Project: Trust-Calibrated RL · Environment: MiniGrid-LavaGapS7-v0 · Schema: research.v1 · Published: 2026-08-21
How does PPO performance on LavaGapS7 degrade under observation corruption, and does corruption-trained robustness transfer to held-out corruption families?
| Arm | Corruption | Attack exposure | Success | Mean return | Safety failures | Episodes | Run |
|---|---|---|---|---|---|---|---|
| Clean PPO | mask @ p=0.5 | HELD OUT | 70.0% | 0.459 | 23.3% | 30 | seed 1 · greedy · rrun_fc7aa8c2b3194989 |
| Mask-trained PPO | shuffle @ p=0.5 | HELD OUT | 40.0% | 0.368 | 60.0% | 20 | seed 1 · greedy · rrun_da0905c5bde04d41 |
| Mask-trained PPO | mask @ p=0.5 | SEEN IN TRAINING | 100.0% | 0.897 | 0.0% | 20 | seed 1 · greedy · rrun_aae3690edc724057 |
| Clean PPO | mask @ p=0.5 | HELD OUT | 65.0% | 0.457 | 30.0% | 20 | seed 1 · greedy · rrun_ac511e21bbfb4edd |
| Clean PPO | none | — | 100.0% | 0.949 | 0.0% | 20 | seed 1 · greedy · rrun_55e7887b7c234d6f |
Training and evaluation metrics are recorded separately; a stochastic training rollout is never presented as a controlled evaluation. Trust calibration and protective actions (ACT/VERIFY/DEFER/ABSTAIN) are research milestones — not implemented, and no values for them are shown anywhere.
Every run records seeds, package versions, checkpoint SHA-256 hashes, and upstream code provenance. Manifest: manifest.json · Full record: experiment.json · Citation: citation.json
Gate: corruption robustness (LavaGapS7). TraceVox Research experiment record, 2026. https://tracevox.ai/research/experiments/gate-corruption-robustness-lavagaps7
@misc{tracevoxgatecorruptionrobustnesslavagaps,
title = {Gate: corruption robustness (LavaGapS7)},
howpublished = {\url{https://tracevox.ai/research/experiments/gate-corruption-robustness-lavagaps7}},
year = {2026},
note = {TraceVox Research experiment record, bundle tracevox.public.bundle.v1}
}
TraceVox Research home · Public Research Library · Documentation · llms.txt