auto-sync 2026-07-04T05:50:49Z workspace
Browse files- workspace/README.md +22 -0
workspace/README.md
CHANGED
|
@@ -675,6 +675,28 @@ Key values:
|
|
| 675 |
Interpretation: the LCB fallback records the requested safety/calibration
|
| 676 |
fields, but it does not solve dominance. It is a negative Part-G diagnostic.
|
| 677 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 678 |
## How To Run Core Commands
|
| 679 |
|
| 680 |
Use the local virtual environment when possible:
|
|
|
|
| 675 |
Interpretation: the LCB fallback records the requested safety/calibration
|
| 676 |
fields, but it does not solve dominance. It is a negative Part-G diagnostic.
|
| 677 |
|
| 678 |
+
### Active utility-energy selector job
|
| 679 |
+
|
| 680 |
+
Submitted on 2026-07-04:
|
| 681 |
+
|
| 682 |
+
```text
|
| 683 |
+
15177546 scripts/slurm/train_utility_energy_selector.sbatch array=0-2
|
| 684 |
+
```
|
| 685 |
+
|
| 686 |
+
Purpose: train three `base_context_obs` utility-energy checkpoints on the
|
| 687 |
+
train-only RGB-ref chart database, then evaluate checkpoint-scored calibrated
|
| 688 |
+
dominance on the strongest current measured support setting:
|
| 689 |
+
|
| 690 |
+
```text
|
| 691 |
+
calibration: runs/ctt_base_context_obs_train_cal_envclip_k16_rollout_comparison
|
| 692 |
+
evaluation: runs/ctt_base_context_obs_test_envclip_k16_rollout_comparison
|
| 693 |
+
K: 16
|
| 694 |
+
```
|
| 695 |
+
|
| 696 |
+
This job targets the current selector gap, not the support side. The run must
|
| 697 |
+
be interpreted as a Part-G diagnostic unless selected success and selector gap
|
| 698 |
+
beat the existing train-clean K=16 selector.
|
| 699 |
+
|
| 700 |
## How To Run Core Commands
|
| 701 |
|
| 702 |
Use the local virtual environment when possible:
|