Published October 28, 2026
| Version 1.3
Preprint
Open
同源失败相关实验协议(预注册 v1.3)
Authors/Creators
- 1. Anhui Institute of Information Technology, School of Computer Science and Technology
Description
Preregistered protocol for a five-family bare-task failure-correlation experiment. Version 1.3 adds a dual-anchor design (cross-model arms anchored on both GLM-5.3 and DeepSeek-Flash, so the contrast does not depend on one anchor family) and records six corrected task defects found in pilot calibration; predictions R1-R3 and all statistical clauses are unchanged since v1.0.
Arms: same-model pairs (GLM-5.3, DeepSeek-Flash), cross-model pairs (DS×Qwen, DS×Kimi, DS×GPT, GLM×DS/Qwen/Kimi/GPT), within-family variant pair (GLM-5.3×GLM-5.3-Flash, exploratory), single-seat baselines. 30 bare tasks × 30 repetitions per arm. Scoring 100 percent scripted, no LLM judge. Skill covariate and conditional analysis per Jha 2026.