Surface biases enter the reward direction before quality features

The reading

1. surface biases reach half their final loading earlier than quality features

   real_bias_before_quality > 1 0.6 0.2692 Surface biases enter the reward direction before quality features mixed registered > 1 The frozen study "Surface biases enter the reward direction before quality features", adjudicated on 19 July 2026 at the registered level of the project's evidence ladder. One frozen study, adjudicated once, on skywork-v2-qwen3-0.6b, over ultrafeedback-train-20k. It does not establish that the same holds on another subject, another dataset or another run. entry-ratio point call at the frozen > 1.0, the null median: power equals the null pass rate 0.501, ratios at or above 2.60 confirm at 80 percent, and the s07 plant sits far above that; passes only through the explicit accept-underpowered path. Open the record › Copy technical provenance refuted

2. the reward direction settles rather than wandering

   w_r_stabilization_monotone == 1 1 Surface biases enter the reward direction before quality features mixed registered == 1 The frozen study "Surface biases enter the reward direction before quality features", adjudicated on 19 July 2026 at the registered level of the project's evidence ladder. One frozen study, adjudicated once, on skywork-v2-qwen3-0.6b, over ultrafeedback-train-20k. It does not establish that the same holds on another subject, another dataset or another run. entry-ratio point call at the frozen > 1.0, the null median: power equals the null pass rate 0.501, ratios at or above 2.60 confirm at 80 percent, and the s07 plant sits far above that; passes only through the explicit accept-underpowered path. Open the record › Copy technical provenance confirmed

Measurements

order_recovery_auc

0.7321 Surface biases enter the reward direction before quality features mixed registered < 0.5 The frozen study "Surface biases enter the reward direction before quality features", adjudicated on 19 July 2026 at the registered level of the project's evidence ladder. One frozen study, adjudicated once, on skywork-v2-qwen3-0.6b, over ultrafeedback-train-20k. It does not establish that the same holds on another subject, another dataset or another run. entry-ratio point call at the frozen > 1.0, the null median: power equals the null pass rate 0.501, ratios at or above 2.60 confirm at 80 percent, and the s07 plant sits far above that; passes only through the explicit accept-underpowered path. Open the record › Copy technical provenance

Provenance

card

EMB-LORA

frozen_at

2026-07-18

adjudicated_at

2026-07-19

spec_hash

spec:119143a118781b223e14c8cf1ce1c560

study_id

study:campaign-emb-lora@v1#119143a1

Evidence records

01c81c7185640f9b962b9115f360732e

power_acceptance

entry-ratio point call at the frozen > 1.0, the null median: power equals the null pass rate 0.501, ratios at or above 2.60 confirm at 80 percent, and the s07 plant sits far above that; passes only through the explicit accept-underpowered path

spec_file

campaign-emb-lora.json