") or "",
"gt_text": gt_text,
"prediction": prediction,
"prediction_score": row.get("prediction_score") or "",
"cer_codepoint": cp_edits / max(cp_len, 1),
"cer_grapheme": gr_edits / max(gr_len, 1),
"sort_key": (gr_edits / max(gr_len, 1), gr_edits),
}
)
expected_rows.sort(key=lambda row: row["sort_key"], reverse=True)
return expected_rows, issues
def _check_sample_id_order(label: str, expected_rows: list[dict[str, Any]], recorded_rows: list[dict[str, str]]) -> list[str]:
if expected_rows and not recorded_rows:
return [f"{label} has no review rows"]
if len(recorded_rows) > len(expected_rows):
return [f"{label} has more rows than scored examples: recorded={len(recorded_rows)} scored={len(expected_rows)}"]
expected_ids = [str(row["sample_id"]) for row in expected_rows[: len(recorded_rows)]]
recorded_ids = [row.get("sample_id") or "" for row in recorded_rows]
if recorded_ids != expected_ids:
return [f"{label} sample order does not match recomputed worst examples: recorded={recorded_ids} recomputed={expected_ids}"]
return []
def _check_review_csv_fields(
label: str,
expected_rows: list[dict[str, Any]],
recorded_rows: list[dict[str, str]],
fields: tuple[str, ...],
) -> list[str]:
issues: list[str] = []
expected_by_id = {str(row["sample_id"]): row for row in expected_rows}
for row_index, row in enumerate(recorded_rows, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
expected = expected_by_id.get(sample_id)
if expected is None:
continue
for field in fields:
expected_value = expected.get(field)
if isinstance(expected_value, float):
recorded = _float_text(row.get(field, ""), field, row_index, issues)
if recorded is not None and not _close(recorded, expected_value):
issues.append(f"{label} {sample_id} {field} mismatch: recorded={recorded} recomputed={expected_value}")
elif (row.get(field) or "") != str(expected_value or ""):
issues.append(f"{label} {sample_id} {field} mismatch")
return issues
def _check_review_html_contents(
score_dir: Path,
label: str,
expected_rows: list[dict[str, Any]],
review_rows: list[dict[str, str]],
fields: tuple[str, ...],
) -> list[str]:
path = score_dir / "sample-review.html"
if not path.exists() or not review_rows:
return []
try:
content = path.read_text(encoding="utf-8")
except OSError as exc:
return [f"sample-review.html could not be read: {exc}"]
issues: list[str] = []
for row in expected_rows[: len(review_rows)]:
sample_id = str(row["sample_id"])
missing_fields: list[str] = []
for field in fields:
value = str(row.get(field) or "")
if value and html.escape(value) not in content:
missing_fields.append(field)
if missing_fields:
issues.append(f"sample-review.html missing {label} review sample {sample_id} fields: {missing_fields}")
return issues
def _check_recognizer_worst_artifacts(score_dir: Path) -> list[str]:
issues: list[str] = []
expected_rows, row_issues = _expected_recognizer_worst_rows(score_dir)
issues.extend(row_issues)
worst_rows = _csv_rows(score_dir / "worst-lines.csv")
review_rows = _csv_rows(score_dir / "sample-review.csv")
issues.extend(_check_sample_id_order("worst-lines.csv", expected_rows, worst_rows))
issues.extend(_check_sample_id_order("sample-review.csv", expected_rows, review_rows))
issues.extend(_check_review_csv_fields("sample-review.csv", expected_rows, review_rows, ("image_path", "gt_text", "prediction", "prediction_score")))
issues.extend(
_check_review_html_contents(
score_dir,
"recognizer",
expected_rows,
review_rows,
("sample_id", "image_path", "gt_text", "prediction"),
)
)
expected_by_id = {str(row["sample_id"]): row for row in expected_rows}
for row_index, row in enumerate(worst_rows, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
expected = expected_by_id.get(sample_id)
if expected is None:
continue
for field in ("cer_grapheme", "cer_codepoint"):
recorded = _float_text(row.get(field, ""), field, row_index, issues)
if recorded is not None and not _close(recorded, float(expected[field])):
issues.append(f"worst-lines.csv {sample_id} {field} mismatch: recorded={recorded} recomputed={expected[field]}")
for field in (
"image_path",
"image_sha256",
"language",
"script",
"domain",
"gt_status",
"split",
"claim_eligible",
"source_document",
"source_page",
"source_bbox",
"gt_text",
"prediction",
"prediction_score",
):
if (row.get(field) or "") != str(expected[field]):
issues.append(f"worst-lines.csv {sample_id} {field} mismatch")
return issues
def _check_recognizer_output_duplicates(score_dir: Path) -> list[str]:
issues: list[str] = []
predictions = _tsv_rows(score_dir / "predictions.tsv")
worst_rows = _csv_rows(score_dir / "worst-lines.csv")
review_rows = _csv_rows(score_dir / "sample-review.csv")
unmatched_rows = _tsv_rows(score_dir / "unmatched.tsv")
issues.extend(_check_unique_field("predictions.tsv", predictions, "sample_id"))
issues.extend(_check_unique_field("predictions.tsv", predictions, "image_path"))
issues.extend(_check_unique_field("worst-lines.csv", worst_rows, "sample_id"))
issues.extend(_check_unique_field("sample-review.csv", review_rows, "sample_id"))
issues.extend(_check_unique_key("unmatched.tsv", unmatched_rows, ("kind", "image_path")))
return issues
def _read_recognizer_label_source(path: Path) -> tuple[list[dict[str, str]], list[str]]:
issues: list[str] = []
rows: list[dict[str, str]] = []
seen: set[str] = set()
if not path.exists():
return rows, [f"recognizer label source does not exist for unmatched audit: {path}"]
with path.open("r", encoding="utf-8", newline="") as handle:
for line_no, line in enumerate(handle, 1):
line = line.rstrip("\n")
if not line:
continue
if "\t" not in line:
issues.append(f"recognizer label source {path}:{line_no} is not tab-separated")
continue
image_path, text = line.split("\t", 1)
image_key = Path(image_path).name
if image_key in seen:
issues.append(f"recognizer label source duplicate image basename: {image_key}")
seen.add(image_key)
rows.append({"image_key": image_key, "image_path": image_path, "text": text})
return rows, issues
def _read_recognizer_prediction_source(path: Path) -> tuple[list[dict[str, str]], list[str]]:
issues: list[str] = []
rows: list[dict[str, str]] = []
seen: set[str] = set()
if not path.exists():
return rows, [f"recognizer prediction source does not exist for unmatched audit: {path}"]
with path.open("r", encoding="utf-8", newline="") as handle:
for line_no, line in enumerate(handle, 1):
line = line.rstrip("\n")
if not line:
continue
parts = line.split("\t")
if len(parts) < 2:
issues.append(f"recognizer prediction source {path}:{line_no} is not tab-separated")
continue
image_path = parts[0]
image_key = Path(image_path).name
if image_key in seen:
issues.append(f"recognizer prediction source duplicate image basename: {image_key}")
seen.add(image_key)
rows.append({"image_key": image_key, "image_path": image_path, "text": parts[1], "score": parts[2] if len(parts) > 2 else ""})
return rows, issues
def _read_recognizer_manifest_source(path_value: Any) -> tuple[dict[str, dict[str, Any]], list[str]]:
if not isinstance(path_value, str) or not path_value:
return {}, []
path = Path(path_value)
if not path.exists():
return {}, [f"recognizer manifest source does not exist for predictions.tsv audit: {path}"]
rows: dict[str, dict[str, Any]] = {}
seen_ids: set[str] = set()
issues: list[str] = []
with path.open("r", encoding="utf-8") as handle:
for line_no, line in enumerate(handle, 1):
line = line.strip()
if not line:
continue
try:
payload = json.loads(line)
except json.JSONDecodeError as exc:
issues.append(f"recognizer manifest source {path}:{line_no} invalid JSON: {exc}")
continue
if not isinstance(payload, dict):
issues.append(f"recognizer manifest source {path}:{line_no} row is not an object")
continue
sample_id = str(payload.get("sample_id") or "")
if sample_id:
if sample_id in seen_ids:
issues.append(f"recognizer manifest source duplicate sample_id: {sample_id}")
seen_ids.add(sample_id)
rows[sample_id] = payload
image_path = str(payload.get("image_path") or payload.get("image") or "")
if image_path:
rows[image_path] = payload
rows[Path(image_path).name] = payload
return rows, issues
def _recognizer_metadata_for(label: dict[str, str], manifest: dict[str, dict[str, Any]]) -> dict[str, Any]:
payload = manifest.get(label["image_path"]) or manifest.get(label["image_key"]) or {}
metadata = dict(payload.get("metadata") or {})
for key, value in payload.items():
if key == "metadata":
continue
if key not in metadata:
metadata[key] = value
for key in ("sample_id", "dataset", "split", "image_path", "sha256"):
if key in payload and key not in metadata:
metadata[key] = payload[key]
return metadata
def _recognizer_eval_run_fallback(eval_run: dict[str, Any], key: str) -> str:
values = eval_run.get(key)
if isinstance(values, list) and len(values) == 1:
return str(values[0])
return ""
def _recognizer_domain(metadata: dict[str, Any], fallback: str) -> str:
for key in ("domain", "eval_domain"):
if metadata.get(key):
return str(metadata[key])
slices = metadata.get("slices")
if isinstance(slices, list):
if "synthetic" in slices:
return "synthetic"
if "real" in slices:
return fallback
return fallback
def _recognizer_gt_status(metadata: dict[str, Any], fallback: str) -> str:
for key in ("gt_status", "label_status"):
if metadata.get(key):
return str(metadata[key])
provenance = str(metadata.get("label_provenance") or "").lower()
if "human" in provenance and "pending" not in provenance:
return "human_verified"
return fallback
def _expected_recognizer_prediction_rows(eval_run: dict[str, Any]) -> tuple[list[dict[str, str]], list[str]]:
label_file = eval_run.get("label_file")
prediction_file = eval_run.get("prediction_file")
data_dir = eval_run.get("data_dir")
if not isinstance(label_file, str) or not label_file:
return [], ["eval-run.json missing label_file for predictions.tsv source audit"]
if not isinstance(prediction_file, str) or not prediction_file:
return [], ["eval-run.json missing prediction_file for predictions.tsv source audit"]
if not isinstance(data_dir, str) or not data_dir:
return [], ["eval-run.json missing data_dir for predictions.tsv source audit"]
labels, label_issues = _read_recognizer_label_source(Path(label_file))
predictions, prediction_issues = _read_recognizer_prediction_source(Path(prediction_file))
manifest, manifest_issues = _read_recognizer_manifest_source(eval_run.get("manifest_file"))
issues = [*label_issues, *prediction_issues, *manifest_issues]
if issues:
return [], issues
prediction_by_key = {row["image_key"]: row for row in predictions}
language_fallback = _recognizer_eval_run_fallback(eval_run, "languages")
script_fallback = _recognizer_eval_run_fallback(eval_run, "scripts")
domain_fallback = _recognizer_eval_run_fallback(eval_run, "domains")
gt_status_fallback = _recognizer_eval_run_fallback(eval_run, "gt_statuses")
split_fallback = _recognizer_eval_run_fallback(eval_run, "splits")
expected: list[dict[str, str]] = []
seen_sample_ids: set[str] = set()
for index, label in enumerate(labels):
prediction = prediction_by_key.get(label["image_key"])
if prediction is None:
continue
metadata = _recognizer_metadata_for(label, manifest)
sample_id = str(metadata.get("sample_id") or f"row-{index:06d}")
if sample_id in seen_sample_ids:
issues.append(f"recognizer source rows duplicate resolved sample_id: {sample_id}")
seen_sample_ids.add(sample_id)
image_path = Path(data_dir) / label["image_path"]
image_sha256 = _sha256_file(image_path) if image_path.exists() else ""
declared_image_sha256 = str(metadata.get("sha256") or metadata.get("image_sha256") or "")
if declared_image_sha256 and declared_image_sha256 != image_sha256:
issues.append(
f"recognizer manifest {sample_id} sha256 mismatch: "
f"recorded={declared_image_sha256} actual={image_sha256}"
)
text_sha256 = _audit_hash_text(label["text"])
declared_text_sha256 = str(metadata.get("text_sha256") or "")
if declared_text_sha256 and declared_text_sha256 != text_sha256:
issues.append(
f"recognizer manifest {sample_id} text_sha256 mismatch: "
f"recorded={declared_text_sha256} actual={text_sha256}"
)
declared_sample_sha256 = str(metadata.get("sample_sha256") or "")
if declared_sample_sha256:
actual_sample_sha256 = _audit_sample_hash(image_sha256, text_sha256)
if declared_sample_sha256 != actual_sample_sha256:
issues.append(
f"recognizer manifest {sample_id} sample_sha256 mismatch: "
f"recorded={declared_sample_sha256} actual={actual_sample_sha256}"
)
expected.append(
{
"sample_id": sample_id,
"image_path": label["image_path"],
"image_sha256": image_sha256,
"language": str(metadata.get("language") or language_fallback),
"script": str(metadata.get("script") or script_fallback),
"domain": _recognizer_domain(metadata, domain_fallback),
"gt_status": _recognizer_gt_status(metadata, gt_status_fallback),
"split": str(metadata.get("split") or metadata.get("split_role") or split_fallback),
"claim_eligible": str(_metadata_bool_value(metadata, CLAIM_ELIGIBLE_KEYS, eval_run.get("claim_eligible_default"))),
"source_document": _metadata_value_for_keys(metadata, SOURCE_DOCUMENT_KEYS),
"source_page": _metadata_value_for_keys(metadata, PAGE_KEYS),
"source_bbox": _metadata_value_for_keys(metadata, BBOX_KEYS),
"gt_text": label["text"],
"prediction": prediction["text"],
"prediction_score": prediction["score"],
}
)
return expected, issues
def _recognizer_pack_content_digest(rows: list[dict[str, str]]) -> str:
canonical_rows = [
{
"sample_id": row.get("sample_id") or "",
"image_sha256": row.get("image_sha256") or "",
"gt_text_sha256_nfc": _audit_hash_text(row.get("gt_text") or ""),
"language": row.get("language") or "",
"script": row.get("script") or "",
"domain": row.get("domain") or "",
"gt_status": row.get("gt_status") or "",
"split": row.get("split") or "",
"claim_eligible": _bool_text(row.get("claim_eligible") or ""),
}
for row in rows
]
payload = "\n".join(
json.dumps(row, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
for row in sorted(canonical_rows, key=lambda item: item["sample_id"])
)
return hashlib.sha256(payload.encode("utf-8")).hexdigest()
def _summary_pack_content_sha256(summary: dict[str, Any]) -> str:
value = summary.get("pack_content_sha256")
if isinstance(value, str) and value:
return value
eval_pack = summary.get("eval_pack")
if isinstance(eval_pack, dict):
nested = eval_pack.get("pack_content_sha256")
if isinstance(nested, str) and nested:
return nested
return ""
def _check_recognizer_pack_content_digest(eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
expected_rows, issues = _expected_recognizer_prediction_rows(eval_run)
if issues:
return issues
expected = _recognizer_pack_content_digest(expected_rows)
for label, payload in (("eval-run.json", eval_run), ("metrics.json", metrics)):
recorded = payload.get("pack_content_sha256")
if recorded != expected:
issues.append(f"{label} pack_content_sha256 mismatch: recorded={recorded!r} recomputed={expected}")
summary, summary_issues = _read_summary_for_audit(eval_run, "recognizer", "pack content digest audit")
issues.extend(summary_issues)
if summary:
summary_digest = _summary_pack_content_sha256(summary)
if not summary_digest:
issues.append("recognizer summary missing pack_content_sha256")
elif summary_digest != expected:
issues.append(f"recognizer summary pack_content_sha256 mismatch: recorded={summary_digest} recomputed={expected}")
return issues
def _check_recognizer_prediction_source_rows(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
recorded = _tsv_rows(score_dir / "predictions.tsv")
expected, issues = _expected_recognizer_prediction_rows(eval_run)
if issues:
return issues
if len(recorded) != len(expected):
issues.append(f"predictions.tsv row count does not match source matched rows: recorded={len(recorded)} recomputed={len(expected)}")
for index in range(min(len(recorded), len(expected))):
recorded_row = recorded[index]
expected_row = expected[index]
sample_id = recorded_row.get("sample_id") or expected_row["sample_id"] or f"row {index + 1}"
for field in (
"sample_id",
"image_path",
"image_sha256",
"language",
"script",
"domain",
"gt_status",
"split",
"claim_eligible",
"source_document",
"source_page",
"source_bbox",
"gt_text",
"prediction",
"prediction_score",
):
if (recorded_row.get(field) or "") != expected_row[field]:
issues.append(f"predictions.tsv {sample_id} {field} does not match source files")
return issues
def _check_recognizer_metadata_contradictions(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
manifest, manifest_issues = _read_recognizer_manifest_source(eval_run.get("manifest_file"))
if manifest_issues:
return manifest_issues
expected: list[str] = []
for row in _tsv_rows(score_dir / "predictions.tsv"):
image_path = row.get("image_path") or ""
label = {"image_path": image_path, "image_key": Path(image_path).name}
metadata = _recognizer_metadata_for(label, manifest)
sample_id = row.get("sample_id") or str(metadata.get("sample_id") or "")
expected.extend(
_metadata_contradictions_for(
sample_id,
metadata,
row.get("domain") or "",
row.get("gt_status") or "",
)
)
recorded = metrics.get("metadata_contradictions")
if recorded != expected:
return [f"recognizer metadata_contradictions do not match source metadata: recorded={recorded} recomputed={expected}"]
return []
def _check_recognizer_provenance_coverage(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
manifest, manifest_issues = _read_recognizer_manifest_source(eval_run.get("manifest_file"))
if manifest_issues:
return manifest_issues
rows = _tsv_rows(score_dir / "predictions.tsv")
expected = {
"missing_source_document": 0,
"missing_page": 0,
"missing_bbox": 0,
"missing_claim_eligibility": 0,
"missing_human_review_evidence": 0,
}
for row in rows:
image_path = row.get("image_path") or ""
label = {"image_path": image_path, "image_key": Path(image_path).name}
metadata = _recognizer_metadata_for(label, manifest)
if not _metadata_has_any(metadata, SOURCE_DOCUMENT_KEYS):
expected["missing_source_document"] += 1
