-
Notifications
You must be signed in to change notification settings - Fork 2
Discipline Gates
type: concept up: "Three-Operations" extends: "Three-Operations" tags: [policy, safety]
A short list of rationalizations that are universally wrong in wiki ingest work, paired with the discipline that defeats each one. Introduced by PR #1 alongside the Verification-Gate.
When the agent is about to skip a step ("the index does not really need updating this time", "this finding is not novel enough to file"), the rationalization usually falls into one of a few recurring patterns. Naming them makes them visible.
- "This change is small enough to skip the log." No. The log is append-only by design; small changes are exactly what it should record.
- "The cross-reference is implied; readers will infer it." No. Implied means missing; future agents will not infer it.
- "I will fix the frontmatter later." No. Frontmatter added later loses the context that would have informed it.
- "The verification gate is overkill for this one-line edit." No. The gate is cheap; the drift it prevents is not.
- "This finding is half-baked; I will file it once it is stronger." No. The wiki tolerates uncertain claims with appropriate framing; what it cannot tolerate is silent retention in chat context that disappears at session end.
The Verification-Gate checks concrete state: is the index updated, is the log appended. Discipline gates check agent reasoning: am I about to talk myself out of doing the right thing. They are complementary: the verification gate stops bad commits; the disciplines stop bad rationalizations from reaching the gate.
The file wiki/agents/discipline-gates.md is overlay-independent. Every agent's procedure references it, the same DRY mechanism used by the verification gate. See Agent-Overlays.
- Verification-Gate
- Three-Operations
- Lesson-Validation-Methodology
- Limitation-Prompt-Injection-Residual: complementary content-discipline mechanism at the wiki layer
- Lesson-Template-Contribution-Discipline: the parallel set of universal rationalizations for AI agents contributing PRs to the template itself (scope creep, speculative PR numbers, state-locking test assertions, single-writer default)