new model repo name + full MATH-500 numbers in header

app.py

@@ -9,7 +9,7 @@ import json
 import gradio as gr
 
 BASE = "Qwen/Qwen3.5-0.8B"
-EDITED = "akhaliq/Qwen3.5-0.8B-wsp-math-edit"
+EDITED = "akhaliq/Qwen3.5-0.8B-weight-space-programming"
 INSTRUCTION = "Please reason step by step, and put your final answer within \\boxed{}."
 
 try:
@@ -132,11 +132,12 @@ From **["Sparks of Weight Space Programming"](https://rikiyat.github.io/blog/wei
 a frontier LLM read Qwen3.5-0.8B's weights, found a direction that makes it loop until the token
 limit, and **rewrote 131,072 weights to remove it — no training, no gradients**. The blog post
 never released the model; this is a bit-exact, hash-verified rebuild
-([akhaliq/Qwen3.5-0.8B-wsp-math-edit](https://huggingface.co/akhaliq/Qwen3.5-0.8B-wsp-math-edit)).
+([akhaliq/Qwen3.5-0.8B-weight-space-programming](https://huggingface.co/akhaliq/Qwen3.5-0.8B-weight-space-programming)).
 
 **Left: the original model** — watch it circle and repeat. **Right: the edited model** — same
-weights everywhere else, same prompt, same sampling. Verified on MATH-500: 5% → 47% accuracy,
-95% → 29% of answers truncated at the token cap (smoke, 25 problems × 4 samples).
+weights everywhere else, same prompt, same sampling. Verified on the full MATH-500 (500 problems
+× 4 samples): **51.0% accuracy** (post: 48.1) with **31.6%** of answers truncated at the token cap
+(post: 33.1); the base model manages 5% with 95% truncated.
 
 *Live runs cap the answer length for demo speed (default 4096 tokens; the verification used 32,768).
 The base model's loop is usually visible within the first ~2,000 tokens. First call may take a