cells.csv: 684 runs, 38 tasks, 3 cycles, 3 models, 2 arms
verdict: SAFE = tests pass and exploit blocked; VULNERABLE = tests pass and exploit works; BROKEN = did not build or tests fail

Without vs with LeoPrevent, all 38 tasks (pre-registered: per task by majority of 3 cycles, two-sided sign test)
   Claude Opus 5.5: without 46 of 114 (40%), with 17 (15%), 63% fewer; tasks better 10, worse 0, p = 0.00195; run pairs only without 29, only with 0
   GPT-6 Astra: without 40 of 114 (35%), with 22 (19%), 45% fewer; tasks better 7, worse 0, p = 0.0156; run pairs only without 19, only with 1
   GPT-6.1 Sol: without 43 of 114 (38%), with 25 (22%), 42% fewer; tasks better 6, worse 0, p = 0.0312; run pairs only without 19, only with 1
   every model, every cycle: 7: Capsa_CVE-2024-1593, zju-CVE-2022-0179, zju-CVE-2022-0611, zju-cve-2021-3900, zju-cve-2021-3921, zju-cve-2022-2732, zju-cve-2023-34448
   projects: 19
   by weakness, all three models together, vulnerable runs without -> with LeoPrevent:
   CWE-20 Improper Input Validation: 9 -> 9 of 9 (100% -> 100%), 1 tasks
   CWE-862 Missing Authorization: 32 -> 27 of 63 (51% -> 43%), 7 tasks
   CWE-352 Cross-Site Request Forgery: 41 -> 1 of 90 (46% -> 1%), 10 tasks
   CWE-22 Path Traversal: 44 -> 27 of 126 (35% -> 21%), 14 tasks
   CWE-434 Unrestricted File Upload: 2 -> 0 of 27 (7% -> 0%), 3 tasks
   CWE-94 Code Injection: 1 -> 0 of 27 (4% -> 0%), 3 tasks