Rewrite Caveats section as Evaluation details; strip personal-context and apologetic framing

README.md

@@ -87,13 +87,12 @@ Same 20-domain class-agnostic AR@100 protocol as Argus, evaluated live through t
 
 Per-domain numbers live in `rf100vl_results.json`.
 
-## Caveats
+## Evaluation details
 
-These are first-pass trained heads at conservative schedules (one-shot overnight training run). None are at peer-reviewed publication quality:
-
-- Classifier val top-1 is **79.13 %**, top-5 95.53 % on 50K ImageNet val 2012 images, using the TensorFlow Models repo's synset-label mapping for ground truth. Beats the EUPE-ViT-S paper kNN baseline (78.2).
-- Detection head has no mAP evaluation in this release; loss curve converged but COCO-eval requires a separate pass.
-- Depth head is a simpler architecture than Argus-B's DPT; full DPT with hooked intermediate blocks is a natural next iteration.
+- Classifier val top-1 is **79.13 %**, top-5 **95.53 %** on 50K ImageNet val 2012 images, using the TensorFlow Models repo's synset-label mapping for ground truth. Above the EUPE-ViT-S paper kNN baseline (78.2).
+- Detection head ships training-loss curves and the RF100-VL cross-domain AR@100 number above; COCO val2017 mAP is a future addition.
+- Depth head here is a BN + two-conv + 256-bin variant. A full DPT decoder reassembling the four hooked block activations from the cache is the natural next iteration.
+- Segmentation head is a linear probe at 5 epochs; the EUPE-ViT-S paper reports mIoU 0.466 at a much longer schedule.
 
 ## Source backbone