RL Triage Agent · Diagnosis, Retraining & Policy Checks

RL Triage Agent · Diagnosis, Retraining & Policy Checks An architecture diagram generated by Archify. NAT Triage Agent · NeMo Agent Toolkit · OpenShell Sandbox · Config / Reward Edits Blocked NAT Triage Agent NeMo Agent Toolkit Preregister Experiment · write_preregistration · OpenShell Sandbox · Config / Reward Edits Blocked Preregister Experiment write_preregistration Host Submission · Create pending request · Architecture component · submit CLI / POST /eval Host Submission Create pending request submit CLI / POST /eval Human Approval · bridge approve CLI · Architecture component · Approval blocked in sandbox Human Approval bridge approve CLI Approval blocked in sandbox Isaac Lab Retraining · Revert 1 suspect variable · Architecture component Isaac Lab Retraining Revert 1 suspect variable Telemetry Analysis · Baseline / History / Python · OpenShell Sandbox · Config / Reward Edits Blocked Telemetry Analysis Baseline / History / Python Nemotron 3 Super · External NVIDIA API · Architecture component · integrate.api.nvidia.com Nemotron 3 Super External NVIDIA API integrate.api.nvidia.com Diagnosis Verification · Recovered / Not recovered · Architecture component Diagnosis Verification Recovered / Not recovered Static Policy Proposal · policies/proposals/*.yaml · Offline Policy Checks · evals/prove_policies.py · Merged with the base policy Static Policy Proposal policies/proposals/*.yaml Merged with the base policy openshell-prover · Compare to site_boundary · Offline Policy Checks · evals/prove_policies.py openshell-prover Compare to site_boundary Human Policy Review · human_review · Offline Policy Checks · evals/prove_policies.py Human Policy Review human_review Automatic Rejection · auto_reject · Offline Policy Checks · evals/prove_policies.py Automatic Rejection auto_reject Register Read file Pending Approved Verify retraining Call tools POST /v1/chat/completions Check policy within_boundary Exceeds / Unprovable OpenShell Sandbox · Config / Reward Edits Blocked Offline Policy Checks · evals/prove_policies.py