veronica-core includes reproducible evaluation of runtime containment across four canonical runaway failure modes (retry amplification, recursive tools, multi-agent loops, WebSocket runaway):
- Technical paper -- system design, threat model, formal safety guarantees (G1-G6), evaluation
- Baseline comparison -- no containment vs veronica across four scenarios (avg 78.8% call reduction)
- Ablation study -- incremental component contribution (BudgetEnforcer, AgentStepGuard, CircuitBreaker, RetryContainer)
- Real incident reproduction -- five real-world failure scenarios with before/after comparison
- Scale simulation -- 1 to 1000 concurrent agent chains (~83.1% reduction, ~12.63 us/chain overhead)
- Reproducibility guide -- environment, commands, expected output, verification against paper claims
Supporting theory:
- Amplification model -- formal model of retry and agent amplification with worked examples
- Safety guarantees -- cost bound, termination, retry budget, failure isolation proofs