Upload README.md with huggingface_hub

README.md

@@ -107,18 +107,19 @@ Weights learn **how to screen**. Facts that move (SOP revs, notices) stay in the
 
 ## ScreenBench
 
-Same code grader. `verified: false`. Lab board, not the hero.
+Same 12 files. Same grader. `verified: false`.
 
-![Holdout](holdout.png)
+![Frontier vs Trace](compare.png)
 
 | Seat | All-pass | False kill |
 |---|---|---|
-| Qwen 7B base | 0.00 | 0 |
-| Claude | 0.25 | 0 |
-| Weights alone | 0.50 | 0 |
+| Qwen 2.5 7B | 0.00 | 0 |
+| gpt-oss 20B | 0.00 | 0.08 |
+| Gemma 4 31B | 0.00 | 0 |
+| Claude (this login) | 0.08 | 0.08 |
 | **Trace** | **1.00** | **0** |
 
-12/12 is Trace the product (weights + lake + runner). 6/12 is the adapter with the runner off. Codex and Grok are not rows.
+Trace is weights + dated lake + runner (12/12). Naked adapter is 0.50. Codex and Grok are not rows.
 
 ## Quickstart