Research catalog
The 43 VolvenceZero research candidates imported from docs/moving forward/探索方向.md. Each candidate carries adoptable suggestions that can be turned into plugins + A/B runs.
| Code | Title | Group | Priority | R-IDs | Evaluation | Status | ROI |
|---|---|---|---|---|---|---|---|
| COG-1 | 反事实信用分配 + least-control：把"谁造成了长期关系结果"做成 credit 主线3 actions | COG | P0/M | — | Runnable A/B | PROPOSED | — |
| COG-2 | Social Cognition / ToM owner 学习：belief / intent / feeling / preference 在线维护4 actions | COG | P0/M | — | Runnable A/B | PROPOSED | — |
| COG-3 | Persona / Regime Geometry：regime 与 value drift 的 read-only 几何监控3 actions | COG | P0/S-M | — | Runnable A/B | PROPOSED | — |
| EVO-2 | Evaluation cascade（廉价→昂贵）+ LLM-judge **仅 readout**（自然性/简洁性）3 actions | EVO | P0/M | — | Spec-level | PROPOSED | — |
| EVO-6 | Cross-generation winrate / head-to-head 作为 ModificationGate 硬证据（与 DM-7 合并叙述）0 actions | EVO | P0/S | — | Spec-level | PROPOSED | — |
| OA-1 | spec motivation 引用 N3/N4/N6 实证3 actions | OA | P0/S | — | Spec-level | PROPOSED | — |
| OA-2 | 在 expression-layer spec 形式化 Mind/Face 隔离3 actions | OA | P0/S | — | Spec-level | PROPOSED | — |
| OA-3 | ModificationGate 加 framing-aware 检查（inoculation 工程化）6 actions | OA | P0/M | — | Spec-level | PROPOSED | — |
| SYS-1 | 基于 CPD (Change Point Detection) 的 $\beta_t$ 涌现切换：用 PE spike + reward shift 无监督检测边界3 actions | SYS | P0/S | — | Runnable A/B | PROPOSED | — |
| SYS-2 | ModificationGate 引入 VC 容量上限 (Capacity Bound) 双门机制：防过拟合与泛化丧失3 actions | SYS | P0/M | — | Spec-level | PROPOSED | — |
| SYS-5 | Latent Action RL 作为控制器强化基线：在冻结 LLM 上挂极小控制器跑 RL3 actions | SYS | P0/L | — | Spec-level | PROPOSED | — |