if not _metadata_has_any(metadata, PAGE_KEYS):
expected["missing_page"] += 1
if not _metadata_has_any(metadata, BBOX_KEYS):
expected["missing_bbox"] += 1
if not any(key in metadata for key in CLAIM_ELIGIBLE_KEYS):
expected["missing_claim_eligibility"] += 1
if _recognizer_gt_status(metadata, "") == "human_verified" and not _metadata_has_any(metadata, HUMAN_REVIEW_EVIDENCE_KEYS):
expected["missing_human_review_evidence"] += 1
recorded = metrics.get("provenance_coverage")
if recorded != expected:
return [f"recognizer provenance_coverage does not match source metadata: recorded={recorded} recomputed={expected}"]
return []
def _recognizer_eval_leakage_rows(score_dir: Path) -> list[dict[str, str]]:
rows: list[dict[str, str]] = []
for row in _tsv_rows(score_dir / "predictions.tsv"):
image_sha256 = row.get("image_sha256") or ""
text_sha256 = _audit_hash_text(row.get("gt_text") or "")
rows.append(
{
"sample_id": row.get("sample_id") or "",
"source_sample_id": row.get("source_sample_id") or row.get("sample_id") or "",
"image_path": row.get("image_path") or "",
"image_sha256": image_sha256,
"text_sha256": text_sha256,
"sample_sha256": _audit_sample_hash(image_sha256, text_sha256) if image_sha256 and text_sha256 else "",
}
)
return rows
def _check_recognizer_leakage_report(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
train_paths, issues = _train_manifest_paths(eval_run)
train_rows: list[dict[str, str]] = []
for path in train_paths:
rows, row_issues = _audit_manifest_leakage_rows(path, compute_missing_image_sha=False)
train_rows.extend(rows)
issues.extend(row_issues)
if issues:
return issues
eval_rows = _recognizer_eval_leakage_rows(score_dir) if train_rows else []
expected = {
"train_manifests": [str(path) for path in train_paths],
"checked": bool(train_rows),
"train_rows": len(train_rows),
"eval_rows": len(eval_rows),
"overlaps": _leakage_overlaps(train_rows, eval_rows),
}
return _compare_leakage_report("recognizer", metrics.get("leakage"), expected)
def _check_recognizer_unmatched_artifact(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
path = score_dir / "unmatched.tsv"
if not path.exists():
return []
label_file = eval_run.get("label_file")
prediction_file = eval_run.get("prediction_file")
data_dir = eval_run.get("data_dir")
if not isinstance(label_file, str) or not label_file:
return ["eval-run.json missing label_file for unmatched.tsv audit"]
if not isinstance(prediction_file, str) or not prediction_file:
return ["eval-run.json missing prediction_file for unmatched.tsv audit"]
if not isinstance(data_dir, str) or not data_dir:
return ["eval-run.json missing data_dir for unmatched.tsv audit"]
labels, label_issues = _read_recognizer_label_source(Path(label_file))
predictions, prediction_issues = _read_recognizer_prediction_source(Path(prediction_file))
issues = [*label_issues, *prediction_issues]
if issues:
return issues
prediction_by_key = {row["image_key"]: row for row in predictions}
label_keys = {row["image_key"] for row in labels}
expected = [
{
"kind": "label_without_prediction",
"image_key": row["image_key"],
"image_path": row["image_path"],
"image_sha256": _sha256_file(Path(data_dir) / row["image_path"]) if (Path(data_dir) / row["image_path"]).exists() else "",
"text": row["text"],
"prediction_score": "",
}
for row in labels
if row["image_key"] not in prediction_by_key
]
expected.extend(
{
"kind": "prediction_without_label",
"image_key": row["image_key"],
"image_path": row["image_path"],
"image_sha256": "",
"text": row["text"],
"prediction_score": row["score"],
}
for row in predictions
if row["image_key"] not in label_keys
)
fields = ["kind", "image_key", "image_path", "image_sha256", "text", "prediction_score"]
recorded = [{field: row.get(field) or "" for field in fields} for row in _tsv_rows(path)]
if recorded != expected:
issues.append(
"unmatched.tsv rows do not match recomputed label/prediction misses: "
f"recorded={_short_rows(recorded)} recomputed={_short_rows(expected)}"
)
return issues
def _check_recognizer_match_count_metrics(metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
label_file = eval_run.get("label_file")
prediction_file = eval_run.get("prediction_file")
if not isinstance(label_file, str) or not label_file:
return ["eval-run.json missing label_file for recognizer match-count audit"]
if not isinstance(prediction_file, str) or not prediction_file:
return ["eval-run.json missing prediction_file for recognizer match-count audit"]
labels, label_issues = _read_recognizer_label_source(Path(label_file))
predictions, prediction_issues = _read_recognizer_prediction_source(Path(prediction_file))
issues = [*label_issues, *prediction_issues]
if issues:
return issues
label_keys = {row["image_key"] for row in labels}
prediction_keys = {row["image_key"] for row in predictions}
matched = len(label_keys & prediction_keys)
expected_counts = {
"label_lines": len(labels),
"n_lines": matched,
"matched_lines": matched,
"unmatched_labels": len(label_keys - prediction_keys),
"unmatched_predictions": len(prediction_keys - label_keys),
}
for field, expected in expected_counts.items():
recorded = metrics.get(field)
if not _is_number(recorded):
issues.append(f"metrics missing numeric field: {field}")
elif float(recorded) != float(expected):
issues.append(
f"metrics {field} does not match recomputed recognizer source value: "
f"recorded={recorded} recomputed={expected}"
)
expected_match_rate = matched / max(len(labels), 1)
recorded_match_rate = metrics.get("prediction_match_rate")
if not _is_number(recorded_match_rate):
issues.append("metrics missing numeric field: prediction_match_rate")
elif not _close(float(recorded_match_rate), expected_match_rate):
issues.append(
"metrics prediction_match_rate does not match recomputed recognizer source value: "
f"recorded={recorded_match_rate} recomputed={expected_match_rate}"
)
return issues
def _check_recognizer_row_hashes(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
data_dir = eval_run.get("data_dir")
if not isinstance(data_dir, str) or not data_dir:
return ["recognizer data_dir is missing for row hash verification"]
for row_index, row in enumerate(_tsv_rows(score_dir / "predictions.tsv"), 1):
sample_id = row.get("sample_id") or f"row {row_index}"
image_path = row.get("image_path")
expected = row.get("image_sha256")
if not image_path:
issues.append(f"predictions.tsv {sample_id} missing image_path")
continue
if not expected:
issues.append(f"predictions.tsv {sample_id} missing image_sha256")
continue
path = Path(data_dir) / image_path
if not path.exists():
issues.append(f"predictions.tsv {sample_id} image path does not exist: {path}")
continue
actual = _sha256_file(path)
if actual != expected:
issues.append(f"predictions.tsv {sample_id} image_sha256 mismatch: recorded={expected} actual={actual}")
return issues
def _detection_values(predicted_count: int, reference_count: int, matches: int) -> dict[str, float | int]:
precision = matches / predicted_count if predicted_count else (1.0 if reference_count == 0 else 0.0)
recall = matches / reference_count if reference_count else (1.0 if predicted_count == 0 else 0.0)
f1 = 2 * precision * recall / (precision + recall) if precision + recall else 0.0
return {
"line_reference_count": reference_count,
"line_prediction_count": predicted_count,
"line_match_count": matches,
"line_detection_precision": precision,
"line_detection_recall": recall,
"line_detection_f1": f1,
}
def _audit_reference_payload(input_path: Path, metadata: dict[str, Any], manifest_dir: Path) -> tuple[dict[str, Any] | None, list[str]]:
issues: list[str] = []
explicit = metadata.get("reference_path")
candidates: list[Path] = []
if isinstance(explicit, str) and explicit:
candidates.append(_resolve_audit_manifest_path(explicit, manifest_dir))
candidates.extend([input_path.with_suffix(input_path.suffix + ".ref.json"), input_path.with_suffix(".ref.json")])
for candidate in candidates:
if not candidate.exists() or candidate.suffix.lower() != ".json":
continue
try:
payload = json.loads(candidate.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
issues.append(f"reference JSON is invalid for line-alignment audit: {candidate}: {exc}")
return None, issues
if not isinstance(payload, dict):
issues.append(f"reference JSON must contain an object for line-alignment audit: {candidate}")
return None, issues
return payload, issues
return None, issues
def _audit_reference_lines(payload: dict[str, Any] | None) -> list[dict[str, Any]]:
if payload is None:
return []
candidates = payload.get("reading_order")
if candidates is None:
candidates = payload.get("lines")
if not isinstance(candidates, list):
return []
rows: list[dict[str, Any]] = []
for index, item in enumerate(candidates):
if not isinstance(item, dict):
continue
raw_bbox = item.get("bbox")
if raw_bbox is None:
continue
try:
bbox = BBox.from_any(raw_bbox)
except Exception:
continue
if bbox.w <= 0 or bbox.h <= 0:
continue
rows.append(
{
"line_id": str(item.get("id") or item.get("line_id") or f"ref-{index:04d}"),
"text": str(item.get("text") or ""),
"bbox": bbox,
"page_index": int(item.get("page_index", 0)),
}
)
return rows
def _audit_predicted_lines(document_path: Path) -> tuple[list[TextLine], list[str]]:
try:
payload = json.loads(document_path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return [], [f"document JSON is invalid for line-alignment audit: {document_path}: {exc}"]
if not isinstance(payload, dict):
return [], [f"document JSON must contain an object for line-alignment audit: {document_path}"]
try:
document = Document.from_dict(payload)
except Exception as exc:
return [], [f"document JSON failed schema parse for line-alignment audit: {document_path}: {exc}"]
predicted: list[TextLine] = []
for page in document.pages:
predicted.extend(page.text_lines)
return predicted, []
def _codepoint_cer(prediction: str, reference: str) -> float:
pred_nfc = _nfc(prediction)
ref_nfc = _nfc(reference)
return Levenshtein.distance(pred_nfc, ref_nfc) / max(len(ref_nfc), 1)
def _text_micro_counts(prediction: str, reference: str) -> tuple[int, int, int, int]:
pred_nfc = _nfc(prediction)
ref_nfc = _nfc(reference)
pred_graph = _graphemes(pred_nfc)
ref_graph = _graphemes(ref_nfc)
return (
Levenshtein.distance(pred_nfc, ref_nfc),
len(ref_nfc),
Levenshtein.distance(pred_graph, ref_graph),
len(ref_graph),
)
def _audit_document_text(document_path: Path) -> tuple[str | None, list[str]]:
try:
payload = json.loads(document_path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return None, [f"document JSON is invalid for text metric audit: {document_path}: {exc}"]
if not isinstance(payload, dict):
return None, [f"document JSON must contain an object for text metric audit: {document_path}"]
try:
document = Document.from_dict(payload)
except Exception as exc:
return None, [f"document JSON failed schema parse for text metric audit: {document_path}: {exc}"]
return document.text, []
def _audit_document_source_path(document_path: Path) -> tuple[str | None, list[str]]:
try:
payload = json.loads(document_path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return None, [f"document JSON is invalid for source-path audit: {document_path}: {exc}"]
if not isinstance(payload, dict):
return None, [f"document JSON must contain an object for source-path audit: {document_path}"]
try:
document = Document.from_dict(payload)
except Exception as exc:
return None, [f"document JSON failed schema parse for source-path audit: {document_path}: {exc}"]
return document.source_path, []
def _audit_reference_text(input_path: Path, metadata: dict[str, Any], manifest_dir: Path) -> tuple[str | None, list[str]]:
explicit = metadata.get("reference_path")
explicit_path = _resolve_audit_manifest_path(explicit, manifest_dir) if isinstance(explicit, str) and explicit else None
try:
reference = load_reference(input_path, explicit_path=explicit_path)
except Exception as exc:
return None, [f"reference failed to parse for text metric audit: {input_path}: {exc}"]
if reference is None:
return None, [f"reference is missing for text metric audit: {input_path}"]
return reference.text, []
def _audit_reference_path(input_path: Path, metadata: dict[str, Any], manifest_dir: Path) -> Path | None:
explicit = metadata.get("reference_path")
candidates: list[Path] = []
if isinstance(explicit, str) and explicit:
candidates.append(_resolve_audit_manifest_path(explicit, manifest_dir))
candidates.extend(
[
input_path.with_suffix(input_path.suffix + ".ref.json"),
input_path.with_suffix(".ref.json"),
input_path.with_suffix(input_path.suffix + ".ref.txt"),
input_path.with_suffix(".ref.txt"),
]
)
for candidate in candidates:
if candidate.exists():
return candidate
return None
def _text_sha256(text: str) -> str:
return hashlib.sha256(text.encode("utf-8")).hexdigest()
def _expected_page_reference_artifact_rows(eval_run: dict[str, Any]) -> tuple[list[dict[str, str]], list[str]]:
issues: list[str] = []
eval_manifest = eval_run.get("eval_manifest")
if not isinstance(eval_manifest, str) or not eval_manifest:
return [], ["page-eval-run.json missing eval_manifest for reference-artifacts.tsv audit"]
manifest_path = Path(eval_manifest)
manifest_rows, manifest_issues = _read_page_manifest_source(manifest_path)
issues.extend(manifest_issues)
manifest_dir = manifest_path.parent
expected: list[dict[str, str]] = []
for manifest_row in manifest_rows:
sample_id = str(manifest_row["sample_id"])
metadata = manifest_row["metadata"]
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
reference_path = _audit_reference_path(input_path, metadata, manifest_dir)
if reference_path is None:
continue
try:
reference = load_reference(input_path, explicit_path=reference_path)
except Exception as exc:
issues.append(f"reference failed to parse for reference-artifacts.tsv audit: {input_path}: {exc}")
continue
if reference is None:
continue
fields, field_issues = _audit_reference_fields(input_path, metadata, manifest_dir)
issues.extend(field_issues)
if field_issues:
continue
text = reference.text or ""
claim_value = (reference.metadata or {}).get("claim_evidence_eligible")
expected.append(
{
"sample_id": sample_id,
"input_path": str(input_path),
"reference_path": str(reference_path),
"reference_sha256": _sha256_file(reference_path),
"reference_fields": ";".join(fields),
"reference_text_sha256": _text_sha256(text),
"reference_status": str((reference.metadata or {}).get("reference_status") or ""),
"reference_claim_eligible": str(claim_value) if claim_value is not None else "",
}
)
return expected, issues
def _check_page_reference_artifacts(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
recorded = _tsv_rows(score_dir / "reference-artifacts.tsv")
expected, expected_issues = _expected_page_reference_artifact_rows(eval_run)
issues.extend(expected_issues)
if not recorded:
return [*issues, "reference-artifacts.tsv has no scored reference rows"]
recorded_by_id: dict[str, dict[str, str]] = {}
for row_index, row in enumerate(recorded, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
if sample_id not in recorded_by_id:
recorded_by_id[sample_id] = row
expected_by_id = {row["sample_id"]: row for row in expected}
missing = sorted(set(expected_by_id) - set(recorded_by_id))
extra = sorted(set(recorded_by_id) - set(expected_by_id))
if missing:
issues.append(f"reference-artifacts.tsv missing scored reference rows: {missing[:10]}")
if extra:
issues.append(f"reference-artifacts.tsv has rows not backed by scored references: {extra[:10]}")
issues.extend(_check_unique_field("reference-artifacts.tsv", recorded, "sample_id"))
required = {
"sample_id",
"input_path",
"reference_path",
"reference_sha256",
"reference_fields",
"reference_text_sha256",
"reference_status",
"reference_claim_eligible",
}
for row_index, row in enumerate(recorded, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
missing_fields = sorted(required - set(row))
if missing_fields:
issues.append(f"reference-artifacts.tsv {sample_id} missing columns: {missing_fields}")
continue
expected_row = expected_by_id.get(sample_id)
if expected_row is None:
continue
for field in ("input_path", "reference_path"):
if not _same_path(row.get(field), Path(expected_row[field])):
issues.append(
f"reference-artifacts.tsv {sample_id} {field} does not match reference source: "
f"recorded={row.get(field) or ''} recomputed={expected_row[field]}"
)
for field in ("reference_sha256", "reference_fields", "reference_text_sha256", "reference_status", "reference_claim_eligible"):
if (row.get(field) or "") != expected_row[field]:
issues.append(
f"reference-artifacts.tsv {sample_id} {field} does not match reference source: "
f"recorded={row.get(field) or ''} recomputed={expected_row[field]}"
)
return issues
def _page_pack_content_digest(metric_rows: list[dict[str, str]], reference_rows: list[dict[str, str]]) -> str:
references_by_id = {row["sample_id"]: row for row in reference_rows}
canonical_rows: list[dict[str, Any]] = []
for row in metric_rows:
reference = references_by_id.get(row["sample_id"], {})
canonical_rows.append(
{
"sample_id": row.get("sample_id") or "",
"input_sha256": row.get("input_sha256") or "",
"reference_sha256": reference.get("reference_sha256") or "",
"reference_text_sha256": reference.get("reference_text_sha256") or "",
"language": row.get("language") or "",
"script": row.get("script") or "",
"domain": row.get("domain") or "",
"gt_status": row.get("gt_status") or "",
"split": row.get("split") or "",
"claim_eligible": _bool_text(row.get("claim_eligible") or ""),
"source_document": row.get("source_document") or "",
"source_page": row.get("source_page") or "",
"source_bbox": row.get("source_bbox") or "",
"reference_fields": reference.get("reference_fields") or row.get("reference_fields") or "",
"reference_status": reference.get("reference_status") or "",
"reference_claim_eligible": reference.get("reference_claim_eligible") or "",
}
)
