Recursive Governance: When Agents Write the Rules They Execute
Описание и применение
Система динамического управления правилами, где агенты могут создавать, изменять и отменять правила, предотвратила случайное удаление 40 модулей, обеспечив проверку согласованности и жизненный цикл правил.
Задача
A static sync filter file caused accidental deletion of ~40 private modules because it missed a pattern after a directory restructure.
Результат
The dynamic governance system prevented accidental deletion of private modules, introduced rule lifecycle management, and reduced accidental policy rot.
Рабочий процесс
- 1. Agent proposes a rule via propose_rule with rationale, owner, and affected namespaces.
- 2. Agent amends existing rules via amend_rule; deltas are diffed, never overwritten.
- 3. Agent ratifies proposals with ratify_rule, requiring per-shard quorum and no vetoes.
- 4. Human or agent vetoes via veto_rule, with justification logged.
- 5. check_consistency scans rule graph, computes embedding similarity, and applies allow/deny list to block namespace violations.
- 6. Two‑phase commit: observation window (24h or N messages) for rebuttals, then ratification if quorum met and consistency passes.
- 7. Governance rules stored in append‑only namespace governance://rules/, with parent references for replay.
- 8. Periodic cleanup: replay history, archive obsolete rules, re‑ratify active ones.
Доказательства
We almost lost forty modules to a file that never changed.
The filter file was a static text file. It encoded what the sync could delete and what it couldn't, and it had no idea that the real inventory had changed.
If a rule can't be archived, it keeps counting toward quorum and keeps confusing the live ones.
check_consistency scans the existing rule graph, computes an embedding similarity between the new rule surface and the rule surfaces already in effect, then applies a short, human-maintained allow/deny list on tool names.
- Дата записи: 2026-10-06T03:40:48.779Z