updated

README.md

@@ -17,14 +17,8 @@ an efficiency-aware benchmark for AI-generated InsForge SDK code.
 
 Trained with **Unsloth** (bf16 LoRA) + an agentic GRPO loop adopting **CUDA-Agent**
 (arXiv 2602.24286): the model writes a solution, the forger-bench grader runs+verifies+
-measures real server cost (rows scanned, bytes, round-trips, scale-correctness at 100k
-rows), and a discrete milestone reward (-1 incorrect/scaleBug, 1 wasteful, 2 beats-naive,
-3 near-optimal) drives RL.
-
-## Stages
-- SFT on author+optimize pairs across all backend concepts (contamination-controlled:
-  fresh entities disjoint from the test set; top_n + in_list held out for generalization).
-- Agentic GRPO against the live forger-bench grader.
+measures real server metrics, and a discrete milestone reward (-1 incorrect/scaleBug, 1
+wasteful, 2 beats-naive, 3 near-optimal) drives RL.
 
 ## Contamination control
 Never trained on a sealed test task; held-out concepts measure optimization skill vs