Link EpiWM blog and correct score data reference

README.md

@@ -10,9 +10,11 @@ tags:
 
 # EpiWM: LeWorldModel trained with epiplexity instead of SIGReg
 
+[Blog post](https://the-puzzler.github.io/blog/epiwm/) · [Code and results](https://github.com/the-puzzler/epijepa/tree/main/worldmodel)
+
 EpiWM: these are the best checkpoints from replacing [LeWorldModel](https://github.com/lucas-maes/le-wm)'s SIGReg
 anti-collapse term with the epiplexity score from EpiJEPA. The code, the full analysis and our replication of LeWM's
-SIGReg recipe are in [the-puzzler/epijepa](https://github.com/the-puzzler/epijepa) (`worldmodel/`).
+SIGReg recipe are in [the-puzzler/epijepa](https://github.com/the-puzzler/epijepa/tree/main/worldmodel).
 
 There is one folder per environment, laid out like LeWM's own releases:
 
@@ -24,8 +26,8 @@ There is one folder per environment, laid out like LeWM's own releases:
 | `reacher/` | Reacher (DMC) | 0.3 | 200k | 1 | 76 | 72.0 | 76.6 | 52 / 62.0 / 60.8 |
 
 These are planning success rates (%) with LeWM's CEM planner and evaluation (`eval.py`). The 3-seed means of the
-same recipe are TwoRoom 99.9, Push-T 88.5, Cube 71.5 and Reacher 73.5 on n=500. LeWM's SIGReg recipe retrained by us at the same budget (2 seeds) gives 87.9, 88.6, 65.5 and 62.2. The released LeWM column is our measurement with LeWM's released evaluation; on Reacher the release uses random-policy data, not the paper's SAC data. Each folder holds the seed with the
-best n=500 score; every seed's number is in the repo's `worldmodel/analysis/scores/all_scores.csv`.
+same recipe are TwoRoom 99.9, Push-T 88.5, Cube 71.5 and Reacher 73.5 on n=500. LeWM's SIGReg recipe retrained by us at the same budget (2 seeds) gives 87.9, 88.6, 65.5 and 62.2. The released LeWM column is our measurement with LeWM's released evaluation; Reacher evaluations use random-policy data, while the paper describes SAC-collected data. Each folder holds the seed with the
+best n=500 score; every seed's number is in [analysis/scores/all_scores.csv](https://huggingface.co/basilboy/epiwm/blob/main/analysis/scores/all_scores.csv).
 
 ## Model