E56: on Hopper, the step ledger catches and locates every fault in what crosses; curve checks miss all 69 at one value

[tactino]

The question

E54 and E55 injected faults on Pendulum and HalfCheetah, which never end an episode early. Hopper ends an episode when it falls, and it is harder to learn. On Hopper, do E54's and E55's findings hold? And what do faults in terminations do?

The answer

They hold. The faults were E54's 21 client faults plus two in terminations (terminated:drop, terminated:as-trunc) and the two log faults, at doses one, 0.01 and 1.0. There were 244 runs of 1,024,000 steps.

Client-fault runs with a changed value, 69 at each dose:

dose	status rule	strict	byte identity	record	step ledger
one value	0	0	57	19	69
0.01	3	3	69	21	69
1.0	33	34	69	18	69

Every check and prediction holds (results/verdicts.txt).

• One changed value moved the final return by a median of 197, against a fault-free seed s.d. of 259. Neither curve check caught any of the 69 runs.
• At 1.0, 11 faults that change every value still passed both curve checks at all three seeds.
• The ledger caught and located all 207 client-fault runs. It found no difference in the 27 control and log-fault runs. The strict check flagged none of the 10 fault-free seeds.
• New on Hopper: a fault can cross without changing the weights. At one value, each final-obs fault changed the final observation of an episode that ended in a fall. SB3 bootstraps from it only at a time limit, so the weights stayed none's in 12 of 12 runs. Only the ledger caught them. At 0.01 and 1.0 the changed values land at time limits too, and byte identity caught 24 of 24.
• terminated:as-trunc changes nothing in the record's books, so the record missed it at every dose. With terminated:drop at 1.0, the trainer's log held no episode, while the environment's record held 7,178.

verdicts.py was not in the protocol commit. It was last changed after the registered runs started and before the first one ended. FINDINGS.md gives the times, and the file is committed unchanged.

The per-run directories stay on guangzhao. Their traces and ledger arrays (150 MB a run) were deleted once judged, as run.sh says.