EMB-LORA S7 · Reward embryology and pretraining origins
Surface biases enter the reward direction before quality features
Mixed
frozen 2026-07-18 · adjudicated 2026-07-19 · ev:01c81c7185640f9b962b9115f360732e · spec:119143a118781
The frozen predictions
- H-bias-first surface biases reach half their final loading earlier than quality features
real_bias_before_quality > 1registered p 0.6 real_bias_before_quality Value 0.2692307692307692 Subject EMB-LORA Threshold > 1 Trust Adjudicated Evidenceev:01c81c7185640f9b962b9115f360732eRefuted - H-stabilize the reward direction settles rather than wandering
w_r_stabilization_monotone == 1w_r_stabilization_monotone Subject EMB-LORA Threshold == 1 Trust Adjudicated Evidenceev:01c81c7185640f9b962b9115f360732eConfirmed
Registered companions
-
order_recovery_aucorder_recovery_auc Value 0.7321428571428571 Subject EMB-LORA Trust Adjudicated Evidenceev:01c81c7185640f9b962b9115f360732e
