Update docs with latest architecture and results

2026-01-28 11:23:50 +08:00
parent e2974342d5
commit 34d6f0d808
5 changed files with 37 additions and 4 deletions
--- a/docs/ideas.md
+++ b/docs/ideas.md
@@ -11,3 +11,6 @@

 ## Two-stage training with curriculum
 - Hypothesis: train diffusion on residuals only after temporal GRU converges to low error.
+
+## Discrete calibration
+- Hypothesis: post-hoc calibration on discrete marginals can reduce JSD without harming KS.