The experiment evaluates LLM-agent replanning under a permanent byzantine / adversarial agents injection. Fault onset is randomized between environment steps 40 and 60.
Expected evaluation record
Metric
Recorded value
Degradation-area
58.51 ± 8.82 (corrected; mean ± 95% CI, n=72)
Safety-violation excess
290.2 ± 50.8 (weighted, control-subtracted; mean ± 95% CI, n=72)
Fraction of time healthy
0.326 ± 0.044 (mean ± 95% CI, n=72)
Relapse count
0.24 ± 0.12 (mean ± 95% CI, n=72)
False-recovery rate
0.013 ± 0.007 (mean ± 95% CI, n=72)
Intervention recall
0.643 ± 0.061 (mean ± 95% CI, n=72)
Values are means with 95% confidence intervals across every target × duration-fixed combination shown on this page (see graphs below), 8 seeds each. See the Results page for the cross-mechanism comparison.
Experimental trajectory graphs
One representative seed (seed 0 of 8) 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)
LLM-agent replanning - Sorting Agent Target (fixed malicious mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Pressing Agent Target (fixed malicious mode)
LLM-agent replanning - Pressing Agent Target (fixed malicious mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Both Agents Target (fixed malicious mode)
LLM-agent replanning - Both Agents Target (fixed malicious mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Sorting Agent Target (random malicious mode)
LLM-agent replanning - Sorting Agent Target (random malicious mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Pressing Agent Target (random malicious mode)
LLM-agent replanning - Pressing Agent Target (random malicious mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Both Agents Target (random malicious mode)
LLM-agent replanning - Both Agents Target (random malicious mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Sorting Agent Target (worst action mode)
LLM-agent replanning - Sorting Agent Target (worst action mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Pressing Agent Target (worst action mode)
LLM-agent replanning - Pressing Agent Target (worst action mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
Both Agents Target (worst action mode)
LLM-agent replanning - Both Agents Target (worst action mode) - evaluation trajectory under permanent byzantine / adversarial agents (fault onset randomized between steps 40-60), seed 0 of the 8-seed sweep.
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.
