Upload README.md

README.md

@@ -1,5 +1,4 @@
 ---
-license: cc-by-nc-2.0
 language:
 - tr
 pretty_name: Turkish Last Exam (TLE)
@@ -35,14 +34,13 @@ One attempt per task. No tools, no web access, no system prompt. A task counts a
 
 | Model | Tasks solved (pass@1) | Sub-answer accuracy |
 |---|---|---|
-| **GPT 6.1 Sol** | **8 / 41 (19.5%)** | 343 / 497 (%69.0) |
+| GPT 6 Astra | 12 / 41 (29.3%) | 339 / 497 (%68.2) |
+| GPT 6.1 Sol | 8 / 41 (19.5%) | 343 / 497 (%69.0) |
 | Claude Opus 5.5 | 5 / 41 (12.2%) | 301 / 497 (60.6%) |
 | Gemini 3.7 Flash | 4 / 41 (9.8%) | 286 / 497 (57.5%) |
 | Claude Sonnet 5.5 | 0 / 41 (0.0%) | 221 / 497 (44.5%) |
 
-The gap between the two columns is the point of TLE. GPT sol gets most individual steps right, but almost never an entire chain. One wrong official rule early in a task (a threshold, an annex table, a scoring rule) carries through every later step.
-
-Gemini 3.7 Flash was run through Kaggle Benchmarks with the same protocol. It solved four tasks that Claude did not, which shows TLE is not tuned to a single model family. Evaluation of more model families on the full set is in progress.
+The gap between the two columns is the point of TLE. GPT 6 astra gets most individual steps right, but almost never an entire chain. One wrong official rule early in a task (a threshold, an annex table, a scoring rule) carries through every later step.
 
 ---