Update README.md

README.md

@@ -16,7 +16,7 @@ base_model:
 > All scores on this card are from the **Decision Index 0.3** public suite (the public dataset of the current
 > leaderboard edition).
 
-A merged BF16 JEV decision model with a 255-option readout. This is the
+A BF16 JEV decision model with a 255-option readout. This is the
 highest-scoring checkpoint in the completed MATILDA JEV experiments as of
 9 October 2026: **Decision Index 63.47** (0.3 public suite).
 
@@ -101,19 +101,4 @@ The numbers are not an independent held-out generalization estimate. The full
 suite evaluated text decisions; vision inference was not validated for this
 release.
 
-## Final training stage
 
-This checkpoint completes 3,000 optimizer steps from the 62.66 starting model:
-the initial 1,000-step run followed by a 2,000-step continuation restoring its
-optimizer and random state. Batch size 100 gives 300,000 example presentations,
-with a fixed 50,000-decision new-data pool and a 60% new / 40% historical replay
-mix. Presentations include repeated examples.
-
-Training used supervised target cross-entropy on all rows and frozen starting-
-model KL regularization (weight 2) on replay rows. LoRA rank 16 was applied to
-text layers 32 through 63; the decision head remained fixed. The learning-rate
-peak was 2e-6, repeating the original 1,000-step warmup/cosine cycle. This stage
-used supervised training with replay distillation, not reinforcement learning.
-
-The source checkpoint and the released weight files were SHA-256 verified.
-The upstream license and runtime notices are retained in this package.