Catalog
SYS-2SYSP0/MSpec-levelPROPOSED

ModificationGate 引入 VC 容量上限 (Capacity Bound) 双门机制:防过拟合与泛化丧失

Evaluation modality

Spec-level

A spec-motivation / governance borrow. Evaluated by spec review + contract tests, not A/B or ablation.

Primary owner
Phase-A verdict
Shadow profile
Source papers
Two-Gate Guardrail 2025
Specs
docs/specs/credit-and-self-modification.md

Blind spot (现状盲点)

DM-4 强调了 ModificationGate 必须有"可验证的 Evaluation"(Validation Margin)。但这只防了"改错了",没防"改得过拟合"。如果一次 rare-heavy 更新改变了太多的模型参数或逻辑容量,系统就会丧失泛化保证。

Adoptable suggestions (可落地动作)

  1. 1.在 [`docs/specs/credit-and-self-modification.md`](../specs/credit-and-self-modification.md) 的 ModificationGate 中引入 **Capacity Cap(修改容量上限)** 概念。PROPOSED

    Not a runnable A/B candidate — evaluated by the path above, not ablation.

  2. 2.将 ModificationGate 升级为"双门"机制:第一道门看分数(Validation Margin,即 DM-4 的可验证提升),第二道门看容量(Capacity Cap,即本次修改的 VC 维或参数变动量是否超标)。PROPOSED

    Not a runnable A/B candidate — evaluated by the path above, not ablation.

  3. 3.结合 EWC(Elastic Weight Consolidation)等算法,量化并限制每次自修改对系统整体容量的占用。PROPOSED

    Not a runnable A/B candidate — evaluated by the path above, not ablation.

Traceability

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

Expected benefit (预期收益)

- 在 PAC 学习框架下为系统的长期自修改提供严格的泛化保证。 - 防止系统在持续学习中因为单次过度拟合而导致灾难性遗忘或能力塌陷。

Cited paper (引用论文)

**C1-01. Two-Gate Guardrail for Self-Modifying Agents. arXiv:2510.04399, 2025.** - 关键观点:形式化证明自修改 agent 要保留 PAC 学习保证,需要 policy-reachable 模型族 VC 维有界。提出"validation margin + capacity cap"双门机制。 ---