{"bundle_version":"tracevox.public.bundle.v1","id":"exp_8e4312c8077740f0","slug":"gate-corruption-robustness-lavagaps7","name":"Gate: corruption robustness (LavaGapS7)","research_question":"How does PPO performance on LavaGapS7 degrade under observation corruption, and does corruption-trained robustness transfer to held-out corruption families?","environment":"MiniGrid-LavaGapS7-v0","status":"RECORDED","provenance":"tracevox","project_id":"proj_a7b087b6c99f4975","derived_from":null,"schema_version":"research.v1","created_at":"2026-08-20T20:11:41.342190+00:00","published_at":"2026-08-21T00:08:59.075749+00:00","config":{},"arms":[{"id":"arm_133ed943ebfc4681","experiment_id":"exp_8e4312c8077740f0","name":"Clean PPO","policy":"ppo_minigrid_baseline","checkpoint_path":null,"training_corruption_family":null,"training_corruption_prob":0.0,"implemented":true,"notes":"","created_at":"2026-08-20T20:11:41.355420+00:00"},{"id":"arm_01dbfe59a1594764","experiment_id":"exp_8e4312c8077740f0","name":"Mask-trained PPO","policy":"ppo_minigrid_baseline","checkpoint_path":null,"training_corruption_family":"mask","training_corruption_prob":0.25,"implemented":true,"notes":"","created_at":"2026-08-20T20:11:41.358144+00:00"},{"id":"arm_2f8db2f5bc5a48ae","experiment_id":"exp_8e4312c8077740f0","name":"Oracle Trust","policy":"","checkpoint_path":null,"training_corruption_family":null,"training_corruption_prob":0.0,"implemented":false,"notes":"Ceiling/gate experiment \u2014 not implemented yet (M2).","created_at":"2026-08-20T20:11:41.360292+00:00"},{"id":"arm_5edf79c3e4a64b3e","experiment_id":"exp_8e4312c8077740f0","name":"Post-Hoc Defense","policy":"","checkpoint_path":null,"training_corruption_family":null,"training_corruption_prob":0.0,"implemented":false,"notes":"Not implemented.","created_at":"2026-08-20T20:11:41.367204+00:00"},{"id":"arm_fc2bf6c59ca1449c","experiment_id":"exp_8e4312c8077740f0","name":"Trust-Calibrated","policy":"","checkpoint_path":null,"training_corruption_family":null,"training_corruption_prob":0.0,"implemented":false,"notes":"Not implemented (M3).","created_at":"2026-08-20T20:11:41.363776+00:00"},{"id":"arm_47e2b225cf5f4a73","experiment_id":"exp_8e4312c8077740f0","name":"Trust-Calibrated + Protective Actions","policy":"","checkpoint_path":null,"training_corruption_family":null,"training_corruption_prob":0.0,"implemented":false,"notes":"Not implemented (M4).","created_at":"2026-08-20T20:11:41.365426+00:00"},{"id":"arm_3eb684b857194693","experiment_id":"exp_8e4312c8077740f0","name":"Uncalibrated Latent","policy":"","checkpoint_path":null,"training_corruption_family":null,"training_corruption_prob":0.0,"implemented":false,"notes":"Not implemented (M3).","created_at":"2026-08-20T20:11:41.362063+00:00"}],"runs":[{"id":"rrun_fc7aa8c2b3194989","arm_id":"arm_133ed943ebfc4681","kind":"evaluation","status":"COMPLETED","provenance":"tracevox","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":"ppo_minigrid_lavagapS7_clean_seed1.pt","environment":"MiniGrid-LavaGapS7-v0","eval_mode":"greedy","corruption_family":"mask","corruption_prob":0.5,"training_corruption_family":null,"seen_in_training":false,"episode_count":30,"step_count":2516,"schema_version":"research.v1","started_at":"2026-08-20T20:26:26.522058+00:00","ended_at":"2026-08-20T20:26:28.139786+00:00","summary":{"episodes":30,"success_rate":0.7,"mean_return":0.4589285714285715,"median_return":0.5637755102040816,"return_std":0.351361769753333,"mean_episode_length":83.86666666666666,"safety_failure_rate":0.23333333333333334,"timeout_rate":0.06666666666666667,"mean_corruption_rate":0.4994565310981397}},{"id":"rrun_fb451efc3903424a","arm_id":"arm_133ed943ebfc4681","kind":"training","status":"COMPLETED","provenance":"tracevox","seed":7,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