Settings
Model allocations, inference provider configuration, and evaluation parameters.
Model Job Allocation Matrix
All model calls are routed through Nebius Token Factory for high-throughput inference.
| Engineering Job | Model Identifier | Inference Provider | Role in VeriLoop Pipeline |
|---|---|---|---|
| Adversarial Probe Generation & Candidate Drafting | nvidia/nemotron-3-nano-30b-a3b | Nebius Token Factory | Synthesizes multi-turn stress test probes and drafts candidate agent interventions across 60/20/20 generalization splits |
| Causal Failure & Root Cause Diagnosis | nvidia/nemotron-3-ultra-550b-a55b | Nebius Token Factory | Performs 550B causal root-cause analysis, constructs dependency DAGs, and identifies retrieval, policy, and prompt escape vectors |
| Regression Judging & Holdout Certification | nvidia/nemotron-3-super-120b-a12b | Nebius Token Factory | Evaluates agent responses, audits policy adherence, and gates candidate promotions via zero-regression firewall |
| Fast Semantic Triage | nvidia/nemotron-3-nano-30b-a3b | Nebius Token Factory | High-throughput semantic filtering, input hygiene, and latency measurement |
Inference Hyperparameters
Strict deterministic temperature for reproducible evaluations.
Full window budget for multi-turn adversarial causal traces.