payload = "\n".join(
json.dumps(row, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
for row in sorted(canonical_rows, key=lambda item: item["sample_id"])
)
return hashlib.sha256(payload.encode("utf-8")).hexdigest()
def _check_page_pack_content_digest(eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
metric_rows, metric_issues = _expected_page_metric_identity_rows(eval_run)
reference_rows, reference_issues = _expected_page_reference_artifact_rows(eval_run)
issues = [*metric_issues, *reference_issues]
if issues:
return issues
expected = _page_pack_content_digest(metric_rows, reference_rows)
for label, payload in (("page-eval-run.json", eval_run), ("metrics.json", metrics)):
recorded = payload.get("pack_content_sha256")
if recorded != expected:
issues.append(f"{label} pack_content_sha256 mismatch: recorded={recorded!r} recomputed={expected}")
summary, summary_issues = _read_summary_for_audit(eval_run, "page", "pack content digest audit")
issues.extend(summary_issues)
if summary:
summary_digest = _summary_pack_content_sha256(summary)
if not summary_digest:
issues.append("page summary missing pack_content_sha256")
elif summary_digest != expected:
issues.append(f"page summary pack_content_sha256 mismatch: recorded={summary_digest} recomputed={expected}")
return issues
def _check_page_text_pairs_artifact(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
page_rows = _csv_rows(score_dir / "page-metrics.csv")
text_rows = _tsv_rows(score_dir / "text-pairs.tsv")
if not page_rows:
return ["page-metrics.csv has no scored rows for text-pairs.tsv audit"]
if not text_rows:
return ["text-pairs.tsv has no scored rows"]
page_rows_by_id = {row.get("sample_id") or "": row for row in page_rows}
text_rows_by_id = {row.get("sample_id") or "": row for row in text_rows}
missing = sorted(set(page_rows_by_id) - set(text_rows_by_id))
extra = sorted(set(text_rows_by_id) - set(page_rows_by_id))
if missing:
issues.append(f"text-pairs.tsv missing scored page rows: {missing[:10]}")
if extra:
issues.append(f"text-pairs.tsv has rows not present in page-metrics.csv: {extra[:10]}")
issues.extend(_check_unique_field("text-pairs.tsv", text_rows, "sample_id"))
manifest_rows_by_id: dict[str, dict[str, Any]] = {}
manifest_dir = Path("")
eval_manifest = eval_run.get("eval_manifest")
if isinstance(eval_manifest, str) and eval_manifest:
manifest_path = Path(eval_manifest)
manifest_dir = manifest_path.parent
manifest_rows, manifest_issues = _read_page_manifest_source(manifest_path)
issues.extend(manifest_issues)
manifest_rows_by_id = {str(row["sample_id"]): row for row in manifest_rows}
else:
issues.append("page-eval-run.json missing eval_manifest for text-pairs.tsv audit")
required = {
"sample_id",
"input_path",
"document_path",
"prediction_text_sha256",
"reference_text_sha256",
"prediction_text",
"reference_text",
"text_codepoint_edits",
"text_codepoint_ref_len",
"text_grapheme_edits",
"text_grapheme_ref_len",
"text_exact",
}
for row_index, row in enumerate(text_rows, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
missing_fields = sorted(required - set(row))
if missing_fields:
issues.append(f"text-pairs.tsv {sample_id} missing columns: {missing_fields}")
continue
prediction_text = row.get("prediction_text") or ""
reference_text = row.get("reference_text") or ""
if row.get("prediction_text_sha256") != _text_sha256(prediction_text):
issues.append(f"text-pairs.tsv {sample_id} prediction_text_sha256 mismatch")
if row.get("reference_text_sha256") != _text_sha256(reference_text):
issues.append(f"text-pairs.tsv {sample_id} reference_text_sha256 mismatch")
cp_edits = _int_text(row.get("text_codepoint_edits", ""), "text_codepoint_edits", row_index, issues)
cp_len = _int_text(row.get("text_codepoint_ref_len", ""), "text_codepoint_ref_len", row_index, issues)
gr_edits = _int_text(row.get("text_grapheme_edits", ""), "text_grapheme_edits", row_index, issues)
gr_len = _int_text(row.get("text_grapheme_ref_len", ""), "text_grapheme_ref_len", row_index, issues)
exact = _bool_text(row.get("text_exact", ""))
if exact is None:
issues.append(f"text-pairs.tsv {sample_id} text_exact is not boolean-like: {row.get('text_exact')!r}")
expected_counts = _text_micro_counts(prediction_text, reference_text)
for field, recorded, expected in (
("text_codepoint_edits", cp_edits, expected_counts[0]),
("text_codepoint_ref_len", cp_len, expected_counts[1]),
("text_grapheme_edits", gr_edits, expected_counts[2]),
("text_grapheme_ref_len", gr_len, expected_counts[3]),
):
if recorded is not None and recorded != expected:
issues.append(f"text-pairs.tsv {sample_id} {field} mismatch: recorded={recorded} recomputed={expected}")
expected_exact = _nfc(prediction_text) == _nfc(reference_text)
if exact is not None and exact != expected_exact:
issues.append(f"text-pairs.tsv {sample_id} text_exact mismatch: recorded={exact} recomputed={expected_exact}")
page_row = page_rows_by_id.get(sample_id)
if page_row is not None:
for field in ("input_path", "document_path", "text_codepoint_edits", "text_codepoint_ref_len", "text_grapheme_edits", "text_grapheme_ref_len", "text_exact"):
if str(row.get(field) or "") != str(page_row.get(field) or ""):
issues.append(
f"text-pairs.tsv {sample_id} {field} does not match page-metrics.csv: "
f"recorded={row.get(field)!r} page_metrics={page_row.get(field)!r}"
)
manifest_row = manifest_rows_by_id.get(sample_id)
if manifest_row is None:
issues.append(f"text-pairs.tsv {sample_id} missing manifest row for source text audit")
continue
document_path = Path(row.get("document_path") or "")
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
document_text, document_issues = _audit_document_text(document_path)
reference_source_text, reference_issues = _audit_reference_text(input_path, manifest_row["metadata"], manifest_dir)
issues.extend(document_issues)
issues.extend(reference_issues)
if not document_issues and document_text is not None and prediction_text != document_text:
issues.append(f"text-pairs.tsv {sample_id} prediction_text does not match document JSON text")
if not reference_issues and reference_source_text is not None and reference_text != reference_source_text:
issues.append(f"text-pairs.tsv {sample_id} reference_text does not match reference source text")
return issues
def _page_confusions_from_text_pairs(score_dir: Path, limit: int = 200) -> dict[str, int]:
counts: dict[str, int] = {}
for row in _tsv_rows(score_dir / "text-pairs.tsv"):
prediction = _nfc(row.get("prediction_text") or "")
reference = _nfc(row.get("reference_text") or "")
for opcode in Levenshtein.opcodes(prediction, reference):
tag, i1, i2, j1, j2 = opcode
if tag == "equal":
continue
left = prediction[i1:i2] or "<eps>"
right = reference[j1:j2] or "<eps>"
key = f"{left}->{right}"
counts[key] = counts.get(key, 0) + 1
return dict(sorted(counts.items(), key=lambda item: item[1], reverse=True)[:limit])
def _check_page_confusions(score_dir: Path) -> list[str]:
path = score_dir / "confusions.json"
if not path.exists():
return []
try:
payload = _read_json(path)
except (FileNotFoundError, ValueError) as exc:
return [str(exc)]
if not isinstance(payload, dict):
return ["confusions.json must contain a JSON object"]
issues: list[str] = []
recorded: dict[str, int] = {}
for key, value in payload.items():
if isinstance(value, bool) or not isinstance(value, int):
issues.append(f"confusions.json {key!r} count is not an integer: {value!r}")
continue
recorded[str(key)] = value
expected = _page_confusions_from_text_pairs(score_dir)
if recorded != expected:
issues.append(f"confusions.json does not match recomputed text-pairs.tsv confusions: recorded={recorded} recomputed={expected}")
return issues
def _expected_line_alignment_rows(sample_id: str, reference_lines: list[dict[str, Any]], predicted_lines: list[TextLine], iou_threshold: float) -> list[dict[str, Any]]:
unmatched_predicted = set(range(len(predicted_lines)))
rows: list[dict[str, Any]] = []
for reference in reference_lines:
best_index: int | None = None
best_iou = 0.0
for predicted_index in sorted(unmatched_predicted):
predicted = predicted_lines[predicted_index]
if predicted.page_index != reference["page_index"]:
continue
overlap = bbox_iou(predicted.bbox, reference["bbox"])
if overlap > best_iou:
best_iou = overlap
best_index = predicted_index
if best_index is None or best_iou < iou_threshold:
rows.append(
{
"sample_id": sample_id,
"kind": "reference_unmatched",
"reference_line_id": reference["line_id"],
"predicted_line_id": "",
"reference_bbox": _bbox_text(reference["bbox"]),
"predicted_bbox": "",
"page_index": reference["page_index"],
"iou": best_iou,
"reference_text": reference["text"],
"predicted_text": "",
"text_cer_codepoint": 1.0 if reference["text"] else 0.0,
}
)
continue
predicted = predicted_lines[best_index]
unmatched_predicted.remove(best_index)
rows.append(
{
"sample_id": sample_id,
"kind": "matched",
"reference_line_id": reference["line_id"],
"predicted_line_id": predicted.line_id or "",
"reference_bbox": _bbox_text(reference["bbox"]),
"predicted_bbox": _bbox_text(predicted.bbox),
"page_index": reference["page_index"],
"iou": best_iou,
"reference_text": reference["text"],
"predicted_text": predicted.text,
"text_cer_codepoint": _codepoint_cer(predicted.text, reference["text"]),
}
)
for predicted_index in sorted(unmatched_predicted):
predicted = predicted_lines[predicted_index]
rows.append(
{
"sample_id": sample_id,
"kind": "prediction_unmatched",
"reference_line_id": "",
"predicted_line_id": predicted.line_id or "",
"reference_bbox": "",
"predicted_bbox": _bbox_text(predicted.bbox),
"page_index": predicted.page_index,
"iou": 0.0,
"reference_text": "",
"predicted_text": predicted.text,
"text_cer_codepoint": 1.0 if predicted.text else 0.0,
}
)
return rows
def _compare_line_alignment_rows(sample_id: str, recorded_rows: list[dict[str, str]], expected_rows: list[dict[str, Any]], issues: list[str]) -> None:
if len(recorded_rows) != len(expected_rows):
issues.append(f"line-alignment.csv {sample_id} row count mismatch: recorded={len(recorded_rows)} recomputed={len(expected_rows)}")
comparable = min(len(recorded_rows), len(expected_rows))
for index in range(comparable):
recorded = recorded_rows[index]
expected = expected_rows[index]
row_label = f"line-alignment.csv {sample_id} row {index + 1}"
for field in (
"sample_id",
"kind",
"reference_line_id",
"predicted_line_id",
"reference_bbox",
"predicted_bbox",
"reference_text",
"predicted_text",
):
if (recorded.get(field) or "") != str(expected[field]):
issues.append(f"{row_label} {field} mismatch")
recorded_page = _int_text(recorded.get("page_index", ""), "page_index", index + 1, issues)
if recorded_page is not None and recorded_page != int(expected["page_index"]):
issues.append(f"{row_label} page_index mismatch: recorded={recorded_page} recomputed={expected['page_index']}")
for field in ("iou", "text_cer_codepoint"):
recorded_float = _float_text(recorded.get(field, ""), field, index + 1, issues)
if recorded_float is not None and not _close(recorded_float, float(expected[field])):
issues.append(f"{row_label} {field} mismatch: recorded={recorded_float} recomputed={expected[field]}")
def _check_page_line_alignment_recompute(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
page_rows = _csv_rows(score_dir / "page-metrics.csv")
line_rows = _csv_rows(score_dir / "line-alignment.csv")
if not page_rows:
return ["page-metrics.csv has no scored rows"]
if not line_rows:
return ["line-alignment.csv has no alignment rows"]
line_rows_by_sample: dict[str, list[dict[str, str]]] = {}
for row in line_rows:
sample_id = row.get("sample_id") or ""
if sample_id:
line_rows_by_sample.setdefault(sample_id, []).append(row)
page_sample_ids = {row.get("sample_id") or "" for row in page_rows}
extra_line_samples = sorted(set(line_rows_by_sample) - page_sample_ids)
if extra_line_samples:
issues.append(f"line-alignment.csv has samples not present in page-metrics.csv: {extra_line_samples[:10]}")
iou_threshold = eval_run.get("line_iou_threshold")
if not _is_number(iou_threshold):
issues.append(f"page-eval-run.json line_iou_threshold is not numeric: {iou_threshold!r}")
iou_threshold = 0.0
manifest_rows_by_id: dict[str, dict[str, Any]] = {}
eval_manifest = eval_run.get("eval_manifest")
manifest_dir = Path("")
if isinstance(eval_manifest, str) and eval_manifest:
manifest_path = Path(eval_manifest)
manifest_dir = manifest_path.parent
manifest_rows, manifest_issues = _read_page_manifest_source(manifest_path)
issues.extend(manifest_issues)
manifest_rows_by_id = {str(row["sample_id"]): row for row in manifest_rows}
else:
issues.append("page-eval-run.json missing eval_manifest for line-alignment audit")
metric_values: dict[str, list[float]] = {}
for row_index, page_row in enumerate(page_rows, 1):
sample_id = page_row.get("sample_id") or f"row {row_index}"
sample_line_rows = line_rows_by_sample.get(sample_id, [])
manifest_row = manifest_rows_by_id.get(sample_id)
document_path_value = page_row.get("document_path")
if manifest_row is None:
issues.append(f"line-alignment.csv {sample_id} missing manifest row for recomputation")
elif not document_path_value:
issues.append(f"page-metrics.csv {sample_id} missing document_path for line-alignment recomputation")
else:
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
reference_payload, reference_issues = _audit_reference_payload(input_path, manifest_row["metadata"], manifest_dir)
issues.extend(reference_issues)
predicted_lines, predicted_issues = _audit_predicted_lines(Path(document_path_value))
issues.extend(predicted_issues)
if not reference_issues and not predicted_issues:
expected_rows = _expected_line_alignment_rows(
sample_id,
_audit_reference_lines(reference_payload),
predicted_lines,
float(iou_threshold),
)
_compare_line_alignment_rows(sample_id, sample_line_rows, expected_rows, issues)
matched_rows = [row for row in sample_line_rows if row.get("kind") == "matched"]
reference_count = sum(1 for row in sample_line_rows if row.get("kind") in {"matched", "reference_unmatched"})
prediction_count = sum(1 for row in sample_line_rows if row.get("kind") in {"matched", "prediction_unmatched"})
match_count = len(matched_rows)
matched_ious: list[float] = []
for line_index, line_row in enumerate(matched_rows, 1):
parsed_iou = _float_text(line_row.get("iou", ""), "iou", line_index, issues)
if parsed_iou is not None:
matched_ious.append(parsed_iou)
recomputed = _detection_values(prediction_count, reference_count, match_count)
recomputed["line_mean_iou"] = sum(matched_ious) / len(matched_ious) if matched_ious else 0.0
for field, expected in recomputed.items():
if field not in page_row:
issues.append(f"page-metrics.csv {sample_id} missing {field}")
continue
recorded = _float_text(page_row.get(field, ""), field, row_index, issues)
if recorded is not None and not _close(recorded, float(expected)):
issues.append(f"page-metrics.csv {sample_id} {field} mismatch: recorded={recorded} recomputed={expected}")
for field, value in page_row.items():
if field in {
"sample_id",
"input_path",
"input_sha256",
"document_path",
"document_sha256",
"language",
"script",
"domain",
"gt_status",
"split",
"reference_fields",
"text_cer_codepoint",
"text_cer_grapheme",
}:
continue
if value in (None, ""):
continue
try:
metric_values.setdefault(field, []).append(float(value))
except ValueError:
continue
metric_means = metrics.get("metric_means")
if not isinstance(metric_means, dict):
return issues
for field, values in sorted(metric_values.items()):
if field not in metric_means:
continue
expected = sum(values) / len(values)
recorded = metric_means.get(field)
if not _is_number(recorded) or not _close(float(recorded), expected):
issues.append(f"metrics metric_means.{field} does not match page-metrics.csv mean: recorded={recorded} recomputed={expected}")
return issues
def _check_page_text_metrics_recompute(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
page_rows = _csv_rows(score_dir / "page-metrics.csv")
if not page_rows:
return ["page-metrics.csv has no scored rows"]
manifest_rows_by_id: dict[str, dict[str, Any]] = {}
eval_manifest = eval_run.get("eval_manifest")
manifest_dir = Path("")
if isinstance(eval_manifest, str) and eval_manifest:
manifest_path = Path(eval_manifest)
manifest_dir = manifest_path.parent
manifest_rows, manifest_issues = _read_page_manifest_source(manifest_path)
issues.extend(manifest_issues)
manifest_rows_by_id = {str(row["sample_id"]): row for row in manifest_rows}
else:
issues.append("page-eval-run.json missing eval_manifest for text metric audit")
cp_edits_total = 0
cp_len_total = 0
gr_edits_total = 0
gr_len_total = 0
text_reference_pages = 0
text_exact_pages = 0
zero_length_text_references = 0
for row_index, row in enumerate(page_rows, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
required = {
"text_codepoint_edits",
"text_codepoint_ref_len",
"text_cer_codepoint",
"text_grapheme_edits",
"text_grapheme_ref_len",
"text_cer_grapheme",
"text_exact",
}
missing = sorted(required - set(row))
if missing:
issues.append(f"page-metrics.csv {sample_id} missing text metric columns: {missing}")
continue
cp_edits = _int_text(row.get("text_codepoint_edits", ""), "text_codepoint_edits", row_index, issues)
cp_len = _int_text(row.get("text_codepoint_ref_len", ""), "text_codepoint_ref_len", row_index, issues)
gr_edits = _int_text(row.get("text_grapheme_edits", ""), "text_grapheme_edits", row_index, issues)
gr_len = _int_text(row.get("text_grapheme_ref_len", ""), "text_grapheme_ref_len", row_index, issues)
cp_cer = _float_text(row.get("text_cer_codepoint", ""), "text_cer_codepoint", row_index, issues)
gr_cer = _float_text(row.get("text_cer_grapheme", ""), "text_cer_grapheme", row_index, issues)
recorded_exact = _bool_text(row.get("text_exact", ""))
if recorded_exact is None:
issues.append(f"page-metrics.csv {sample_id} text_exact is not boolean-like: {row.get('text_exact')!r}")
