# VERIFY — moe100m-physics tiny-BPE

**ALL PASS: False**

- [x] **1_generation_sanity**: parsed 1 obj from greedy continuation; sample='Frame 5: in motion\n  obj_0: pos=(94.7987, 92.7295), vel=(223.8441, 234.7186)\nFrame 6:'
- [x] **2_loss_sanity**: train_loss(last50 avg)=1.2204 (raw last-step=1.7107) in band [0.62,6.24]
- [x] **3_eval_tracks_train**: |best_eval 1.6926 - train_loss(last50 avg) 1.2204| = 0.4722 (<0.5)
- [ ] **4_data_consumption**: consumed 1.4e+08/7e+08 = 19.9% (>=70%)
- [ ] **5_stderr_scan**: stderr hits=[]; aborted=True. (NaN-skip+recover is tolerated)
- [x] **6_param_count**: active=92.83M (target 100M +/-15% -> [85,115])

## Eval summary (from EVAL_RESULTS.json)
- trained:  @15f=5.548%  @80f=5.548%
- held_out: @15f=6.753%  @80f=6.753%
- orbit @80f: 0.612%
- LFM2-350M baseline: @15f trained 0.38% / held-out 0.93%; orbit 0.75% @80f