fix: png rename + remove placeholder + fix blog link + add baseline comparison

README.md

@@ -21,7 +21,7 @@ pinned: false
 | Resource | URL |
 |---|---|
 | 🤗 HuggingFace Space (live env) | https://huggingface.co/spaces/CystronCode/cannon-and-wall |
-| 📝 Mini blog post | https://huggingface.co/CystronCode/cannon-and-wall-grpo/blob/main/README.md |
+| 📝 Writeup / Model Card | https://huggingface.co/CystronCode/cannon-and-wall-grpo |
 | 📓 Training notebook (Colab) | https://colab.research.google.com/drive/1uTSt6DahNVXoAZ0hlpyzrXrn-F6ehc9z?usp=sharing|
 | Reward curves | https://wandb.ai/saiamogh7-r-v-c-e/cannon-and-wall/runs/p3uh8q2b |
 
@@ -180,13 +180,11 @@ Escalation triggers when either agent achieves consistently low reward for 3 epi
 
 ## 📈 Results
 
-*(Filled in after training run — April 25-26, 2026)*
-
 ### Reward curves
 
 ![Reward curve](assets/reward_curve.png)
 
-*50 episodes — Cannon (red) averages 1.3 reward, Wall (blue) stable at 0.714. Adversarial self-play loop confirmed working.*
+*Episode 0: Cannon reward ~0 (no valid exploit pattern). Episode 50: Cannon averages 1.3 — consistently identifying and bypassing SQLi vulnerabilities. Wall holds stable at 0.714 (100% patch validity rate). Adversarial self-play loop confirmed working.*
 
 ### Results after 50 episodes
 
@@ -197,9 +195,14 @@ Escalation triggers when either agent achieves consistently low reward for 3 epi
 | Cannon bypass success rate | ~60% |
 | Wall patch validity rate | 100% |
 | Episodes run | 50 |
+| Cannon reward at episode 0 | ~0.0 (untrained baseline) |
+| Cannon reward at episode 50 | ~1.3 (after self-play) |
 | Base model | Qwen/Qwen2.5-3B-Instruct |
 | Environment | Live on HuggingFace Spaces |
 
+> Note: v1.0 uses the same Flask app across all curriculum stages.
+> Stage 2 and 3 variant files are planned for v1.1.
+
 ## ▶️ Running Locally
 
 ```bash