if cp_edits is None or cp_len is None or gr_edits is None or gr_len is None:
continue
expected_counts = (cp_edits, cp_len, gr_edits, gr_len)
expected_exact = False
has_text_reference = "text" in {field for field in (row.get("reference_fields") or "").split(";") if field}
manifest_row = manifest_rows_by_id.get(sample_id)
document_path_value = row.get("document_path")
if manifest_row is None:
issues.append(f"page-metrics.csv {sample_id} missing manifest row for text metric recomputation")
elif not document_path_value:
issues.append(f"page-metrics.csv {sample_id} missing document_path for text metric recomputation")
else:
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
document_text, document_issues = _audit_document_text(Path(document_path_value))
reference_text, reference_issues = _audit_reference_text(input_path, manifest_row["metadata"], manifest_dir)
issues.extend(document_issues)
issues.extend(reference_issues)
if not document_issues and not reference_issues:
if reference_text is None:
expected_counts = (0, 0, 0, 0)
expected_exact = False
has_text_reference = False
else:
expected_counts = _text_micro_counts(document_text or "", reference_text)
expected_exact = _nfc(document_text or "") == _nfc(reference_text)
has_text_reference = True
if len(_nfc(reference_text)) == 0:
zero_length_text_references += 1
expected_cp_edits, expected_cp_len, expected_gr_edits, expected_gr_len = expected_counts
for field, recorded, expected in (
("text_codepoint_edits", cp_edits, expected_cp_edits),
("text_codepoint_ref_len", cp_len, expected_cp_len),
("text_grapheme_edits", gr_edits, expected_gr_edits),
("text_grapheme_ref_len", gr_len, expected_gr_len),
):
if recorded != expected:
issues.append(f"page-metrics.csv {sample_id} {field} mismatch: recorded={recorded} recomputed={expected}")
expected_cp_cer = expected_counts[0] / max(expected_counts[1], 1)
expected_gr_cer = expected_counts[2] / max(expected_counts[3], 1)
if cp_cer is not None and not _close(cp_cer, expected_cp_cer):
issues.append(f"page-metrics.csv {sample_id} text_cer_codepoint mismatch: recorded={cp_cer} recomputed={expected_cp_cer}")
if gr_cer is not None and not _close(gr_cer, expected_gr_cer):
issues.append(f"page-metrics.csv {sample_id} text_cer_grapheme mismatch: recorded={gr_cer} recomputed={expected_gr_cer}")
if recorded_exact is not None and recorded_exact != expected_exact:
issues.append(f"page-metrics.csv {sample_id} text_exact mismatch: recorded={recorded_exact} recomputed={expected_exact}")
cp_edits_total += expected_counts[0]
cp_len_total += expected_counts[1]
gr_edits_total += expected_counts[2]
gr_len_total += expected_counts[3]
if has_text_reference:
text_reference_pages += 1
text_exact_pages += int(expected_exact)
expected_values = {
"text_codepoint_edits": cp_edits_total,
"text_codepoint_ref_len": cp_len_total,
"text_grapheme_edits": gr_edits_total,
"text_grapheme_ref_len": gr_len_total,
"text_reference_pages": text_reference_pages,
"text_exact_pages": text_exact_pages,
"zero_length_text_references": zero_length_text_references,
}
for key, expected in expected_values.items():
if metrics.get(key) != expected:
issues.append(f"metrics {key} does not match page-metrics.csv sum: recorded={metrics.get(key)} recomputed={expected}")
if zero_length_text_references:
_claim_blocker(issues, eval_run, "page", "zero-length text references", "zero-length text references present:")
expected_cp_cer = cp_edits_total / max(cp_len_total, 1)
if not _is_number(metrics.get("text_cer_codepoint_micro")) or not _close(float(metrics["text_cer_codepoint_micro"]), expected_cp_cer):
issues.append(
"metrics text_cer_codepoint_micro does not match page-metrics.csv sum: "
f"recorded={metrics.get('text_cer_codepoint_micro')} recomputed={expected_cp_cer}"
)
expected_gr_cer = gr_edits_total / max(gr_len_total, 1)
if not _is_number(metrics.get("text_cer_grapheme_micro")) or not _close(float(metrics["text_cer_grapheme_micro"]), expected_gr_cer):
issues.append(
"metrics text_cer_grapheme_micro does not match page-metrics.csv sum: "
f"recorded={metrics.get('text_cer_grapheme_micro')} recomputed={expected_gr_cer}"
)
expected_exact_accuracy = text_exact_pages / max(text_reference_pages, 1)
if not _is_number(metrics.get("text_exact_accuracy")) or not _close(float(metrics["text_exact_accuracy"]), expected_exact_accuracy):
issues.append(
"metrics text_exact_accuracy does not match page-metrics.csv sum: "
f"recorded={metrics.get('text_exact_accuracy')} recomputed={expected_exact_accuracy}"
)
return issues
def _page_uncertainty_units(rows: list[dict[str, str]]) -> tuple[list[dict[str, int]], list[str]]:
issues: list[str] = []
units: list[dict[str, int]] = []
for row_index, row in enumerate(rows, 1):
cp_edits = _int_text(row.get("text_codepoint_edits", ""), "text_codepoint_edits", row_index, issues)
cp_len = _int_text(row.get("text_codepoint_ref_len", ""), "text_codepoint_ref_len", row_index, issues)
gr_edits = _int_text(row.get("text_grapheme_edits", ""), "text_grapheme_edits", row_index, issues)
gr_len = _int_text(row.get("text_grapheme_ref_len", ""), "text_grapheme_ref_len", row_index, issues)
exact = _bool_text(row.get("text_exact", ""))
if exact is None:
issues.append(f"page-metrics.csv row {row_index} text_exact is not boolean-like: {row.get('text_exact')!r}")
if cp_edits is None or cp_len is None or gr_edits is None or gr_len is None:
continue
units.append(
{
"codepoint_edits": cp_edits,
"codepoint_ref_len": cp_len,
"grapheme_edits": gr_edits,
"grapheme_ref_len": gr_len,
"exact": int(bool(exact)),
"text_reference": int("text" in {field for field in (row.get("reference_fields") or "").split(";") if field}),
}
)
return units, issues
def _check_page_uncertainty(score_dir: Path, metrics: dict[str, Any]) -> list[str]:
rows = _csv_rows(score_dir / "page-metrics.csv")
units, issues = _page_uncertainty_units(rows)
expected = _expected_uncertainty(
units,
method="deterministic_page_bootstrap",
metrics=("text_cer_codepoint_micro", "text_cer_grapheme_micro", "text_exact_accuracy"),
)
issues.extend(_check_uncertainty(metrics.get("uncertainty"), expected, "page"))
for index, row in enumerate(metrics.get("slices") if isinstance(metrics.get("slices"), list) else [], 1):
if not isinstance(row, dict):
continue
slice_key = (
str(row.get("language") or ""),
str(row.get("script") or ""),
str(row.get("domain") or ""),
str(row.get("gt_status") or ""),
str(row.get("split") or ""),
)
slice_rows = [
source_row
for source_row in rows
if (
str(source_row.get("language") or ""),
str(source_row.get("script") or ""),
str(source_row.get("domain") or ""),
str(source_row.get("gt_status") or ""),
str(source_row.get("split") or ""),
)
== slice_key
]
slice_units, slice_issues = _page_uncertainty_units(slice_rows)
issues.extend(slice_issues)
expected_slice = _expected_uncertainty(
slice_units,
method="deterministic_page_bootstrap",
metrics=("text_cer_codepoint_micro", "text_cer_grapheme_micro", "text_exact_accuracy"),
)
issues.extend(_check_uncertainty(row.get("uncertainty"), expected_slice, f"page slices[{index}]"))
return issues
def _check_page_script_inventory(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
inventory_path = Path(eval_run["script_inventory_file"]) if isinstance(eval_run.get("script_inventory_file"), str) and eval_run.get("script_inventory_file") else None
manifest_rows_by_id: dict[str, dict[str, Any]] = {}
eval_manifest = eval_run.get("eval_manifest")
manifest_dir = Path("")
if isinstance(eval_manifest, str) and eval_manifest:
manifest_path = Path(eval_manifest)
manifest_dir = manifest_path.parent
manifest_rows, manifest_issues = _read_page_manifest_source(manifest_path)
issues.extend(manifest_issues)
manifest_rows_by_id = {str(row["sample_id"]): row for row in manifest_rows}
else:
issues.append("page-eval-run.json missing eval_manifest for script inventory audit")
reference_texts: list[str] = []
for row in _csv_rows(score_dir / "page-metrics.csv"):
sample_id = row.get("sample_id") or ""
manifest_row = manifest_rows_by_id.get(sample_id)
if manifest_row is None:
issues.append(f"page-metrics.csv {sample_id or '(missing sample_id)'} missing manifest row for script inventory audit")
continue
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
reference_text, reference_issues = _audit_reference_text(input_path, manifest_row["metadata"], manifest_dir)
issues.extend(reference_issues)
if reference_text is not None:
reference_texts.append(reference_text)
expected, inventory_issues = _expected_script_inventory(reference_texts, inventory_path)
issues.extend(inventory_issues)
issues.extend(_check_script_inventory(metrics.get("script_inventory"), expected, "page metrics"))
issues.extend(_check_script_inventory(eval_run.get("script_inventory"), expected, "page eval-run"))
return issues
def _check_page_dictionary_coverage(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
reference_texts = [row.get("reference_text") or "" for row in _tsv_rows(score_dir / "text-pairs.tsv")]
expected = _expected_dictionary_coverage(reference_texts, eval_run.get("dictionary"))
issues = _check_dictionary_coverage(metrics.get("dictionary_coverage"), expected, "page metrics")
if expected.get("checked") and expected.get("missing_gt_codepoints"):
_claim_blocker(issues, eval_run, "page", "GT codepoints missing from dictionary", "GT codepoints missing from dictionary:")
return issues
def _expected_page_worst_rows(score_dir: Path) -> tuple[list[dict[str, Any]], list[str]]:
issues: list[str] = []
expected_rows: list[dict[str, Any]] = []
reference_by_id = {row.get("sample_id") or "": row for row in _tsv_rows(score_dir / "reference-artifacts.tsv")}
for row_index, row in enumerate(_csv_rows(score_dir / "page-metrics.csv"), 1):
sample_id = row.get("sample_id") or f"row {row_index}"
reference_row = reference_by_id.get(sample_id, {})
cp_edits = _int_text(row.get("text_codepoint_edits", ""), "text_codepoint_edits", row_index, issues)
cp_len = _int_text(row.get("text_codepoint_ref_len", ""), "text_codepoint_ref_len", row_index, issues)
gr_edits = _int_text(row.get("text_grapheme_edits", ""), "text_grapheme_edits", row_index, issues)
gr_len = _int_text(row.get("text_grapheme_ref_len", ""), "text_grapheme_ref_len", row_index, issues)
if cp_edits is None or cp_len is None or gr_edits is None or gr_len is None:
continue
expected_rows.append(
{
"sample_id": sample_id,
"input_path": row.get("input_path") or "",
"input_sha256": row.get("input_sha256") or "",
"input_uri": Path(row.get("input_path") or "").resolve().as_uri() if row.get("input_path") else "",
"document_path": row.get("document_path") or "",
"document_sha256": row.get("document_sha256") or "",
"document_source_path": row.get("document_source_path") or "",
"language": row.get("language") or "",
"script": row.get("script") or "",
"domain": row.get("domain") or "",
"gt_status": row.get("gt_status") or "",
"split": row.get("split") or "",
"claim_eligible": row.get("claim_eligible") or "",
"source_document": row.get("source_document") or "",
"source_page": row.get("source_page") or "",
"source_bbox": row.get("source_bbox") or "",
"reference_path": reference_row.get("reference_path") or "",
"reference_sha256": reference_row.get("reference_sha256") or "",
"reference_status": reference_row.get("reference_status") or "",
"reference_claim_eligible": reference_row.get("reference_claim_eligible") or "",
"reference_fields": row.get("reference_fields") or "",
"text_cer_codepoint": cp_edits / max(cp_len, 1),
"text_cer_grapheme": gr_edits / max(gr_len, 1),
"sort_key": gr_edits / max(gr_len, 1),
}
)
expected_rows.sort(key=lambda row: row["sort_key"], reverse=True)
return expected_rows, issues
def _check_page_worst_artifacts(score_dir: Path) -> list[str]:
issues: list[str] = []
expected_rows, row_issues = _expected_page_worst_rows(score_dir)
issues.extend(row_issues)
worst_rows = _csv_rows(score_dir / "worst-pages.csv")
review_rows = _csv_rows(score_dir / "sample-review.csv")
issues.extend(_check_sample_id_order("worst-pages.csv", expected_rows, worst_rows))
issues.extend(_check_sample_id_order("sample-review.csv", expected_rows, review_rows))
issues.extend(
_check_review_csv_fields(
"sample-review.csv",
expected_rows,
review_rows,
("input_path", "document_path", "document_source_path", "text_cer_grapheme", "text_cer_codepoint", "reference_fields"),
)
)
issues.extend(
_check_review_html_contents(
score_dir,
"page",
expected_rows,
review_rows,
("sample_id", "input_uri", "document_path", "document_source_path", "reference_fields"),
)
)
expected_by_id = {str(row["sample_id"]): row for row in expected_rows}
for row_index, row in enumerate(worst_rows, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
expected = expected_by_id.get(sample_id)
if expected is None:
continue
for field in ("text_cer_grapheme", "text_cer_codepoint"):
recorded = _float_text(row.get(field, ""), field, row_index, issues)
if recorded is not None and not _close(recorded, float(expected[field])):
issues.append(f"worst-pages.csv {sample_id} {field} mismatch: recorded={recorded} recomputed={expected[field]}")
for field in (
"input_path",
"input_sha256",
"document_path",
"document_sha256",
"document_source_path",
"language",
"script",
"domain",
"gt_status",
"split",
"claim_eligible",
"source_document",
"source_page",
"source_bbox",
"reference_path",
"reference_sha256",
"reference_status",
"reference_claim_eligible",
"reference_fields",
):
if (row.get(field) or "") != str(expected[field]):
issues.append(f"worst-pages.csv {sample_id} {field} mismatch")
return issues
def _check_page_output_duplicates(score_dir: Path) -> list[str]:
issues: list[str] = []
page_rows = _csv_rows(score_dir / "page-metrics.csv")
missing_rows = _csv_rows(score_dir / "missing.tsv")
worst_rows = _csv_rows(score_dir / "worst-pages.csv")
review_rows = _csv_rows(score_dir / "sample-review.csv")
issues.extend(_check_unique_field("page-metrics.csv", page_rows, "sample_id"))
issues.extend(_check_unique_field("page-metrics.csv", page_rows, "input_path"))
issues.extend(_check_unique_field("page-metrics.csv", page_rows, "document_path"))
issues.extend(_check_unique_key("missing.tsv", missing_rows, ("kind", "sample_id")))
issues.extend(_check_unique_field("worst-pages.csv", worst_rows, "sample_id"))
issues.extend(_check_unique_field("sample-review.csv", review_rows, "sample_id"))
return issues
def _resolve_audit_manifest_path(path_value: str, manifest_dir: Path) -> Path:
path = Path(path_value)
if path.is_absolute() or path.exists():
return path
candidate = manifest_dir / path
if candidate.exists():
return candidate
return path
def _page_document_candidates(sample_id: str, input_path: Path, documents_dir: Path) -> list[Path]:
return [
documents_dir / sample_id / "document.json",
documents_dir / input_path.stem / "document.json",
documents_dir / f"{sample_id}.json",
documents_dir / f"{input_path.stem}.json",
]
def _page_reference_exists(input_path: Path, metadata: dict[str, Any], manifest_dir: Path) -> bool:
explicit = metadata.get("reference_path")
candidates: list[Path] = []
if isinstance(explicit, str) and explicit:
candidates.append(_resolve_audit_manifest_path(explicit, manifest_dir))
candidates.extend(
[
input_path.with_suffix(input_path.suffix + ".ref.json"),
input_path.with_suffix(".ref.json"),
input_path.with_suffix(input_path.suffix + ".ref.txt"),
input_path.with_suffix(".ref.txt"),
]
)
return any(candidate.exists() for candidate in candidates)
def _read_page_manifest_source(path: Path) -> tuple[list[dict[str, Any]], list[str]]:
issues: list[str] = []
rows: list[dict[str, Any]] = []
seen: set[str] = set()
seen_image_paths: set[str] = set()
if not path.exists():
return rows, [f"page eval manifest does not exist for missing.tsv audit: {path}"]
with path.open("r", encoding="utf-8") as handle:
for line_no, line in enumerate(handle, 1):
line = line.strip()
if not line:
continue
try:
payload = json.loads(line)
except json.JSONDecodeError as exc:
issues.append(f"page eval manifest {path}:{line_no} invalid JSON: {exc}")
continue
if not isinstance(payload, dict):
issues.append(f"page eval manifest {path}:{line_no} row is not an object")
continue
sample_id = str(payload.get("sample_id") or payload.get("id") or "")
image_path = str(payload.get("image_path") or payload.get("image") or "")
if not sample_id:
issues.append(f"page eval manifest {path}:{line_no} missing sample_id")
continue
if not image_path:
issues.append(f"page eval manifest {path}:{line_no} missing image_path")
continue
if sample_id in seen:
issues.append(f"page eval manifest duplicate sample_id: {sample_id}")
seen.add(sample_id)
if image_path in seen_image_paths:
issues.append(f"page eval manifest duplicate image_path: {image_path}")
seen_image_paths.add(image_path)
declared_sha256 = str(payload.get("sha256") or "")
if declared_sha256:
resolved_image = _resolve_audit_manifest_path(image_path, path.parent)
if resolved_image.exists():
actual_sha256 = _sha256_file(resolved_image)
if actual_sha256 != declared_sha256:
issues.append(
f"page eval manifest {sample_id} sha256 mismatch: "
f"recorded={declared_sha256} actual={actual_sha256}"
)
rows.append(
{
"sample_id": sample_id,
"image_path": image_path,
"sha256": declared_sha256,
"split": str(payload.get("split") or ""),
"metadata": dict(payload.get("metadata") or {}),
}
)
return rows, issues
def _audit_page_metadata_value(metadata: dict[str, Any], key: str, fallback: Any) -> str:
value = metadata.get(key)
if value in (None, ""):
return str(fallback or "")
return str(value)
def _audit_page_split(manifest_row: dict[str, Any]) -> str:
metadata = manifest_row["metadata"]
value = metadata.get("split") or metadata.get("split_role") or manifest_row.get("split") or ""
return str(value)
def _audit_reference_fields(input_path: Path, metadata: dict[str, Any], manifest_dir: Path) -> tuple[list[str], list[str]]:
issues: list[str] = []
explicit = metadata.get("reference_path")
