Context-Fragmented Violations arise when multi-agent AI actions are locally safe yet globally non-compliant due to siloed contexts; Distributed Sentinel with STT Protocol detects them at F1=0.95 on PhantomEcosystem benchmark, outperforming prompt-based and rule-based baselines.
comp data
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.MA 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Beyond Single-Agent Alignment: Preventing Context-Fragmented Violations in Multi-Agent Systems
Context-Fragmented Violations arise when multi-agent AI actions are locally safe yet globally non-compliant due to siloed contexts; Distributed Sentinel with STT Protocol detects them at F1=0.95 on PhantomEcosystem benchmark, outperforming prompt-based and rule-based baselines.