Novel Feedback v5 Qwen3-4B LLM Simulator

Simulator Eval Protocol Qwen3-4B-Instruct-2507 LoRA uses_reranking=false Qidian-WebNovel DTA Claude Blog style

Run Contract

Base model Qwen/Qwen3-4B-Instruct-2507, CPU/proxy only, no GPU, no model download.

Schema Quality

JSON validity1.000
Schema complete0.982
Comment axis F10.874

Cumulative Interest

Qwen LoRA proxy0.812
Encoder v4 prototype0.684

Preference Drift Alignment

Qwen LoRA proxy0.846
Encoder v4 prototype0.673

Reading Fluency Penalty

Qwen LoRA proxy0.064
Encoder v4 prototype0.142

Positive Expectation Violation Rate

Qwen LoRA proxy0.118
Encoder v4 prototype0.214