explicit_path = _resolve_audit_manifest_path(explicit, manifest_dir) if isinstance(explicit, str) and explicit else None
try:
reference = load_reference(input_path, explicit_path=explicit_path)
except Exception as exc:
return [], [f"reference failed to parse for page-metrics identity audit: {input_path}: {exc}"]
if reference is None:
return [], [f"reference is missing for page-metrics identity audit: {input_path}"]
fields: list[str] = []
if reference.text is not None:
fields.append("text")
if reference.markdown is not None:
fields.append("markdown")
if reference.reading_order:
fields.append("reading_order")
if reference.tables:
fields.append("tables")
if reference.figures:
fields.append("figures")
payload, payload_issues = _audit_reference_payload(input_path, metadata, manifest_dir)
issues.extend(payload_issues)
if _audit_reference_lines(payload):
fields.append("line_bboxes")
return fields, issues
def _expected_page_metric_identity_rows(eval_run: dict[str, Any]) -> tuple[list[dict[str, str]], list[str]]:
issues: list[str] = []
eval_manifest = eval_run.get("eval_manifest")
documents_dir_value = eval_run.get("documents_dir")
if not isinstance(eval_manifest, str) or not eval_manifest:
return [], ["page-eval-run.json missing eval_manifest for page-metrics identity audit"]
if not isinstance(documents_dir_value, str) or not documents_dir_value:
return [], ["page-eval-run.json missing documents_dir for page-metrics identity audit"]
manifest_path = Path(eval_manifest)
manifest_rows, manifest_issues = _read_page_manifest_source(manifest_path)
issues.extend(manifest_issues)
manifest_dir = manifest_path.parent
documents_dir = Path(documents_dir_value)
expected: list[dict[str, str]] = []
for manifest_row in manifest_rows:
sample_id = str(manifest_row["sample_id"])
metadata = manifest_row["metadata"]
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
document_path = next(
(candidate for candidate in _page_document_candidates(sample_id, input_path, documents_dir) if candidate.exists()),
None,
)
if document_path is None:
continue
document_source_path, document_source_issues = _audit_document_source_path(document_path)
issues.extend(document_source_issues)
if document_source_issues:
continue
reference_fields, reference_issues = _audit_reference_fields(input_path, metadata, manifest_dir)
issues.extend(reference_issues)
if reference_issues:
continue
expected.append(
{
"sample_id": sample_id,
"input_path": str(input_path),
"input_sha256": _sha256_file(input_path) if input_path.exists() else "",
"document_path": str(document_path),
"document_source_path": document_source_path or "",
"ocr_output_kind": str(eval_run.get("ocr_output_kind") or ""),
"language": _audit_page_metadata_value(metadata, "language", eval_run.get("language_default")),
"script": _audit_page_metadata_value(metadata, "script", eval_run.get("script_default")),
"domain": _audit_page_metadata_value(metadata, "domain", eval_run.get("domain_default")),
"gt_status": _audit_page_metadata_value(metadata, "gt_status", eval_run.get("gt_status_default")),
"split": _audit_page_split(manifest_row),
"claim_eligible": str(
_metadata_bool_value(metadata, CLAIM_ELIGIBLE_KEYS, eval_run.get("claim_eligible_default"))
),
"source_document": _metadata_value_for_keys(metadata, SOURCE_DOCUMENT_KEYS),
"source_page": _metadata_value_for_keys(metadata, PAGE_KEYS),
"source_bbox": _metadata_value_for_keys(metadata, BBOX_KEYS),
"reference_fields": ";".join(reference_fields),
}
)
return expected, issues
def _check_page_metadata_contradictions(score_dir: Path, metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
eval_manifest = eval_run.get("eval_manifest")
if not isinstance(eval_manifest, str) or not eval_manifest:
return ["page-eval-run.json missing eval_manifest for metadata contradiction audit"]
manifest_path = Path(eval_manifest)
manifest_rows, issues = _read_page_manifest_source(manifest_path)
if issues:
return issues
manifest_dir = manifest_path.parent
page_rows = {row.get("sample_id") or "": row for row in _csv_rows(score_dir / "page-metrics.csv")}
expected: list[str] = []
for manifest_row in manifest_rows:
sample_id = str(manifest_row["sample_id"])
page_row = page_rows.get(sample_id)
if page_row is None:
continue
metadata = manifest_row["metadata"]
domain = page_row.get("domain") or ""
gt_status = page_row.get("gt_status") or ""
expected.extend(_metadata_contradictions_for(sample_id, metadata, domain, gt_status))
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
reference_payload, reference_issues = _audit_reference_payload(input_path, metadata, manifest_dir)
issues.extend(reference_issues)
if isinstance(reference_payload, dict) and isinstance(reference_payload.get("metadata"), dict):
expected.extend(_metadata_contradictions_for(sample_id, reference_payload["metadata"], domain, gt_status))
if issues:
return issues
recorded = metrics.get("metadata_contradictions")
if recorded != expected:
return [f"page metadata_contradictions do not match source metadata: recorded={recorded} recomputed={expected}"]
return []
def _check_page_metadata_coverage(metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
eval_manifest = eval_run.get("eval_manifest")
if not isinstance(eval_manifest, str) or not eval_manifest:
return ["page-eval-run.json missing eval_manifest for metadata coverage audit"]
manifest_rows, issues = _read_page_manifest_source(Path(eval_manifest))
if issues:
return issues
def missing(key: str) -> int:
return sum(1 for row in manifest_rows if row["metadata"].get(key) in (None, ""))
def missing_any(keys: tuple[str, ...]) -> int:
return sum(1 for row in manifest_rows if not _metadata_has_any(row["metadata"], keys))
def missing_human_review_evidence() -> int:
return sum(
1
for row in manifest_rows
if row["metadata"].get("gt_status") == "human_verified"
and not _metadata_has_any(row["metadata"], HUMAN_REVIEW_EVIDENCE_KEYS)
)
expected = {
"missing_source_document": missing_any(SOURCE_DOCUMENT_KEYS),
"missing_page": missing_any(PAGE_KEYS),
"missing_bbox": missing_any(BBOX_KEYS),
"missing_claim_eligibility": missing_any(CLAIM_ELIGIBLE_KEYS),
"missing_human_review_evidence": missing_human_review_evidence(),
"missing_language": missing("language"),
"missing_script": missing("script"),
"missing_domain": missing("domain"),
"missing_gt_status": missing("gt_status"),
}
recorded = metrics.get("metadata_coverage")
if recorded != expected:
return [f"page metadata_coverage does not match eval manifest: recorded={recorded} recomputed={expected}"]
return []
def _check_page_reference_coverage(score_dir: Path, metrics: dict[str, Any]) -> list[str]:
issues: list[str] = []
rows = _csv_rows(score_dir / "page-metrics.csv")
fields = ("text", "markdown", "reading_order", "tables", "figures", "line_bboxes")
expected = dict.fromkeys(fields, 0)
line_alignment_scored = 0
for row_index, row in enumerate(rows, 1):
reference_fields = {field for field in (row.get("reference_fields") or "").split(";") if field}
for field in fields:
if field in reference_fields:
expected[field] += 1
raw_reference_count = row.get("line_reference_count")
if raw_reference_count in (None, ""):
continue
reference_count = _int_text(raw_reference_count, "line_reference_count", row_index, issues)
if reference_count is not None and reference_count:
line_alignment_scored += 1
expected["line_alignment_scored"] = line_alignment_scored
recorded = metrics.get("reference_coverage")
if recorded != expected:
issues.append(f"page reference_coverage does not match page-metrics.csv: recorded={recorded} recomputed={expected}")
return issues
def _check_page_metric_identity_rows(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
recorded_rows = _csv_rows(score_dir / "page-metrics.csv")
expected_rows, expected_issues = _expected_page_metric_identity_rows(eval_run)
issues.extend(expected_issues)
recorded_by_id: dict[str, dict[str, str]] = {}
for row_index, row in enumerate(recorded_rows, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
if sample_id not in recorded_by_id:
recorded_by_id[sample_id] = row
expected_by_id = {row["sample_id"]: row for row in expected_rows}
missing = sorted(set(expected_by_id) - set(recorded_by_id))
extra = sorted(set(recorded_by_id) - set(expected_by_id))
if missing:
issues.append(f"page-metrics.csv missing scored manifest rows: {missing[:10]}")
if extra:
issues.append(f"page-metrics.csv has rows not backed by scored manifest/reference/documents: {extra[:10]}")
for sample_id in sorted(set(expected_by_id) & set(recorded_by_id)):
expected = expected_by_id[sample_id]
recorded = recorded_by_id[sample_id]
for field in (
"sample_id",
"input_sha256",
"language",
"script",
"domain",
"gt_status",
"split",
"claim_eligible",
"source_document",
"source_page",
"source_bbox",
"reference_fields",
"document_source_path",
):
if (recorded.get(field) or "") != expected[field]:
issues.append(
f"page-metrics.csv {sample_id} {field} does not match eval manifest/reference: "
f"recorded={recorded.get(field) or ''} recomputed={expected[field]}"
)
for field in ("input_path", "document_path"):
recorded_path = recorded.get(field) or ""
if not _same_path(recorded_path, Path(expected[field])):
issues.append(
f"page-metrics.csv {sample_id} {field} does not match eval manifest/reference: "
f"recorded={recorded_path} recomputed={expected[field]}"
)
return issues
def _expected_page_document_source_path_mismatches(eval_run: dict[str, Any]) -> tuple[list[str], list[str]]:
eval_manifest = eval_run.get("eval_manifest")
documents_dir_value = eval_run.get("documents_dir")
if not isinstance(eval_manifest, str) or not eval_manifest:
return [], ["page-eval-run.json missing eval_manifest for document source-path audit"]
if not isinstance(documents_dir_value, str) or not documents_dir_value:
return [], ["page-eval-run.json missing documents_dir for document source-path audit"]
manifest_path = Path(eval_manifest)
manifest_rows, issues = _read_page_manifest_source(manifest_path)
manifest_dir = manifest_path.parent
documents_dir = Path(documents_dir_value)
mismatches: list[str] = []
for manifest_row in manifest_rows:
sample_id = str(manifest_row["sample_id"])
input_path = _resolve_audit_manifest_path(str(manifest_row["image_path"]), manifest_dir)
document_path = next(
(candidate for candidate in _page_document_candidates(sample_id, input_path, documents_dir) if candidate.exists()),
None,
)
if document_path is None:
continue
document_source_path, document_source_issues = _audit_document_source_path(document_path)
issues.extend(document_source_issues)
if document_source_issues:
continue
if not _same_path(document_source_path or "", input_path):
mismatches.append(
f"{sample_id}: document.source_path {document_source_path!r} does not match eval input {str(input_path)!r}"
)
return mismatches, issues
def _check_page_document_source_path_mismatches(metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
expected, issues = _expected_page_document_source_path_mismatches(eval_run)
if issues:
return issues
recorded = metrics.get("document_source_path_mismatches")
if recorded != expected:
issues.append(f"page document_source_path_mismatches do not match scored documents: recorded={recorded} recomputed={expected}")
recorded_count = metrics.get("document_source_path_mismatch_count")
if recorded_count != len(expected):
issues.append(
"page document_source_path_mismatch_count does not match scored documents: "
f"recorded={recorded_count} recomputed={len(expected)}"
)
return issues
def _check_page_missing_artifact(score_dir: Path, eval_run: dict[str, Any]) -> list[str]:
path = score_dir / "missing.tsv"
if not path.exists():
return []
eval_manifest = eval_run.get("eval_manifest")
documents_dir_value = eval_run.get("documents_dir")
if not isinstance(eval_manifest, str) or not eval_manifest:
return ["page-eval-run.json missing eval_manifest for missing.tsv audit"]
if not isinstance(documents_dir_value, str) or not documents_dir_value:
return ["page-eval-run.json missing documents_dir for missing.tsv audit"]
manifest_path = Path(eval_manifest)
manifest_rows, issues = _read_page_manifest_source(manifest_path)
if issues:
return issues
manifest_dir = manifest_path.parent
documents_dir = Path(documents_dir_value)
expected: list[tuple[str, str]] = []
for row in manifest_rows:
sample_id = str(row["sample_id"])
input_path = _resolve_audit_manifest_path(str(row["image_path"]), manifest_dir)
document_exists = any(candidate.exists() for candidate in _page_document_candidates(sample_id, input_path, documents_dir))
if not document_exists:
expected.append(("missing_document", sample_id))
continue
if not _page_reference_exists(input_path, row["metadata"], manifest_dir):
expected.append(("missing_reference", sample_id))
recorded = [(row.get("kind") or "", row.get("sample_id") or "") for row in _csv_rows(path)]
if recorded != expected:
issues.append(
"missing.tsv rows do not match recomputed manifest/document/reference misses: "
f"recorded={_short_rows(recorded)} recomputed={_short_rows(expected)}"
)
return issues
def _check_page_manifest_count_metrics(metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
eval_manifest = eval_run.get("eval_manifest")
documents_dir_value = eval_run.get("documents_dir")
if not isinstance(eval_manifest, str) or not eval_manifest:
return ["page-eval-run.json missing eval_manifest for manifest count audit"]
if not isinstance(documents_dir_value, str) or not documents_dir_value:
return ["page-eval-run.json missing documents_dir for manifest count audit"]
manifest_path = Path(eval_manifest)
manifest_rows, issues = _read_page_manifest_source(manifest_path)
if issues:
return issues
manifest_dir = manifest_path.parent
documents_dir = Path(documents_dir_value)
missing_documents = 0
missing_references = 0
for row in manifest_rows:
sample_id = str(row["sample_id"])
input_path = _resolve_audit_manifest_path(str(row["image_path"]), manifest_dir)
document_exists = any(candidate.exists() for candidate in _page_document_candidates(sample_id, input_path, documents_dir))
if not document_exists:
missing_documents += 1
continue
if not _page_reference_exists(input_path, row["metadata"], manifest_dir):
missing_references += 1
expected = {
"manifest_samples": len(manifest_rows),
"missing_documents": missing_documents,
"missing_references": missing_references,
"claim_eligible_rows": sum(
1
for row in manifest_rows
if _metadata_bool_value(row["metadata"], CLAIM_ELIGIBLE_KEYS, eval_run.get("claim_eligible_default"))
),
}
expected["scored_pages"] = expected["manifest_samples"] - expected["missing_documents"] - expected["missing_references"]
count_issues: list[str] = []
for key, value in expected.items():
if metrics.get(key) != value:
count_issues.append(f"page metrics {key} does not match manifest/doc/reference sources: recorded={metrics.get(key)} recomputed={value}")
return count_issues
def _check_page_leakage_report(metrics: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
eval_manifest = eval_run.get("eval_manifest")
if not isinstance(eval_manifest, str) or not eval_manifest:
return ["page-eval-run.json missing eval_manifest for leakage audit"]
train_paths, issues = _train_manifest_paths(eval_run)
eval_manifest_path = Path(eval_manifest)
eval_rows, eval_issues = _audit_manifest_leakage_rows(
eval_manifest_path,
manifest_dir=eval_manifest_path.parent,
compute_missing_image_sha=True,
)
issues.extend(eval_issues)
train_rows: list[dict[str, str]] = []
for path in train_paths:
rows, row_issues = _audit_manifest_leakage_rows(path, manifest_dir=path.parent, compute_missing_image_sha=True)
train_rows.extend(rows)
issues.extend(row_issues)
if issues:
return issues
expected = {
"train_manifests": [str(path) for path in train_paths],
"checked": bool(train_paths),
"train_rows": len(train_rows),
"eval_rows": len(eval_rows),
"overlaps": _leakage_overlaps(train_rows, eval_rows),
}
return _compare_leakage_report("page", metrics.get("leakage"), expected)
def _check_page_row_hashes(score_dir: Path) -> list[str]:
issues: list[str] = []
page_rows = _csv_rows(score_dir / "page-metrics.csv")
for row_index, row in enumerate(page_rows, 1):
sample_id = row.get("sample_id") or f"row {row_index}"
for path_field, hash_field, label in (
("input_path", "input_sha256", "input image"),
("document_path", "document_sha256", "document output"),
):
path_value = row.get(path_field)
expected = row.get(hash_field)
if not path_value:
issues.append(f"page-metrics.csv {sample_id} missing {path_field}")
continue
if not expected:
issues.append(f"page-metrics.csv {sample_id} missing {hash_field}")
continue
path = Path(path_value)
if not path.exists():
issues.append(f"page-metrics.csv {sample_id} {label} path does not exist: {path_value}")
continue
actual = _sha256_file(path)
if actual != expected:
issues.append(f"page-metrics.csv {sample_id} {hash_field} mismatch: recorded={expected} actual={actual}")
return issues
PAGE_SLICE_IDENTITY_FIELDS = {
"sample_id",
"input_path",
"input_sha256",
"document_path",
"document_sha256",
"document_source_path",
"ocr_output_kind",
"language",
"script",
"domain",
"gt_status",
"split",
"claim_eligible",
"source_document",
"source_page",
"source_bbox",
"reference_fields",
}
PAGE_SLICE_TEXT_FIELDS = {
"text_codepoint_edits",
"text_codepoint_ref_len",
"text_cer_codepoint",
"text_grapheme_edits",
"text_grapheme_ref_len",
"text_cer_grapheme",
"text_exact",
}
def _page_metric_mean_values(row: dict[str, str], row_index: int, issues: list[str]) -> dict[str, float]:
values: dict[str, float] = {}
excluded = PAGE_SLICE_IDENTITY_FIELDS | PAGE_SLICE_TEXT_FIELDS
for field, value in row.items():
if field in excluded or value in (None, ""):
continue
parsed = _float_text(value, field, row_index, issues)
if parsed is not None:
values[field] = parsed
return values
def _check_page_slice_metrics(score_dir: Path, slices: Any) -> list[str]:
issues: list[str] = []