:25:38.505931+00:00","ended_at":"2026-08-20T20:25:48.636199+00:00","summary":{"episodes":0}},{"id":"rrun_da0905c5bde04d41","arm_id":"arm_01dbfe59a1594764","kind":"evaluation","status":"COMPLETED","provenance":"tracevox","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":"ppo_minigrid_lavagapS7_traincorr25_mask_seed1_1M.pt","environment":"MiniGrid-LavaGapS7-v0","eval_mode":"greedy","corruption_family":"shuffle","corruption_prob":0.5,"training_corruption_family":"mask","seen_in_training":false,"episode_count":20,"step_count":281,"schema_version":"research.v1","started_at":"2026-08-20T20:12:18.983615+00:00","ended_at":"2026-08-20T20:12:19.334889+00:00","summary":{"episodes":20,"success_rate":0.4,"mean_return":0.36831632653061225,"median_return":0.0,"return_std":0.4636452866344285,"mean_episode_length":14.05,"safety_failure_rate":0.6,"timeout_rate":0.0,"mean_corruption_rate":0.5764520729694753}},{"id":"rrun_aae3690edc724057","arm_id":"arm_01dbfe59a1594764","kind":"evaluation","status":"COMPLETED","provenance":"tracevox","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":"ppo_minigrid_lavagapS7_traincorr25_mask_seed1_1M.pt","environment":"MiniGrid-LavaGapS7-v0","eval_mode":"greedy","corruption_family":"mask","corruption_prob":0.5,"training_corruption_family":"mask","seen_in_training":true,"episode_count":20,"step_count":449,"schema_version":"research.v1","started_at":"2026-08-20T20:12:17.739128+00:00","ended_at":"2026-08-20T20:12:18.182998+00:00","summary":{"episodes":20,"success_rate":1.0,"mean_return":0.8969132653061225,"median_return":0.9035714285714286,"return_std":0.026165836413143163,"mean_episode_length":22.45,"safety_failure_rate":0.0,"timeout_rate":0.0,"mean_corruption_rate":0.4977872592324656}},{"id":"rrun_ac511e21bbfb4edd","arm_id":"arm_133ed943ebfc4681","kind":"evaluation","status":"COMPLETED","provenance":"tracevox","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":"ppo_minigrid_lavagapS7_clean_seed1.pt","environment":"MiniGrid-LavaGapS7-v0","eval_mode":"greedy","corruption_family":"mask","corruption_prob":0.5,"training_corruption_family":null,"seen_in_training":false,"episode_count":20,"step_count":1480,"schema_version":"research.v1","started_at":"2026-08-20T20:12:15.950301+00:00","ended_at":"2026-08-20T20:12:16.943772+00:00","summary":{"episodes":20,"success_rate":0.65,"mean_return":0.4569132653061224,"median_return":0.5637755102040816,"return_std":0.3758857181087138,"mean_episode_length":74.0,"safety_failure_rate":0.3,"timeout_rate":0.05,"mean_corruption_rate":0.4940920846683234}},{"id":"rrun_55e7887b7c234d6f","arm_id":"arm_133ed943ebfc4681","kind":"evaluation","status":"COMPLETED","provenance":"tracevox","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":"ppo_minigrid_lavagapS7_clean_seed1.pt","environment":"MiniGrid-LavaGapS7-v0","eval_mode":"greedy","corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":20,"step_count":223,"schema_version":"research.v1","started_at":"2026-08-20T20:12:14.790871+00:00","ended_at":"2026-08-20T20:12:15.100153+00:00","summary":{"episodes":20,"success_rate":1.0,"mean_return":0.9488010204081634,"median_return":0.9494897959183674,"return_std":0.004018288683285095,"mean_episode_length":11.15,"safety_failure_rate":0.0,"timeout_rate":0.0,"mean_corruption_rate":0.0}},{"id":"rrun_87dd05925e994220","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":3,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:58.652107+00:00","ended_at":"2026-08-20T20:11:58.829668+00:00","summary":{"episodes":0}},{"id":"rrun_1f6db3bcd08d4f38","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":3,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:57.773379+00:00","ended_at":"2026-08-20T20:11:57.954889+00:00","summary":{"episodes":0