这是一篇由原始材料转换而来的阅读页，保留了源文件的主要结构，并补充了可追溯的来源说明与链接。

Pattern: Rollback

Purpose

Rollback ensures failed or non-beneficial changes do not poison the current best-known state.

Principle

Autoresearch assumes many experiments will fail. The system must make failure cheap.

When to rollback

Rollback when: - the primary metric gets worse - the run crashes in a way that invalidates the idea - complexity increases without sufficient gain - resource cost rises beyond acceptable bounds - the run violates policy or scope

Keep / discard / retry

These are distinct decisions:

• keep: new best or justified improvement
• discard: valid experiment, but not worth advancing
• retry: execution noise or trivial implementation issue makes one more run reasonable

Retries should be limited.

Implementation strategies

Typical mechanisms: - keep a best-known commit pointer - commit every experiment before running - reset to best-known state after discard - log every rollback decision

Why rollback is essential

Without rollback: - bad changes accumulate - attribution becomes impossible - search quality degrades over time - the system becomes harder to trust

Anti-patterns

• keeping all changes “just in case”
• failing to distinguish crash from discard
• retrying the same weak idea repeatedly
• making rollback manual and inconsistent

Recommendation

Treat rollback as a first-class part of the loop, not as an exception path.