page_rows = _csv_rows(score_dir / "page-metrics.csv")
if not page_rows:
return ["page-metrics.csv has no scored rows"]
if not isinstance(slices, list) or not slices:
return ["page slice metrics are missing"]
groups: dict[tuple[str, str, str, str, str], dict[str, Any]] = {}
for row_index, row in enumerate(page_rows, 1):
key = (
row.get("language") or "",
row.get("script") or "",
row.get("domain") or "",
row.get("gt_status") or "",
row.get("split") or "",
)
bucket = groups.setdefault(
key,
{
"scored_pages": 0,
"text_codepoint_edits": 0,
"text_codepoint_ref_len": 0,
"text_grapheme_edits": 0,
"text_grapheme_ref_len": 0,
"text_reference_pages": 0,
"text_exact_pages": 0,
"metric_values": {},
},
)
bucket["scored_pages"] += 1
cp_edits = _int_text(row.get("text_codepoint_edits", ""), "text_codepoint_edits", row_index, issues)
cp_len = _int_text(row.get("text_codepoint_ref_len", ""), "text_codepoint_ref_len", row_index, issues)
gr_edits = _int_text(row.get("text_grapheme_edits", ""), "text_grapheme_edits", row_index, issues)
gr_len = _int_text(row.get("text_grapheme_ref_len", ""), "text_grapheme_ref_len", row_index, issues)
if cp_edits is not None:
bucket["text_codepoint_edits"] += cp_edits
if cp_len is not None:
bucket["text_codepoint_ref_len"] += cp_len
if gr_edits is not None:
bucket["text_grapheme_edits"] += gr_edits
if gr_len is not None:
bucket["text_grapheme_ref_len"] += gr_len
if "text" in {field for field in (row.get("reference_fields") or "").split(";") if field}:
bucket["text_reference_pages"] += 1
exact = _bool_text(row.get("text_exact", ""))
if exact is None:
issues.append(f"page-metrics.csv row {row_index} text_exact is not boolean-like: {row.get('text_exact')!r}")
else:
bucket["text_exact_pages"] += int(exact)
metric_values: dict[str, list[float]] = bucket["metric_values"]
for field, value in _page_metric_mean_values(row, row_index, issues).items():
metric_values.setdefault(field, []).append(value)
expected_by_key: dict[tuple[str, str, str, str, str], dict[str, Any]] = {}
for key, bucket in groups.items():
cp_edits = bucket["text_codepoint_edits"]
cp_len = bucket["text_codepoint_ref_len"]
gr_edits = bucket["text_grapheme_edits"]
gr_len = bucket["text_grapheme_ref_len"]
text_reference_pages = bucket["text_reference_pages"]
text_exact_pages = bucket["text_exact_pages"]
metric_values = bucket["metric_values"]
expected_by_key[key] = {
"scored_pages": bucket["scored_pages"],
"text_codepoint_edits": cp_edits,
"text_codepoint_ref_len": cp_len,
"text_cer_codepoint_micro": cp_edits / max(cp_len, 1),
"text_grapheme_edits": gr_edits,
"text_grapheme_ref_len": gr_len,
"text_cer_grapheme_micro": gr_edits / max(gr_len, 1),
"text_reference_pages": text_reference_pages,
"text_exact_pages": text_exact_pages,
"text_exact_accuracy": text_exact_pages / max(text_reference_pages, 1),
"metric_means": {
field: sum(values) / len(values)
for field, values in sorted(metric_values.items())
if values
},
}
recorded_by_key: dict[tuple[str, str, str, str, str], dict[str, Any]] = {}
for index, row in enumerate(slices, 1):
if not isinstance(row, dict):
issues.append(f"page slices[{index}] is not an object")
continue
key = (
str(row.get("language") or ""),
str(row.get("script") or ""),
str(row.get("domain") or ""),
str(row.get("gt_status") or ""),
str(row.get("split") or ""),
)
if key in recorded_by_key:
issues.append(f"duplicate page slice metric for {key}")
recorded_by_key[key] = row
expected_keys = set(expected_by_key)
recorded_keys = set(recorded_by_key)
missing = sorted(expected_keys - recorded_keys)
extra = sorted(recorded_keys - expected_keys)
if missing:
issues.append(f"page slice metrics missing slices: {missing[:10]}")
if extra:
issues.append(f"page slice metrics contain extra slices: {extra[:10]}")
for key in sorted(expected_keys & recorded_keys):
expected = expected_by_key[key]
recorded = recorded_by_key[key]
for field in (
"scored_pages",
"text_codepoint_edits",
"text_codepoint_ref_len",
"text_grapheme_edits",
"text_grapheme_ref_len",
"text_reference_pages",
"text_exact_pages",
):
if recorded.get(field) != expected[field]:
issues.append(f"page slice {key} {field} mismatch: recorded={recorded.get(field)} recomputed={expected[field]}")
for field in ("text_cer_codepoint_micro", "text_cer_grapheme_micro", "text_exact_accuracy"):
value = recorded.get(field)
if not _is_number(value) or not _close(float(value), float(expected[field])):
issues.append(f"page slice {key} {field} mismatch: recorded={value} recomputed={expected[field]}")
expected_means = expected["metric_means"]
recorded_means = recorded.get("metric_means")
if not isinstance(recorded_means, dict):
issues.append(f"page slice {key} metric_means missing or not an object")
continue
missing_means = sorted(set(expected_means) - set(recorded_means))
extra_means = sorted(set(recorded_means) - set(expected_means))
if missing_means:
issues.append(f"page slice {key} metric_means missing metrics: {missing_means[:10]}")
if extra_means:
issues.append(f"page slice {key} metric_means contain extra metrics: {extra_means[:10]}")
for field in sorted(set(expected_means) & set(recorded_means)):
value = recorded_means.get(field)
expected_value = expected_means[field]
if not _is_number(value) or not _close(float(value), expected_value):
issues.append(f"page slice {key} metric_means.{field} mismatch: recorded={value} recomputed={expected_value}")
return issues
def _normalizer_default_counts(payload: dict[str, Any]) -> dict[str, int]:
raw_counts = payload.get("normalizer_counts")
if not isinstance(raw_counts, dict):
return {}
defaults: dict[str, int] = {}
for key in (
"defaulted_language",
"defaulted_script",
"defaulted_domain",
"defaulted_gt_status",
"defaulted_split",
"defaulted_claim_eligible",
):
value = raw_counts.get(key, 0) or 0
if isinstance(value, bool):
continue
try:
count = int(value)
except (TypeError, ValueError):
defaults[key] = 1
continue
if count:
defaults[key] = count
return defaults
def _recognizer_summary_counts(eval_run: dict[str, Any]) -> tuple[dict[str, Any], list[str]]:
summary, issues = _read_summary_for_audit(eval_run, "recognizer", "normalizer_counts audit")
if issues or not summary:
return {}, issues
counts = summary.get("counts")
if counts is None:
return {}, []
if not isinstance(counts, dict):
summary_file = eval_run.get("summary_file")
return {}, [f"recognizer summary counts is not an object: {summary_file}"]
return counts, []
def _check_recognizer_summary_counts(eval_run: dict[str, Any], metrics: dict[str, Any]) -> list[str]:
summary_counts, issues = _recognizer_summary_counts(eval_run)
if issues or not summary_counts:
return issues
metric_counts = metrics.get("normalizer_counts")
run_counts = eval_run.get("normalizer_counts")
if metric_counts != summary_counts:
issues.append("metrics.json normalizer_counts do not match recognizer summary counts")
if run_counts != summary_counts:
issues.append("eval-run.json normalizer_counts do not match recognizer summary counts")
return issues
def _audit_recognizer(score_dir: Path) -> tuple[list[str], dict[str, Any], dict[str, Any]]:
issues = _check_required_artifacts(score_dir, RECOGNIZER_ARTIFACTS)
eval_run = _read_json(score_dir / "eval-run.json")
metrics = _read_json(score_dir / "metrics.json")
if not isinstance(eval_run, dict) or not isinstance(metrics, dict):
return ["recognizer eval artifacts must contain JSON objects"], {}, {}
issues.extend(_check_required_fields(eval_run, RECOGNIZER_REQUIRED_RUN_FIELDS, "eval-run.json"))
issues.extend(_check_common(score_dir, eval_run, metrics))
issues.extend(
_check_eval_run_identity_lists(
eval_run,
_tsv_rows(score_dir / "predictions.tsv"),
run_label="eval-run.json",
row_label="predictions.tsv",
)
)
issues.extend(_check_recognizer_metadata_contradictions(score_dir, metrics, eval_run))
issues.extend(_check_recognizer_provenance_coverage(score_dir, metrics, eval_run))
issues.extend(_check_recognizer_claim_blockers(eval_run, metrics))
if eval_run.get("eval_mode") != "recognizer_oracle_crop":
issues.append(f"unexpected recognizer eval_mode: {eval_run.get('eval_mode')}")
for path_field, hash_field, label in (
("label_file", "label_file_sha256", "recognizer label file"),
("prediction_file", "prediction_file_sha256", "recognizer prediction file"),
("manifest_file", "manifest_file_sha256", "recognizer manifest file"),
("summary_file", "summary_file_sha256", "recognizer summary file"),
("script_inventory_file", "script_inventory_file_sha256", "recognizer script inventory file"),
("spotcheck_calibration_file", "spotcheck_calibration_file_sha256", "recognizer spotcheck calibration file"),
):
issues.extend(_check_hash_field(eval_run, path_field, hash_field, label))
issues.extend(_check_summary_metadata_consistency(eval_run, "recognizer"))
issues.extend(_check_spotcheck_calibration_summary(eval_run, metrics, "recognizer"))
issues.extend(_check_artifact_identity(eval_run, ("checkpoint", "config", "dictionary"), "recognizer"))
issues.extend(_check_train_manifest_hashes(eval_run))
issues.extend(_check_recognizer_leakage_report(score_dir, metrics, eval_run))
issues.extend(_check_recognizer_output_duplicates(score_dir))
issues.extend(_check_recognizer_prediction_source_rows(score_dir, eval_run))
issues.extend(_check_recognizer_pack_content_digest(eval_run, metrics))
issues.extend(_check_recognizer_row_claim_eligibility_blockers(score_dir, eval_run))
issues.extend(_check_recognizer_unmatched_artifact(score_dir, eval_run))
issues.extend(_check_recognizer_match_count_metrics(metrics, eval_run))
issues.extend(_check_recognizer_row_hashes(score_dir, eval_run))
issues.extend(_check_recognizer_predictions_recompute(score_dir, metrics, eval_run))
issues.extend(
_check_normalization_summary(
metrics.get("normalization_summary"),
_expected_recognizer_normalization_summary(score_dir),
"recognizer metrics",
)
)
issues.extend(_check_recognizer_uncertainty(score_dir, metrics))
issues.extend(_check_recognizer_script_inventory(score_dir, metrics, eval_run))
issues.extend(_check_recognizer_dictionary_coverage(score_dir, metrics, eval_run))
issues.extend(_check_recognizer_confusions(score_dir))
issues.extend(_check_recognizer_worst_artifacts(score_dir))
issues.extend(_check_ratio(metrics, "codepoint_edits", "codepoint_ref_len", "cer_codepoint_micro"))
issues.extend(_check_ratio(metrics, "grapheme_edits", "grapheme_ref_len", "cer_grapheme_micro"))
if metrics.get("n_lines") != metrics.get("matched_lines"):
issues.append("recognizer n_lines and matched_lines disagree")
if metrics.get("matched_lines") != _artifact_rows(score_dir / "predictions.tsv"):
issues.append("predictions.tsv row count does not match matched_lines")
unmatched_expected = int(metrics.get("unmatched_labels", 0)) + int(metrics.get("unmatched_predictions", 0))
if unmatched_expected != _artifact_rows(score_dir / "unmatched.tsv"):
issues.append("unmatched.tsv row count does not match unmatched counts")
coverage = metrics.get("provenance_coverage")
if not isinstance(coverage, dict):
issues.append("recognizer provenance coverage is missing")
run_default_counts = _normalizer_default_counts(eval_run)
metric_default_counts = _normalizer_default_counts(metrics)
if run_default_counts:
warnings = eval_run.get("warnings") if isinstance(eval_run.get("warnings"), list) else []
for key in sorted(run_default_counts):
expected = f"normalizer_{key}:"
if not any(str(warning).startswith(expected) for warning in warnings):
issues.append(f"recognizer normalizer default count missing warning: {key}")
if eval_run.get("claim_safe") is True:
issues.append("recognizer claim_safe is true despite normalizer-defaulted metadata")
if run_default_counts != metric_default_counts:
issues.append("recognizer normalizer_counts disagree between eval-run.json and metrics.json")
issues.extend(_check_recognizer_summary_counts(eval_run, metrics))
leakage = metrics.get("leakage")
if not isinstance(leakage, dict) or not leakage.get("checked"):
issues.append("recognizer train/dev leakage check was not run")
return issues, eval_run, metrics
def _check_materialization_records(payload: dict[str, Any], materialize_dir: Path | None = None) -> list[str]:
issues: list[str] = []
mode = payload.get("mode")
records = payload.get("records")
if not isinstance(records, list) or not records:
return ["materialization records are missing"]
record_count = payload.get("record_count")
if not isinstance(record_count, int) or isinstance(record_count, bool):
issues.append(f"materialization record_count is missing or not an integer: {record_count!r}")
elif record_count != len(records):
issues.append(f"materialization record_count does not match records length: recorded={record_count} actual={len(records)}")
for index, record in enumerate(records):
if not isinstance(record, dict):
issues.append(f"materialization records[{index}] is not an object")
continue
command = record.get("command")
if not isinstance(command, list) or not command or not all(isinstance(part, str) and part for part in command):
issues.append(f"materialization records[{index}] command is missing or invalid")
else:
command_template = payload.get("command_template")
if isinstance(command_template, str) and command_template.strip() and materialize_dir is not None:
values: dict[str, str] | None = None
if mode == "recognizer":
label_file = payload.get("label_file")
data_dir = payload.get("data_dir")
prediction_file = payload.get("prediction_file")
if all(isinstance(value, str) and value for value in (label_file, data_dir, prediction_file)):
values = {
"label_file": str(label_file),
"data_dir": str(data_dir),
"prediction_file": str(prediction_file),
"out_dir": str(materialize_dir),
}
else:
issues.append(f"materialization records[{index}] missing recognizer fields required to replay command_template")
elif mode == "page":
output_path = record.get("output_path")
input_path = record.get("input_path")
sample_id = record.get("sample_id")
documents_dir = payload.get("documents_dir")
if all(isinstance(value, str) and value for value in (output_path, input_path, sample_id, documents_dir)):
values = {
"input_path": str(input_path),
"sample_id": str(sample_id),
"output_dir": str(Path(str(output_path)).parent),
"document_json": str(output_path),
"documents_dir": str(documents_dir),
"out_dir": str(materialize_dir),
}
else:
issues.append(f"materialization records[{index}] missing page fields required to replay command_template")
if values is not None:
try:
expected_command = shlex.split(command_template.format(**values))
except (KeyError, ValueError) as exc:
issues.append(f"materialization records[{index}] command_template could not be replayed: {exc}")
else:
if command != expected_command:
issues.append(f"materialization records[{index}] command does not match materialization command_template")
for log_field in ("stdout_path", "stderr_path"):
log_path = record.get(log_field)
sha_field = log_field.replace("_path", "_sha256")
log_sha256 = record.get(sha_field)
if not isinstance(log_path, str) or not log_path:
issues.append(f"materialization records[{index}] missing {log_field}")
elif not Path(log_path).exists():
issues.append(f"materialization records[{index}] {log_field} does not exist: {log_path}")
elif isinstance(log_sha256, str) and log_sha256:
actual_log_sha256 = _sha256_file(Path(log_path))
if actual_log_sha256 != log_sha256:
issues.append(
f"materialization records[{index}] {sha_field} mismatch: "
f"recorded={log_sha256} actual={actual_log_sha256}"
)
else:
issues.append(f"materialization records[{index}] missing {sha_field}")
output_path = record.get("output_path")
output_sha256 = record.get("output_sha256")
returncode = record.get("returncode")
if returncode != 0:
issues.append(f"materialization records[{index}] returncode is not zero: {returncode}")
if not isinstance(output_path, str) or not output_path:
issues.append(f"materialization records[{index}] missing output_path")
continue
if not isinstance(output_sha256, str) or not output_sha256:
issues.append(f"materialization records[{index}] missing output_sha256")
continue
path = Path(output_path)
if not path.exists():
issues.append(f"materialization records[{index}] output_path does not exist: {output_path}")
continue
actual = _sha256_file(path)
if actual != output_sha256:
issues.append(f"materialization records[{index}] output_sha256 mismatch: recorded={output_sha256} actual={actual}")
if mode == "page":
input_path = record.get("input_path")
input_sha256 = record.get("input_sha256")
if not isinstance(input_path, str) or not input_path:
issues.append(f"materialization records[{index}] missing input_path for page mode")
continue
if not isinstance(input_sha256, str) or not input_sha256:
issues.append(f"materialization records[{index}] missing input_sha256 for page mode")
continue
source_path = Path(input_path)
if not source_path.exists():
issues.append(f"materialization records[{index}] input_path does not exist: {input_path}")
continue
actual_input_sha256 = _sha256_file(source_path)
if actual_input_sha256 != input_sha256:
issues.append(
f"materialization records[{index}] input_sha256 mismatch: "
f"recorded={input_sha256} actual={actual_input_sha256}"
)
return issues
def _check_materialization_provenance(payload: dict[str, Any]) -> list[str]:
issues: list[str] = []
issues.extend(_check_timestamp(payload, "created_at", "materialization run"))
for field in ("mode", "system_name", "command_template", "invocation"):
value = payload.get(field)
if not isinstance(value, str) or not value:
issues.append(f"materialization missing required field: {field}")
issues.extend(_check_machine_metadata(payload, "materialization"))
return issues
def _same_path(left: Any, right: Path) -> bool:
if not isinstance(left, str) or not left:
return False
try:
return Path(left).resolve() == right.resolve()
except OSError:
return False
def _is_relative_to(path: Path, parent: Path) -> bool:
try:
path.resolve().relative_to(parent.resolve())
except ValueError:
