| # Architecture |
|
|
| POLYGUARD-RL uses an OpenEnv-first monorepo architecture with six layers: |
|
|
| 1. Data ingestion and retrieval index. |
| 2. Predictive safety, graph, tabular risk, and dosing models. |
| 3. Multi-agent orchestration graph. |
| 4. Hierarchical RL training stack. |
| 5. Safety governance and anti-cheat controls. |
| 6. FastAPI, OpenEnv, and React deployment surfaces. |
|
|
| ## Data Flow |
|
|
| ```text |
| raw/local knowledge -> processed tables -> scenarios -> SFT/GRPO corpora |
| | |
| v |
| PolyGuardEnv reset/step/state |
| | |
| v |
| agent stack -> verifier reward -> training/evaluation reports |
| | |
| v |
| docs/results + README + HF Space |
| ``` |
|
|
| ## Runtime Boundaries |
|
|
| - Environment code owns state transition, legality, rewards, anti-cheat, and traces. |
| - Agent code owns candidate interpretation, routing, planning, critique, and explanation. |
| - Training code owns SFT, GRPO, reward logging, adapters, and registry metadata. |
| - Evaluation code owns baselines, perturbations, reports, and plots. |
| - Deployment code owns OpenEnv validation and HF Space push. |
|
|
