---
license: mit
tags: [icml2026, reproduction, safe-rl, trackio]
---

# CSPO reproduction artifacts

Artifacts from an independent reproduction of **CSPO: Constraint-Sensitive Policy Optimization
for Safe Reinforcement Learning** (ICML 2026 spotlight, OpenReview `3ySR3TCMRP`).

* Paper code: https://github.com/serval-uni-lu/CSPO (commit `962e696eb1f07b47ac0094bf1a2aa1299f97cc9b`)
* Full logbook: https://huggingface.co/spaces/nmaher/repro-cspo-constraint-sensitive-policy-optimization-for-safe-reinforcement-learning

## Contents

| Path | What it is |
|---|---|
| `analysis/` | All recomputed metric tables: recovery metrics (TTS/RP/VF), Claim 5 sensitivity sweep, Claim 1 derivation + live-probe audits, Claim 2 KKT equivalence + rate audits, Claim 3 environment/baseline coverage |
| `retraining_pointgoal/<ALGO>/seed-<N>/progress.csv` | Per-epoch OmniSafe training logs from our own CSPO vs PPO-Lag retraining on `SafetyPointGoal1-v0` (paper hyperparameters, 10M-step budget, 3 seeds, CPU-only) |

All analysis scripts live in the logbook Space's Workspace tab.