E57: on a clean environment, faults the curve checks miss leave policies as good as fault-free ones; every wrecked policy was already caught

[tactino]

Stacked on #123 (E56): it reads E56's runs.csv.

The question

Reviewers of the workshop draft asked whether the boundary faults matter at all. E54–E56 judged each run by what the trainer saw. This experiment judges the run's outcome instead: the policy it produced, evaluated on a clean environment. Is a policy trained through a fault worse, in particular one whose fault both curve checks miss?

The answer

No, in these tasks. Each of the 834 final policies from E54 and E56 played 50 deterministic episodes in a fresh gymnasium env.

• Which runs were harmed. "Harmed" means more than 3 s.d. below the fault-free band. With every value faulted, 24, 28 and 19 runs were harmed on Pendulum, HalfCheetah and Hopper. The curve checks had already caught every one of them.
• Runs both curve checks missed. On HalfCheetah and Hopper, none was harmed, at any dose. On Pendulum, 5 of 255 scored below all 13 fault-free runs. Thirteen fault-free runs are too few to tell rare harm from a rare bad seed.
• Exploratory. The runs silent on the curve sit 0.03–0.65 band s.d. above the band. None is significantly below it.
• P3. Pendulum's reward:zero logged 0, the best possible return, while its policies scored -1196 to -1639.

Every check and prediction holds (results/verdicts.txt). There is one amendment. summarise.py could not parse the blank cells of the loud faults, which the protocol excludes. It was fixed before any output was seen, and no count changed.

What this means for the paper: a silent fault is silent on a clean evaluation too. What it changes is the run itself (its weights and its logs), not the policy's quality. Only a check of the boundary itself, the step ledger or the episode record, tells such a run apart.