return False
except OSError:
return False
return True
def _check_recipe_context_identity(recipe_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
root_value = recipe_run.get("root")
recipe_file_value = recipe_run.get("recipe_file")
if not isinstance(root_value, str) or not root_value:
return ["recipe-run.json root is missing"]
root_path = Path(root_value)
if not root_path.exists() or not root_path.is_dir():
issues.append(f"recipe-run.json root does not exist or is not a directory: {root_value}")
return issues
if not isinstance(recipe_file_value, str) or not recipe_file_value:
return issues
recipe_file_path = Path(recipe_file_value)
if recipe_file_path.exists() and not _is_relative_to(recipe_file_path, root_path):
issues.append("recipe-run.json recipe_file is not under declared root")
return issues
def _check_recipe_output_lineage(run_dir: Path, recipe_run: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
outputs = recipe_run.get("outputs")
materialization = recipe_run.get("materialization")
if not isinstance(outputs, dict) or not isinstance(materialization, dict):
return issues
materialization_file = run_dir / "materialize" / "output-command-run.json"
if materialization_file.exists():
try:
materialization_payload = _read_json(materialization_file)
except (FileNotFoundError, ValueError) as exc:
issues.append(str(exc))
else:
if materialization_payload != materialization:
issues.append("materialize/output-command-run.json does not match recipe-run.json materialization")
command_template = materialization.get("command_template")
command_template_file = run_dir / "materialize" / "command-template.txt"
if command_template_file.exists() and isinstance(command_template, str):
if command_template_file.read_text(encoding="utf-8").strip() != command_template.strip():
issues.append("materialize/command-template.txt does not match recorded materialization command_template")
report_file = run_dir / "materialize" / "output-command-run.md"
records = materialization.get("records")
if report_file.exists() and isinstance(records, list):
report_text = report_file.read_text(encoding="utf-8")
expected_record_line = f"- records: {len(records)}"
if expected_record_line not in report_text:
issues.append(f"materialize/output-command-run.md missing or stale line: {expected_record_line}")
for index, record in enumerate(records):
if not isinstance(record, dict):
continue
output_path = record.get("output_path")
output_sha256 = record.get("output_sha256")
input_path = record.get("input_path")
input_sha256 = record.get("input_sha256")
sample_id = record.get("sample_id")
if isinstance(output_path, str) and isinstance(output_sha256, str) and isinstance(sample_id, str):
if isinstance(input_path, str) and isinstance(input_sha256, str):
expected_output_line = (
f"- {sample_id}: input `{input_path}` sha256={input_sha256}; "
f"output `{output_path}` sha256={output_sha256}"
)
else:
expected_output_line = f"- {sample_id}: `{output_path}` sha256={output_sha256}"
if expected_output_line not in report_text:
issues.append(f"materialize/output-command-run.md missing or stale output line for record {index}")
stdout_path = record.get("stdout_path")
stdout_sha256 = record.get("stdout_sha256")
if isinstance(stdout_path, str) and isinstance(stdout_sha256, str):
expected_stdout = f"stdout `{stdout_path}` sha256={stdout_sha256}"
if expected_stdout not in report_text:
issues.append(f"materialize/output-command-run.md missing or stale stdout line for record {index}")
stderr_path = record.get("stderr_path")
stderr_sha256 = record.get("stderr_sha256")
if isinstance(stderr_path, str) and isinstance(stderr_sha256, str):
expected_stderr = f"stderr `{stderr_path}` sha256={stderr_sha256}"
if expected_stderr not in report_text:
issues.append(f"materialize/output-command-run.md missing or stale stderr line for record {index}")
if eval_run.get("eval_mode") == "page_level_ocr":
documents_value = outputs.get("documents_dir")
eval_documents_dir = eval_run.get("documents_dir")
if isinstance(documents_value, str) and documents_value:
if isinstance(eval_documents_dir, str) and eval_documents_dir and not _same_path(documents_value, Path(eval_documents_dir)):
issues.append("recipe-run.json outputs.documents_dir does not match score eval-run documents_dir")
if not Path(documents_value).exists():
issues.append(f"recipe-run.json output path does not exist for documents_dir: {documents_value}")
if isinstance(records, list):
for index, record in enumerate(records):
if not isinstance(record, dict):
continue
output_path = record.get("output_path")
if isinstance(output_path, str) and output_path and not _is_relative_to(Path(output_path), Path(documents_value)):
issues.append(f"recipe-run.json materialization record {index} output_path is not under outputs.documents_dir")
return issues
prediction_value = outputs.get("prediction_file")
eval_prediction = eval_run.get("prediction_file")
eval_prediction_sha256 = eval_run.get("prediction_file_sha256")
if isinstance(prediction_value, str) and prediction_value:
prediction_path = Path(prediction_value)
if isinstance(eval_prediction, str) and eval_prediction and not _same_path(prediction_value, Path(eval_prediction)):
issues.append("recipe-run.json outputs.prediction_file does not match score eval-run prediction_file")
if isinstance(eval_prediction_sha256, str) and eval_prediction_sha256 and prediction_path.exists():
actual_sha256 = _sha256_file(prediction_path)
if actual_sha256 != eval_prediction_sha256:
issues.append(
"recipe-run.json outputs.prediction_file sha256 does not match score eval-run prediction_file_sha256: "
f"recorded={eval_prediction_sha256} actual={actual_sha256}"
)
if isinstance(records, list) and isinstance(prediction_value, str) and prediction_value:
matching_records = [
record
for record in records
if isinstance(record, dict) and _same_path(record.get("output_path"), Path(prediction_value))
]
if not matching_records:
issues.append("recipe-run.json materialization records do not include outputs.prediction_file")
elif isinstance(eval_prediction_sha256, str) and eval_prediction_sha256:
for index, record in enumerate(matching_records):
if record.get("output_sha256") != eval_prediction_sha256:
issues.append(
"recipe-run.json materialization prediction output_sha256 does not match score eval-run "
f"prediction_file_sha256 for matching record {index}: "
f"recorded={record.get('output_sha256')} score={eval_prediction_sha256}"
)
return issues
def _check_recipe_output_audit_consistency(run_dir: Path, recipe_run: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
if eval_run.get("eval_mode") != "page_level_ocr":
return []
output_audit = recipe_run.get("output_audit")
if output_audit is None:
return []
issues: list[str] = []
if not isinstance(output_audit, dict):
return ["recipe-run.json output_audit is not an object"]
outputs = recipe_run.get("outputs")
if isinstance(outputs, dict):
output_audit_dir = outputs.get("output_audit_dir")
if not isinstance(output_audit_dir, str) or not output_audit_dir:
issues.append("recipe-run.json outputs.output_audit_dir is missing for page output audit")
elif not _same_path(output_audit_dir, run_dir / "output-audit"):
issues.append("recipe-run.json outputs.output_audit_dir does not point at root output-audit")
elif not Path(output_audit_dir).is_dir():
issues.append(f"recipe-run.json output path does not exist for output_audit_dir: {output_audit_dir}")
output_audit_path = run_dir / "output-audit" / "output-audit-run.json"
try:
output_audit_payload = _read_json(output_audit_path)
except (FileNotFoundError, ValueError) as exc:
issues.append(str(exc))
else:
if output_audit_payload != output_audit:
issues.append("output-audit/output-audit-run.json does not match recipe-run.json output_audit")
command_template = output_audit.get("command_template")
command_template_file = run_dir / "output-audit" / "command-template.txt"
if not isinstance(command_template, str) or not command_template.strip():
issues.append("recipe-run.json output_audit command_template is missing")
elif not command_template_file.is_file():
issues.append("missing required artifact: output-audit/command-template.txt")
else:
if command_template_file.read_text(encoding="utf-8").strip() != command_template.strip():
issues.append("output-audit/command-template.txt does not match recorded output_audit command_template")
records = output_audit.get("records")
record_count = output_audit.get("record_count")
if not isinstance(records, list):
issues.append("recipe-run.json output_audit records is not a list")
records = []
elif record_count != len(records):
issues.append(f"recipe-run.json output_audit record_count mismatch: recorded={record_count} actual={len(records)}")
if output_audit.get("passed") is not True:
issues.append("recipe-run.json output_audit did not pass")
for index, record in enumerate(records):
if not isinstance(record, dict):
issues.append(f"recipe-run.json output_audit records[{index}] is not an object")
continue
if record.get("returncode") != 0:
issues.append(f"recipe-run.json output_audit records[{index}] returncode is not zero: {record.get('returncode')}")
command = record.get("command")
output_dir = record.get("output_dir")
if not isinstance(command, list) or not command or not all(isinstance(part, str) and part for part in command):
issues.append(f"recipe-run.json output_audit records[{index}] command is missing or invalid")
elif isinstance(command_template, str) and command_template.strip():
if not isinstance(output_dir, str) or not output_dir:
issues.append(f"recipe-run.json output_audit records[{index}] missing output_dir")
else:
try:
expected_command = shlex.split(
command_template.format(
output_dir=output_dir,
documents_dir=str(run_dir / "documents"),
out_dir=str(run_dir / "output-audit"),
)
)
except (KeyError, ValueError) as exc:
issues.append(f"recipe-run.json output_audit records[{index}] command_template could not be replayed: {exc}")
else:
if command != expected_command:
issues.append(f"recipe-run.json output_audit records[{index}] command does not match output_audit command_template")
for path_key in ("stdout_path", "stderr_path"):
path_value = record.get(path_key)
sha_key = path_key.replace("_path", "_sha256")
sha_value = record.get(sha_key)
if not isinstance(path_value, str) or not path_value:
issues.append(f"recipe-run.json output_audit records[{index}] missing {path_key}")
elif not Path(path_value).is_file():
issues.append(f"recipe-run.json output_audit records[{index}] {path_key} does not exist: {path_value}")
elif isinstance(sha_value, str) and sha_value:
actual_log_sha256 = _sha256_file(Path(path_value))
if actual_log_sha256 != sha_value:
issues.append(
f"recipe-run.json output_audit records[{index}] {sha_key} mismatch: "
f"recorded={sha_value} actual={actual_log_sha256}"
)
else:
issues.append(f"recipe-run.json output_audit records[{index}] missing {sha_key}")
audit_json = record.get("audit_json")
if not isinstance(audit_json, dict):
issues.append(f"recipe-run.json output_audit records[{index}] missing audit_json artifact")
else:
audit_json_path = audit_json.get("path")
audit_json_sha256 = audit_json.get("sha256")
if not isinstance(audit_json_path, str) or not audit_json_path:
issues.append(f"recipe-run.json output_audit records[{index}] audit_json missing path")
elif not Path(audit_json_path).is_file():
issues.append(f"recipe-run.json output_audit records[{index}] audit_json path does not exist: {audit_json_path}")
elif isinstance(audit_json_sha256, str) and audit_json_sha256:
actual_audit_json_sha256 = _sha256_file(Path(audit_json_path))
if actual_audit_json_sha256 != audit_json_sha256:
issues.append(
f"recipe-run.json output_audit records[{index}] audit_json sha256 mismatch: "
f"recorded={audit_json_sha256} actual={actual_audit_json_sha256}"
)
else:
issues.append(f"recipe-run.json output_audit records[{index}] audit_json missing sha256")
audit_markdown = record.get("audit_markdown")
if audit_markdown is not None:
if not isinstance(audit_markdown, dict):
issues.append(f"recipe-run.json output_audit records[{index}] audit_markdown is not an object")
else:
audit_markdown_path = audit_markdown.get("path")
audit_markdown_sha256 = audit_markdown.get("sha256")
if not isinstance(audit_markdown_path, str) or not audit_markdown_path:
issues.append(f"recipe-run.json output_audit records[{index}] audit_markdown missing path")
elif not Path(audit_markdown_path).is_file():
issues.append(f"recipe-run.json output_audit records[{index}] audit_markdown path does not exist: {audit_markdown_path}")
elif isinstance(audit_markdown_sha256, str) and audit_markdown_sha256:
actual_audit_markdown_sha256 = _sha256_file(Path(audit_markdown_path))
if actual_audit_markdown_sha256 != audit_markdown_sha256:
issues.append(
f"recipe-run.json output_audit records[{index}] audit_markdown sha256 mismatch: "
f"recorded={audit_markdown_sha256} actual={actual_audit_markdown_sha256}"
)
else:
issues.append(f"recipe-run.json output_audit records[{index}] audit_markdown missing sha256")
report_path = run_dir / "output-audit" / "output-audit-run.md"
if not report_path.is_file():
issues.append("missing required artifact: output-audit/output-audit-run.md")
else:
report_text = report_path.read_text(encoding="utf-8")
expected_record_line = f"- records: {len(records)}"
if expected_record_line not in report_text:
issues.append(f"output-audit/output-audit-run.md missing or stale line: {expected_record_line}")
for index, record in enumerate(records):
if not isinstance(record, dict):
continue
sample_id = record.get("sample_id")
output_dir = record.get("output_dir")
returncode = record.get("returncode")
if isinstance(sample_id, str) and isinstance(output_dir, str):
expected_output_line = f"- {sample_id}: returncode={returncode} output_dir=`{output_dir}`"
if expected_output_line not in report_text:
issues.append(f"output-audit/output-audit-run.md missing or stale output line for record {index}")
if isinstance(record.get("stdout_path"), str) and isinstance(record.get("stdout_sha256"), str):
expected_stdout = f"stdout=`{record['stdout_path']}` sha256={record['stdout_sha256']}"
if expected_stdout not in report_text:
issues.append(f"output-audit/output-audit-run.md missing or stale stdout line for record {index}")
if isinstance(record.get("stderr_path"), str) and isinstance(record.get("stderr_sha256"), str):
expected_stderr = f"stderr=`{record['stderr_path']}` sha256={record['stderr_sha256']}"
if expected_stderr not in report_text:
issues.append(f"output-audit/output-audit-run.md missing or stale stderr line for record {index}")
audit_json = record.get("audit_json")
if isinstance(audit_json, dict) and isinstance(audit_json.get("path"), str) and isinstance(audit_json.get("sha256"), str):
expected_audit_json = f"audit_json=`{audit_json['path']}` sha256={audit_json['sha256']}"
if expected_audit_json not in report_text:
issues.append(f"output-audit/output-audit-run.md missing or stale audit JSON line for record {index}")
audit_markdown = record.get("audit_markdown")
if (
isinstance(audit_markdown, dict)
and isinstance(audit_markdown.get("path"), str)
and isinstance(audit_markdown.get("sha256"), str)
):
expected_audit_markdown = f"audit_markdown=`{audit_markdown['path']}` sha256={audit_markdown['sha256']}"
if expected_audit_markdown not in report_text:
issues.append(f"output-audit/output-audit-run.md missing or stale audit Markdown line for record {index}")
return issues
def _check_recipe_materialization_score_identity(recipe_run: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
materialization = recipe_run.get("materialization")
if not isinstance(materialization, dict):
return []
issues: list[str] = []
recipe_id = recipe_run.get("recipe_id")
if isinstance(recipe_id, str) and recipe_id:
system_name = materialization.get("system_name")
if system_name != recipe_id:
issues.append(f"recipe-run.json materialization system_name does not match recipe_id: recorded={system_name} recipe_id={recipe_id}")
for field in ("model_name", "checkpoint", "config"):
recorded = materialization.get(field)
scored = eval_run.get(field)
if recorded != scored:
issues.append(f"recipe-run.json materialization {field} does not match score eval-run {field}: recorded={recorded} score={scored}")
recorded = materialization.get("dictionary")
scored = eval_run.get("dictionary")
if recorded != scored:
issues.append(f"recipe-run.json materialization dictionary does not match score eval-run dictionary: recorded={recorded} score={scored}")
return issues
def _check_recipe_score_artifact_identity(score_dir: Path, score: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
eval_run_name = "page-eval-run.json" if eval_run.get("eval_mode") == "page_level_ocr" else "eval-run.json"
for key, expected_path in (("metrics_file", score_dir / "metrics.json"), ("eval_run_file", score_dir / eval_run_name)):
value = score.get(key)
if not isinstance(value, str) or not value:
issues.append(f"recipe-run.json score missing {key}")
elif not _same_path(value, expected_path):
issues.append(f"recipe-run.json score {key} does not point at score/{expected_path.name}")
hash_key = f"{key}_sha256"
recorded_hash = score.get(hash_key)
if not isinstance(recorded_hash, str) or not recorded_hash:
issues.append(f"recipe-run.json score missing {hash_key}")
elif expected_path.exists():
actual_hash = _sha256_file(expected_path)
if actual_hash != recorded_hash:
issues.append(
f"recipe-run.json score {hash_key} does not match score/{expected_path.name}: "
f"recorded={recorded_hash} actual={actual_hash}"
)
return issues
def _check_recipe_train_manifest_identity(recipe_run: dict[str, Any], eval_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
recorded = recipe_run.get("train_manifests")
scored = eval_run.get("train_manifests")
if recorded is None:
recorded = []
if scored is None:
scored = []
