E54: a fault at the boundary passes the status rule almost always; byte-identical weights or an env-side record catches every one

[tactino]

The question

Across machines, E43, E44, E48 and E49 judge a boundary by a status rule: the policy still learns. When something goes wrong at the environment-trainer boundary, does that rule notice? What does?

The answer

Almost never: the status rule misses nearly every fault. Byte-identical weights, or reconciliation with a record kept at the environment, catch every one.

SB3's PPO trained Pendulum and HalfCheetah through the bridge, as in E50. Meanwhile the env client changed one field of what crossed, in one way: 21 faults, covering every field with zero, stale, swap, float16 and clip, plus episode-end and step faults. Two more faults sat in the trainer's log. Doses ran from one value to every value. 590 registered runs, seeds 0-2.

env	dose	status rule	strict (10 seeds)	weights	record	weights or record
Pendulum	one value	0%	0%	91%	28%	100%
	1%	4%	6%	91%	35%	100%
	every value	39%	55%	91%	30%	100%
HalfCheetah	one value	4%	4%	91%	33%	100%
	every value	32%	46%	91%	33%	100%

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

• Weights caught all 504 client-fault runs at every dose, and no control or log fault. The 91% in the table is client and log faults together.
• Record caught what changes the books: rewards beyond its tolerance, episode ends, steps, the log. It did not catch faults in observations or actions.
• Between them, the two checks caught all 546 silent runs.
• Localisation: the first difference from the fault-free run named the step, env and field in all 351 traced runs.
• Loud faults: the two faults the bridge rejects raised at once.

One changed value moved the final return by about as much as a change of seed. That is why a curve cannot tell a fault from a seed.

The per-run logs (100 MB) stay on guangzhao. results/runs.jsonl and runs.csv carry every run's result and every check's verdict.