update make_card.py

code/make_card.py

@@ -49,7 +49,22 @@ language:
 - en
 ---
 
-# openjev — Qwen3.5 as an NLI cross-encoder
+# openjev — Qwen3.5 trained as jev model
+
+Zero-shot Doom, played by the cross-encoder from the text state (hypotheses "the nearest enemy is left of / right of /
+exactly on the crosshair" → turn left / turn right / attack; 11 kills on average, oracle 18.8, ~57 ms per decision):
+
+<video controls src="https://huggingface.co/AlexWortega/openjev/resolve/main/videos/doom_zs_position.mp4" width="720"></video>
+
+From pixels (the frame goes in as image tokens through the Qwen3.5 vision tower, hypotheses about the monster's
+x-position; 5.2 kills, still zero-shot):
+
+<video controls src="https://huggingface.co/AlexWortega/openjev/resolve/main/videos/doom_vision_zeroshot_pixels.mp4" width="720"></video>
+
+Zero-shot multiple-choice accuracy against Jev and Terra (their numbers read off their published chart; Chess and
+GSM8K-k-choice are our constructions):
+
+![radar](assets/radar_openjev.png)
 
 A Qwen3.5 decoder turned into a 3-way NLI cross-encoder in the spirit of
 [dleemiller's "NLI cross encoders: ways to use them"](https://huggingface.co/blog/dleemiller/nli-xenc-ways-to-use).