The experiment evaluates Fault-tolerant-trained MARL under a transient byzantine / adversarial agents injection. Fault onset is randomized between environment steps 40 and 60.
Expected evaluation record
Metric
Recorded value
Degradation-area
28.37 ± 2.90 (corrected; mean ± 95% CI, n=216)
Safety-violation excess
168.4 ± 19.2 (weighted, control-subtracted; mean ± 95% CI, n=216)
Fraction of time healthy
0.533 ± 0.017 (mean ± 95% CI, n=216)
Relapse count
0.70 ± 0.10 (mean ± 95% CI, n=216)
False-recovery rate
n/a - no fault detector (trained policy only)
Intervention recall
n/a - no fault detector (trained policy only)
Values are means with 95% confidence intervals across every target × duration-fixed combination shown on this page (see graphs below), 8 evaluation seeds each, for each of 3 training seeds. See the Results page for the cross-mechanism comparison.
Experimental trajectory graphs
One representative run (evaluation seed 0 of 8, training seed 1 of 3) per target × mode combination from the full sweep. See the Results page for the multi-seed aggregate comparison across mechanisms. The graphs for every seed are archived at doi:10.5281/zenodo.23304785 (files available from 10 November 2026).
Sorting Agent Target (fixed malicious mode)
Fault-tolerant-trained MARL - Sorting Agent Target (fixed malicious mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Pressing Agent Target (fixed malicious mode)
Fault-tolerant-trained MARL - Pressing Agent Target (fixed malicious mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Both Agents Target (fixed malicious mode)
Fault-tolerant-trained MARL - Both Agents Target (fixed malicious mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Sorting Agent Target (random malicious mode)
Fault-tolerant-trained MARL - Sorting Agent Target (random malicious mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Pressing Agent Target (random malicious mode)
Fault-tolerant-trained MARL - Pressing Agent Target (random malicious mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Both Agents Target (random malicious mode)
Fault-tolerant-trained MARL - Both Agents Target (random malicious mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Sorting Agent Target (worst action mode)
Fault-tolerant-trained MARL - Sorting Agent Target (worst action mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Pressing Agent Target (worst action mode)
Fault-tolerant-trained MARL - Pressing Agent Target (worst action mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Both Agents Target (worst action mode)
Fault-tolerant-trained MARL - Both Agents Target (worst action mode) - evaluation trajectory under transient byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep, training seed 1 of 3.
Notes
This page is intentionally separated from the paper-facing results so that full-resolution plots, additional seeds, and diagnostic figures can be documented without overloading the publication.
