Corpora.ai today announced RACE, a modular multi‑agent framework that blends formal ontology grounding, dynamic reputation, and adversarial training to keep autonomous systems coordinated even when a subset of agents are compromised. RACE’s layered architecture delivers provable Byzantine‑resilient convergence, real‑time trust‑aware sensor fusion, and randomized smoothing for large‑language‑model agents—all while maintaining sub‑linear overhead for thousands of nodes.
At its core, RACE integrates Dynamic Role‑Based Adversarial Training (DRAT), which continuously exposes agents to an evolutionary attacker generator and enforces role specialization (Orchestrator, Executor, Critic, etc.) to prevent hallucination propagation. This adaptive loop hardens policies against unseen coordination attacks, a critical advantage over static signature‑based defenses.
The engine’s trust layer fuses Hybrid Reputation Aggregation (HRA) and Trust‑Aware Sensor Fusion with Dynamic Field‑of‑View (TASF‑DFOV). HRA combines geometric anomaly detection with momentum‑based reputation scores, achieving 98.66 % accuracy in federated retraining scenarios versus 84.77 % for anomaly‑only methods. TASF‑DFOV, grounded in a hidden‑Markov‑model, detects >95 % of spoofing, jamming, and replay attacks while keeping localization error below 0.8 m, even when multiple sensors are compromised.
Randomized Smoothing for LLM‑Based MAS (RS‑LLM‑MAS) provides a certified radius that bounds the influence of malicious hallucinations in language‑driven coordination. Integrated with MPAC’s multi‑principal governance, RS‑LLM‑MAS ensures that no single principal can dictate the joint policy, preserving robustness in distributed decision‑making.
Looking ahead, Corpora.ai will deploy RACE in a series of field trials: a UAV swarm for search‑and‑rescue, a cyber‑physical sensor mesh for critical infrastructure, and a decentralized finance protocol for secure transaction orchestration. The modular design also allows rapid adaptation to new threat models, positioning RACE as the foundation for the next generation of trustworthy autonomous systems.
Key Facts
- RACE guarantees convergence under bounded Byzantine attacks, a property unattainable with conventional consensus protocols.
- Hybrid Reputation Aggregation achieves 98.66 % accuracy in federated learning, outperforming anomaly‑only or reputation‑only baselines.
- Trust‑Aware Sensor Fusion detects >95 % of spoofing/jamming attacks while keeping localization error under 0.8 m.
About Corpora.ai: Corpora.ai is a frontier deep‑tech venture dedicated to building resilient, interpretable AI systems that can operate safely in hostile, dynamic environments. By combining formal methods, advanced machine learning, and robust governance, Corpora.ai empowers organizations to deploy autonomous agents at scale with full auditability and provable safety.