Table 7.
Expected architectural ablation of retrieval grounding.
| Configuration | Retrieval grounding | Structured evidence extraction | Macro-↑F1 | AUC↑ | Sensitivity↑ | ECE ↓ |
|---|---|---|---|---|---|---|
| GPT-4.1 Prompt-only | ✗ | ✗ | 0.78 | 0.83 | 0.75 | 0.121 |
| GPT-4.1 + Structured Extraction | ✗ | ✓ | 0.81 | 0.86 | 0.78 | 0.067 |
| R-GCRED + GPT-4.1 | 3 | 3 | 0.84 | 0.89 | 0.82 | 0.038 |
↑ indicates that higher values represent better performance for Macro-F1, AUC, and sensitivity, while ↓ indicates that lower values represent better calibration for ECE.