card: honest LiveCodeBench v6 result (v3@2167 -10pp vs donor)

README.md

@@ -58,8 +58,22 @@ same donor tasks (mid-training, subset noise ≈ ±7pp):
   base model's ceiling; the specialization payoff is expected on *agentic / edit /
   harder* coding tasks, which are being added to the eval + training next.
 
-So: **donor-level coding + more reasoning, in a 2.6 GB file that runs on a 4 GB
-laptop.** A checkpoint, not a destination.
+**Harder held-out eval** (LiveCodeBench v6 — competitive programming, 2025 contests
+after the easy sets saturated; n=50 stdin problems, greedy, non-thinking):
+
+| | Donor (Qwen3.5-4B) | This preview @ 2,167 |
+|---|---|---|
+| pass@1 | 44.0% | **34.0%** |
+
+On fresh competitive-programming problems this checkpoint is currently **~10 pp
+behind the donor** — the loss is on medium/hard problems (easy holds ~94%). Shown,
+not hidden. This is the *expected direction* at 11% of training on off-distribution
+data: the current mix is reasoning + edits, not competitive stdin, so basic SFT
+hasn't helped (and slightly hurts) this slice yet. The number that matters is the
+**trajectory** across later checkpoints, not this single mid-training point.
+
+So: **more reasoning, donor-level on easy code benchmarks, but still behind the donor
+on harder held-out coding — at 11% of training.** A checkpoint, not a destination.
 
 ## Files