Really is a truth-grounded execution protocol for AI agents.
Core line:
Facts before conclusions. Evidence before confidence. Verification before completion.
Highlights:
- Intent Lock: classify whether the user wants an answer, analysis, plan, scan, execution, verification, or artifact delivery.
- Evidence Budget: choose Light, Standard, or Strict evidence depth based on task risk.
- Truth Discipline: separate facts, inferences, unknowns, and recommendations.
- Source Audit: inspect relevant sources before changes and review actual changes afterward.
- Verification Gates: match verification to code, UI, state, data, files, or artifacts.
- Stop Conditions: stop and report when evidence, permission, or scope is insufficient.
中文短句：真查、真想、真验证、真说明限制。