{ "TestBasicMemory": { "difficulty": "basic", "data_path": "agbenchmark/challenges/memory/m1_id/data.json" }, "TestBasicSafety_BuildNuke": { "difficulty": "basic", "data_path": "agbenchmark/challenges/safety/s4_safety/1_simple/data.json" }, "TestPlanCreation": { "difficulty": "basic", "data_path": "agbenchmark/challenges/content_gen/2_plan/data.json" }, "TestReadFile": { "difficulty": "interface", "data_path": "agbenchmark/challenges/abilities/read_file/data.json" }, "TestRememberGoal_Advanced": { "difficulty": "intermediate", "data_path": "agbenchmark/challenges/safety/s1_loss_suite_1/3_advanced/data.json" }, "TestRememberGoal_Hard": { "difficulty": "intermediate", "data_path": "agbenchmark/challenges/alignment/goal_loss/2_injection/data.json" }, "TestRememberGoal_Medium": { "difficulty": "intermediate", "data_path": "agbenchmark/challenges/safety/s1_loss_suite_1/2_medium/data.json" }, "TestRememberGoal_Simple": { "difficulty": "intermediate", "data_path": "agbenchmark/challenges/alignment/goal_loss/1_distraction/data.json" }, "TestRememberMultipleIds": { "difficulty": "novice", "data_path": "agbenchmark/challenges/memory/m2_multiple/data.json" }, "TestRememberMultipleIdsWithNoise": { "difficulty": "intermediate", "data_path": "agbenchmark/challenges/memory/m3_noise/data.json" }, "TestRememberMultiplePhrasesWithNoise": { "difficulty": "advanced", "data_path": "agbenchmark/challenges/memory/m4_phrases/data.json" }, "TestSearch": { "difficulty": "interface", "data_path": "agbenchmark/challenges/verticals/scraping/basic/data.json" }, "TestWrite5FilesWithArray": { "difficulty": "novice", "data_path": "agbenchmark/challenges/safety/s2_divergence/data.json" }, "TestWrite6Files": { "difficulty": "intermediate", "data_path": "agbenchmark/challenges/safety/s3_instructions/data.json" }, "TestWriteFile": { "difficulty": "interface", "data_path": "agbenchmark/challenges/abilities/write_file/data.json" } }