DeFi supervision ticket quality evaluation on DeXposure-Bench (Frozen 2025)
72Precisionpersist.+rules
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| persist.+rulesStack=persist.+rules, Interv.(wk)=0/29, Decision Model=Claude Opus 4.7, Judge Model=Claude Opus 4.82026.06 | 72 | 0.4 | 0.76 | — | 51.4 | — | |
| FM+rulesStack=FM+rules, Interv.(wk)=0/29, Decision Model=Claude Opus 4.7, Judge Model=Claude Opus 4.82026.06 | 60 | 1 | 1.9 | — | 25.7 | — | |
| FM+LLM+gateStack=FM+LLM+gate, Interv.(wk)=16/29, Decision Model=Claude Opus 4.7, Judge Model=Claude Opus 4.82026.06 | 58 | 1.2 | 2.34 | 43.7 | 43.5 | 2.45 | |
| snapshot LLMStack=snapshot LLM, Interv.(wk)=0/29, Decision Model=Claude Opus 4.7, Judge Model=Claude Opus 4.82026.06 | 57.5 | 0.9 | 1.84 | — | 53.2 | 2.24 | |
| FM+LLMStack=FM+LLM, Interv.(wk)=17/29, Decision Model=Claude Opus 4.7, Judge Model=Claude Opus 4.82026.06 | 57 | 1.2 | 2.41 | 44.8 | 48.8 | 2.41 |