COG-1COGP0/MRunnable A/BPROPOSED

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

—

Evaluation modality

Runnable A/B

Compiles to a runnable VolvenceZero profile. This is the only modality that goes to SHADOW A/B + ablation.

Primary owner: —
Phase-A verdict: —
Shadow profile: counterfactual-credit
Source papers: COCOA 2023 + Least-control 2022 + Deep Feedback Control 2021
Specs: docs/specs/credit-and-self-modification.mddocs/specs/dual-track-learning.mddocs/specs/evaluation.md

Blind spot (现状盲点)

当前 26 条里，R10 的焦点主要是 **ModificationGate 是否能开门**，R12 的焦点主要是 **评估是否可验证**。但关系型数字生命体还需要回答更细的问题：一次 repair 成功 / 关系破裂 / 承诺兑现，究竟是哪一个 turn、哪个 regime、哪条 commitment、哪类 social action 贡献了结果？如果只看终局分数或 generation winrate，credit 会过粗，慢反思无法知道该强化什么。

Adoptable suggestions (可落地动作)

1.在 [`docs/specs/credit-and-self-modification.md`](../specs/credit-and-self-modification.md) 增加 "counterfactual contribution credit" 候选：以 COCOA 的反事实问题为模板，记录"若不采取该 action / regime / commitment update，结果是否仍会发生"。PROPOSED
A/B (dry-run SMOKE)
2.将 **least-control principle** 作为 report-only 指标：好的适应不是上层干预越来越强，而是系统达成同等关系 / 任务结果所需控制量下降。PROPOSED
A/B (dry-run SMOKE)
3.第一阶段只进入 SHADOW evidence，不直接影响 online-fast 更新；与 DM-7 / EVO-6 的 head-to-head 评估串联，给 ModificationGate 提供更细粒度归因。PROPOSED
A/B (dry-run SMOKE)

Traceability

No plugins / runs linked yet. Scaffold a suggestion to start.

Expected benefit (预期收益)

- 把 R9 从"PE 聚合"推进到"长期关系结果的可解释贡献链"。 - 让 ReflectionEngine 不只总结"哪里 PE 高"，还能给出"哪类内部动作真的造成了结果变化"的 evidence。 - 避免把一整段关系改善错误归因给最后一句表达层措辞。

Cited paper (引用论文)

**COCOA: Counterfactual Contribution Analysis**（NeurIPS 2023 spotlight）；**The least-control principle for local learning at equilibrium**（NeurIPS 2022）；**Deep Feedback Control**（2021）。详见 [`research/core-author-paper-assessment-2026-05.md`](../../research/core-author-paper-assessment-2026-05.md) §四。 ---