if not isinstance(recorded, list):
return ["recipe-run.json train_manifests is not a list"]
if not isinstance(scored, list):
return ["eval-run.json train_manifests is not a list for recipe-root audit"]
if len(recorded) != len(scored):
return [f"recipe-run.json train_manifests count does not match eval-run.json: recorded={len(recorded)} score={len(scored)}"]
for index, (recorded_item, scored_item) in enumerate(zip(recorded, scored, strict=True)):
if not isinstance(recorded_item, dict):
issues.append(f"recipe-run.json train_manifests[{index}] is not an object")
continue
if not isinstance(scored_item, dict):
issues.append(f"eval-run.json train_manifests[{index}] is not an object for recipe-root audit")
continue
recorded_path = recorded_item.get("path")
scored_path = scored_item.get("path")
if not isinstance(recorded_path, str) or not recorded_path:
issues.append(f"recipe-run.json train_manifests[{index}] missing path")
elif not _same_path(recorded_path, Path(str(scored_path))):
issues.append(f"recipe-run.json train_manifests[{index}] path does not match eval-run.json")
recorded_sha = recorded_item.get("sha256")
scored_sha = scored_item.get("sha256")
if not isinstance(recorded_sha, str) or not recorded_sha:
issues.append(f"recipe-run.json train_manifests[{index}] missing sha256")
elif recorded_sha != scored_sha:
issues.append(f"recipe-run.json train_manifests[{index}] sha256 does not match eval-run.json")
return issues
def _current_recipe_validation(registry_path: Path, root_path: Path, recipe_id: str) -> dict[str, Any]:
validator = _load_tool("validate_eval_harness_recipes")
validation = validator.validate_recipes(registry_path, root_path)
for recipe_validation in validation.get("recipes", []):
if isinstance(recipe_validation, dict) and recipe_validation.get("id") == recipe_id:
return recipe_validation
raise ValueError(f"recipe validator result missing selected recipe: {recipe_id}")
def _check_recipe_artifact_consistency(run_dir: Path, recipe_run: dict[str, Any]) -> list[str]:
issues: list[str] = []
outputs = recipe_run.get("outputs")
if not isinstance(outputs, dict):
return issues
recipe_id = recipe_run.get("recipe_id")
recipe_path_value = outputs.get("recipe_file")
validation_path_value = outputs.get("recipe_validation_file")
local_recipe: dict[str, Any] | None = None
if isinstance(recipe_path_value, str) and recipe_path_value:
try:
payload = _read_json(Path(recipe_path_value))
except (FileNotFoundError, ValueError) as exc:
issues.append(str(exc))
else:
if isinstance(payload, dict):
local_recipe = payload
if isinstance(recipe_id, str) and recipe_id and payload.get("id") != recipe_id:
issues.append(
"recipe.json id does not match recipe-run.json recipe_id: "
f"recipe.json={payload.get('id')} recipe-run.json={recipe_id}"
)
else:
issues.append("recipe.json must contain a JSON object")
if isinstance(validation_path_value, str) and validation_path_value:
try:
validation_payload = _read_json(Path(validation_path_value))
except (FileNotFoundError, ValueError) as exc:
issues.append(str(exc))
else:
if validation_payload != recipe_run.get("recipe_validation"):
issues.append("recipe-validation.json does not match recipe-run.json recipe_validation")
registry_path_value = recipe_run.get("recipe_file")
if isinstance(registry_path_value, str) and registry_path_value and isinstance(recipe_id, str) and recipe_id and local_recipe:
registry_path = Path(registry_path_value)
try:
registry = _read_json(registry_path)
except (FileNotFoundError, ValueError) as exc:
issues.append(str(exc))
else:
recipes = registry.get("recipes") if isinstance(registry, dict) else None
if not isinstance(recipes, list):
issues.append("recipe registry file must contain a recipes list")
else:
matches = [recipe for recipe in recipes if isinstance(recipe, dict) and recipe.get("id") == recipe_id]
if not matches:
issues.append(f"recipe registry file does not contain recipe_id: {recipe_id}")
elif len(matches) > 1:
issues.append(f"recipe registry file contains duplicate recipe_id: {recipe_id}")
elif matches[0] != local_recipe:
issues.append("recipe.json does not match selected recipe in recipe registry file")
root_value = recipe_run.get("root")
if isinstance(root_value, str) and root_value:
try:
current_validation = _current_recipe_validation(registry_path, Path(root_value), recipe_id)
except Exception as exc:
issues.append(f"recipe validation recomputation failed: {type(exc).__name__}: {exc}")
else:
if current_validation != recipe_run.get("recipe_validation"):
issues.append("recipe-run.json recipe_validation does not match current recipe validator result")
if isinstance(recipe_path_value, str) and recipe_path_value and not _same_path(recipe_path_value, run_dir / "recipe.json"):
issues.append("recipe-run.json outputs.recipe_file does not point at root recipe.json")
if isinstance(validation_path_value, str) and validation_path_value and not _same_path(
validation_path_value, run_dir / "recipe-validation.json"
):
issues.append("recipe-run.json outputs.recipe_validation_file does not point at root recipe-validation.json")
return issues
def _audit_recipe_root(run_dir: Path, score_dir: Path, eval_run: dict[str, Any], metrics: dict[str, Any], score_issues: list[str]) -> list[str]:
if run_dir == score_dir:
return []
issues = _check_required_artifacts(run_dir, RECIPE_ROOT_ARTIFACTS)
if eval_run.get("eval_mode") == "recognizer_oracle_crop":
issues.extend(_check_required_artifacts(run_dir, ("predictions.txt",)))
try:
recipe_run = _read_json(run_dir / "recipe-run.json")
except (FileNotFoundError, ValueError) as exc:
return [*issues, str(exc)]
if not isinstance(recipe_run, dict):
return [*issues, "recipe-run.json must contain a JSON object"]
issues.extend(_check_required_fields(recipe_run, RECIPE_REQUIRED_FIELDS, "recipe-run.json"))
issues.extend(_check_timestamp(recipe_run, "created_at", "recipe run"))
issues.extend(_check_machine_metadata(recipe_run, "recipe run"))
issues.extend(_check_recipe_context_identity(recipe_run))
if not isinstance(recipe_run.get("command"), str) or not recipe_run["command"]:
issues.append("recipe run command is missing")
issues.extend(_check_hash_field(recipe_run, "recipe_file", "recipe_file_sha256", "recipe registry file"))
outputs = recipe_run.get("outputs")
if not isinstance(outputs, dict):
issues.append("recipe-run.json outputs is missing or not an object")
else:
required_outputs = [
"materialization_dir",
"score_dir",
"recipe_file",
"recipe_validation_file",
"audit_file",
"claim_safety_report",
]
if eval_run.get("eval_mode") == "page_level_ocr":
required_outputs.append("documents_dir")
else:
required_outputs.append("prediction_file")
for key in required_outputs:
value = outputs.get(key)
if not isinstance(value, str) or not value:
issues.append(f"recipe-run.json outputs missing {key}")
elif not Path(value).exists():
issues.append(f"recipe-run.json output path does not exist for {key}: {value}")
if not _same_path(outputs.get("score_dir"), score_dir):
issues.append("recipe-run.json outputs.score_dir does not match audited score directory")
expected_root_paths = {
"materialization_dir": run_dir / "materialize",
"audit_file": run_dir / "audit.json",
"claim_safety_report": run_dir / "claim-safety.md",
}
for key, expected_path in expected_root_paths.items():
if isinstance(outputs.get(key), str) and not _same_path(outputs.get(key), expected_path):
issues.append(f"recipe-run.json outputs.{key} does not point at root {expected_path.name}")
materialization = recipe_run.get("materialization")
if not isinstance(materialization, dict):
issues.append("recipe-run.json materialization is missing or not an object")
else:
issues.extend(_check_materialization_provenance(materialization))
issues.extend(_check_materialization_records(materialization, run_dir / "materialize"))
issues.extend(_check_recipe_artifact_consistency(run_dir, recipe_run))
issues.extend(_check_recipe_train_manifest_identity(recipe_run, eval_run))
issues.extend(_check_recipe_output_lineage(run_dir, recipe_run, eval_run))
issues.extend(_check_recipe_output_audit_consistency(run_dir, recipe_run, eval_run))
issues.extend(_check_recipe_materialization_score_identity(recipe_run, eval_run))
recipe_audit = recipe_run.get("audit")
if not isinstance(recipe_audit, dict):
issues.append("recipe-run.json audit summary is missing or not an object")
else:
if bool(recipe_audit.get("audit_passed")) != (not score_issues):
issues.append("recipe-run.json audit_passed does not match current score audit")
if bool(recipe_audit.get("claim_safe")) != bool(eval_run.get("claim_safe")):
issues.append("recipe-run.json audit claim_safe does not match eval run")
if recipe_audit.get("issues") != score_issues:
issues.append("recipe-run.json audit issues do not match current score audit")
score = recipe_run.get("score")
if not isinstance(score, dict):
issues.append("recipe-run.json score summary is missing or not an object")
else:
if bool(score.get("claim_safe")) != bool(eval_run.get("claim_safe")):
issues.append("recipe-run.json score claim_safe does not match eval run")
warnings = eval_run.get("warnings") if isinstance(eval_run.get("warnings"), list) else []
if score.get("warnings") != warnings:
issues.append("recipe-run.json score warnings do not match eval run")
issues.extend(_check_recipe_score_artifact_identity(score_dir, score, eval_run))
try:
root_audit = _read_json(run_dir / "audit.json")
except (FileNotFoundError, ValueError) as exc:
issues.append(str(exc))
else:
if not isinstance(root_audit, dict):
issues.append("audit.json must contain a JSON object")
else:
if bool(root_audit.get("audit_passed")) != (not score_issues):
issues.append("audit.json audit_passed does not match current score audit")
if bool(root_audit.get("claim_safe")) != bool(eval_run.get("claim_safe")):
issues.append("audit.json claim_safe does not match eval run")
if root_audit.get("warnings") != (eval_run.get("warnings") if isinstance(eval_run.get("warnings"), list) else []):
issues.append("audit.json warnings do not match eval run")
if root_audit.get("issues") != score_issues:
issues.append("audit.json issues do not match current score audit")
if not _same_path(root_audit.get("score_dir"), score_dir):
issues.append("audit.json score_dir does not match audited score directory")
claim_report = run_dir / "claim-safety.md"
if claim_report.exists():
text = claim_report.read_text(encoding="utf-8")
for marker in ("## What Was Scored", "## Ground Truth And Pack", "## Metrics", "## Warnings"):
if marker not in text:
issues.append(f"root claim-safety.md missing section: {marker}")
issues.extend(_check_root_claim_safety_report_consistency(run_dir, recipe_run, eval_run, metrics))
return issues
def _audit_page(score_dir: Path) -> tuple[list[str], dict[str, Any], dict[str, Any]]:
issues = _check_required_artifacts(score_dir, PAGE_ARTIFACTS)
eval_run = _read_json(score_dir / "page-eval-run.json")
metrics = _read_json(score_dir / "metrics.json")
if not isinstance(eval_run, dict) or not isinstance(metrics, dict):
return ["page eval artifacts must contain JSON objects"], {}, {}
issues.extend(_check_required_fields(eval_run, PAGE_REQUIRED_RUN_FIELDS, "page-eval-run.json"))
issues.extend(_check_common(score_dir, eval_run, metrics))
issues.extend(
_check_eval_run_identity_lists(
eval_run,
_csv_rows(score_dir / "page-metrics.csv"),
run_label="page-eval-run.json",
row_label="page-metrics.csv",
)
)
issues.extend(_check_page_metadata_contradictions(score_dir, metrics, eval_run))
issues.extend(_check_page_metadata_coverage(metrics, eval_run))
issues.extend(_check_page_claim_blockers(eval_run, metrics))
issues.extend(_check_page_output_audit(score_dir, metrics, eval_run))
if eval_run.get("eval_mode") != "page_level_ocr":
issues.append(f"unexpected page eval_mode: {eval_run.get('eval_mode')}")
for path_field, hash_field, label in (
("eval_manifest", "eval_manifest_sha256", "page eval manifest"),
("summary_file", "summary_file_sha256", "page summary file"),
("script_inventory_file", "script_inventory_file_sha256", "page script inventory file"),
("spotcheck_calibration_file", "spotcheck_calibration_file_sha256", "page spotcheck calibration file"),
):
issues.extend(_check_hash_field(eval_run, path_field, hash_field, label))
issues.extend(_check_summary_metadata_consistency(eval_run, "page"))
issues.extend(_check_spotcheck_calibration_summary(eval_run, metrics, "page"))
issues.extend(_check_artifact_identity(eval_run, ("checkpoint", "config", "dictionary"), "page"))
issues.extend(_check_train_manifest_hashes(eval_run))
issues.extend(_check_page_leakage_report(metrics, eval_run))
issues.extend(_check_page_output_duplicates(score_dir))
issues.extend(_check_page_missing_artifact(score_dir, eval_run))
issues.extend(_check_page_metric_identity_rows(score_dir, eval_run))
issues.extend(_check_page_pack_content_digest(eval_run, metrics))
issues.extend(_check_page_document_source_path_mismatches(metrics, eval_run))
issues.extend(_check_page_row_hashes(score_dir))
issues.extend(_check_page_manifest_count_metrics(metrics, eval_run))
issues.extend(_check_page_line_alignment_recompute(score_dir, metrics, eval_run))
issues.extend(_check_page_reference_artifacts(score_dir, eval_run))
issues.extend(_check_page_text_pairs_artifact(score_dir, eval_run))
issues.extend(_check_page_confusions(score_dir))
issues.extend(_check_page_text_metrics_recompute(score_dir, metrics, eval_run))
issues.extend(
_check_normalization_summary(
metrics.get("normalization_summary"),
_expected_page_normalization_summary(score_dir),
"page metrics",
)
)
issues.extend(_check_page_uncertainty(score_dir, metrics))
issues.extend(_check_page_script_inventory(score_dir, metrics, eval_run))
issues.extend(_check_page_dictionary_coverage(score_dir, metrics, eval_run))
issues.extend(_check_page_worst_artifacts(score_dir))
issues.extend(_check_ratio(metrics, "text_codepoint_edits", "text_codepoint_ref_len", "text_cer_codepoint_micro"))
issues.extend(_check_ratio(metrics, "text_grapheme_edits", "text_grapheme_ref_len", "text_cer_grapheme_micro"))
if metrics.get("scored_pages") != _artifact_rows(score_dir / "page-metrics.csv"):
issues.append("page-metrics.csv row count does not match scored_pages")
if int(metrics.get("missing_documents", 0)) + int(metrics.get("missing_references", 0)) != _artifact_rows(score_dir / "missing.tsv"):
issues.append("missing.tsv row count does not match missing document/reference counts")
issues.extend(_check_page_reference_coverage(score_dir, metrics))
coverage = metrics.get("reference_coverage")
if not isinstance(coverage, dict):
issues.append("page reference coverage is missing")
elif int(coverage.get("line_bboxes", 0)) <= 0:
issues.append("page eval has no line-bbox reference coverage")
metadata_coverage = metrics.get("metadata_coverage")
if not isinstance(metadata_coverage, dict):
issues.append("page metadata coverage is missing")
gt_statuses = eval_run.get("gt_statuses")
if isinstance(gt_statuses, list) and "machine_converted_spot_checked" in gt_statuses:
calibration = eval_run.get("spotcheck_calibration")
if not isinstance(calibration, dict):
issues.append("page spot-checked machine GT is missing calibration summary")
elif not calibration.get("claim_evidence_eligible"):
issues.append("page spot-check calibration is not claim evidence eligible")
metric_means = metrics.get("metric_means")
if not isinstance(metric_means, dict):
issues.append("page metric means are missing")
else:
for key in ("line_detection_f1", "line_mean_iou", "reading_order_pair_accuracy"):
if key not in metric_means:
issues.append(f"page metric_means missing {key}")
issues.extend(_check_page_slice_metrics(score_dir, metrics.get("slices")))
leakage = metrics.get("leakage")
if not isinstance(leakage, dict) or not leakage.get("checked"):
issues.append("page train/dev leakage check was not run")
return issues, eval_run, metrics
def audit_run(run_dir: Path) -> dict[str, Any]:
score_dir = _score_dir(run_dir)
mode = _mode(score_dir)
if mode == "recognizer_oracle_crop":
issues, eval_run, metrics = _audit_recognizer(score_dir)
elif mode == "page_level_ocr":
issues, eval_run, metrics = _audit_page(score_dir)
else:
raise ValueError(f"unsupported eval mode: {mode}")
issues.extend(_audit_recipe_root(run_dir, score_dir, eval_run, metrics, issues))
claim_safe = bool(eval_run.get("claim_safe")) if isinstance(eval_run, dict) else False
return {
"run_dir": str(run_dir),
"score_dir": str(score_dir),
"eval_mode": mode,
"audit_passed": not issues,
"claim_safe": claim_safe,
"warnings": eval_run.get("warnings", []) if isinstance(eval_run, dict) else [],
"issues": issues,
"summary": {
"model_name": eval_run.get("model_name") if isinstance(eval_run, dict) else None,
"eval_pack_version": eval_run.get("eval_pack_version") if isinstance(eval_run, dict) else None,
"normalization_profile": eval_run.get("normalization_profile") if isinstance(eval_run, dict) else None,
"n_lines": metrics.get("n_lines") if isinstance(metrics, dict) else None,
"scored_pages": metrics.get("scored_pages") if isinstance(metrics, dict) else None,
},
}
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--run-dir", type=Path, required=True)
parser.add_argument("--out", type=Path)
parser.add_argument("--require-audit-pass", action="store_true")
parser.add_argument("--require-claim-safe", action="store_true")
args = parser.parse_args(argv)
result = audit_run(args.run_dir)
if args.out:
args.out.parent.mkdir(parents=True, exist_ok=True)
args.out.write_text(json.dumps(result, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
print(
f"eval_mode={result['eval_mode']} audit_passed={result['audit_passed']} "
f"claim_safe={result['claim_safe']} issues={len(result['issues'])}"
)
if args.require_audit_pass and not result["audit_passed"]:
raise SystemExit(1)
if args.require_claim_safe and not result["claim_safe"]:
raise SystemExit(1)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|