The environment has no explicit message channel between agents; the only cross-agent information is the press agent's view of what the sorting stage is currently producing. Communication loss cuts that link (randomized onset between steps 40-60), in one of two modes: stale (the press agent keeps seeing the last value from before the link dropped) or blackout (the value is zeroed out). The press agent's own local sensing and the sort agent's observation are untouched, so the fault only affects the press channel.
Experimental trajectory graphs
Stale Mode - Permanent Duration
Stale Mode - Pressing Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (communication loss only ever affects the press channel - see injection definition above), seed 0 of the 8-seed sweep.
Stale Mode - Transient Duration (20 steps, onset between steps 40-60)
Stale Mode - Pressing Agent Target (Transient): evaluation trajectory with no recovery mechanism active (communication loss only ever affects the press channel - see injection definition above), seed 0 of the 8-seed sweep.
Blackout Mode - Permanent Duration
Blackout Mode - Pressing Agent Target (Permanent): evaluation trajectory with no recovery mechanism active (communication loss only ever affects the press channel - see injection definition above), seed 0 of the 8-seed sweep.
Blackout Mode - Pressing Agent Target (Transient): evaluation trajectory with no recovery mechanism active (communication loss only ever affects the press channel - see injection definition above), seed 0 of the 8-seed sweep.
Duration modes
Permanent: the fault persists until the end of the episode.