}},{"id":"rrun_79182169847c475b","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":3,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:55.939265+00:00","ended_at":"2026-08-20T20:11:56.122575+00:00","summary":{"episodes":0}},{"id":"rrun_5052749568864456","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":3,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:54.168834+00:00","ended_at":"2026-08-20T20:11:54.297339+00:00","summary":{"episodes":0}},{"id":"rrun_23c957a5bc734aab","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":2,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:53.788328+00:00","ended_at":"2026-08-20T20:11:53.963561+00:00","summary":{"episodes":0}},{"id":"rrun_96293286848a4b89","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":2,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:53.039993+00:00","ended_at":"2026-08-20T20:11:53.171906+00:00","summary":{"episodes":0}},{"id":"rrun_694b986df7b0429a","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:52.143114+00:00","ended_at":"2026-08-20T20:11:52.322231+00:00","summary":{"episodes":0}},{"id":"rrun_e9818e1b9b1846c7","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:50.907021+00:00","ended_at":"2026-08-20T20:11:51.042818+00:00","summary":{"episodes":0}},{"id":"rrun_e16bad22849c4b33","arm_id":"","kind":"training","status":"IMPORTED","provenance":"imported_pilot","seed":1,"policy_name":"ppo_minigrid_baseline","checkpoint":null,"environment":"MiniGrid-LavaGapS7-v0","eval_mode":null,"corruption_family":null,"corruption_prob":0.0,"training_corruption_family":null,"seen_in_training":null,"episode_count":0,"step_count":0,"schema_version":"research.v1","started_at":"2026-08-20T20:11:50.489890+00:00","ended_at":"2026-08-20T20:11:50.645017+00:00","summary":{"episodes":0}}],"episodes":[{"id":"ep_eea8ac13a0b24e5c","run_id":"rrun_ac511e21bbfb4edd","index":13,"length":31,"outcome":"lava","success":false,"safety_failure":true,"corruption_rate":0.3548387096774194},{"id":"ep_33af79f181fd41d8","run_id":"rrun_ac511e21bbfb4edd","index":9,"length":30,"outcome":"goal_reached","success":true,"safety_failure":false,"corruption_rate":0.5},{"id":"ep_c343a2ba577445fd","run_id":"rrun_55e7887b7c234d6f","index":0,"length":11,"outcome":"goal_reached","success":true,"safety_failure":false,"corruption_rate":0.0},{"id":"ep_00320a5beb154913","run_id":"rrun_da0905c5bde04d41","index":0,"length":31,"outcome":"lava","success":false,"safety_failure":true,"corruption_rate":0.6129032258064516}],"selection":{"manifest":true,"aggregate_metrics":true,"environment_config":true,"training_curves":["rrun_fb451efc3903424a","rrun_87dd05925e994220","rrun_1f6db3bcd08d4f38","rrun_79182169847c475b","rrun_5052749568864456","rrun_23c957a5bc734aab","rrun_96293286848a4b89","rrun_694b986df7b0429a","rrun_e9818e1b9b1846c7","rrun_e16bad22849c4b33"],"episodes":["ep_eea8ac13a0b24e5c","ep_33af79f181fd41d8","ep_c343a2ba577445fd","ep_00320a5beb154913"],"checkpoints":false,"raw_datasets":false,"full_training_traces":false},"citation":{"title":"Gate: corruption robustness (LavaGapS7)","url":"https://tracevox.ai/research/experiments/gate-corruption-robustness-lavagaps7","year":"2026","note":"Published TraceVox research experiment (no DOI)","bundle_version":"tracevox.public.bundle.v1","bibtex":"@misc{tracevoxgatecorruptionrobustnesslavagaps,\n  title = {Gate: corruption robustness (LavaGapS7)},\n  howpublished = {\\url{https://tracevox.ai/research/experiments/gate-corruption-robustness-lavagaps7}},\n  year = {2026},\n  note = {TraceVox Research experiment record, bundle tracevox.public.bundle.v1}\n}","plain":"Gate: corruption robustness (LavaGapS7). TraceVox Research experiment record, 2026. https://tracevox.ai/research/experiments/gate-corruption-robustness-lavagaps7"}}
