Judgment Gate
Explicit independent judgment for planning decisions — the planning
counterpart of judgment-day (which owns adversarial code review).
Never run both on the same target.
HARD-GATE
- Judgments inform; the human decides. A judgment carries no delivery
authority and satisfies no commit, push, or release gate by itself.
- Frame questions narrow enough to answer independently. "Is this plan
good?" is not a question; "Does the plan name an owner for every P0
task?" is.
- At most two rounds of fix-and-re-judge. Anything unresolved after
that escalates to a human; never extend the budget.
- Judges are independent: separate context, identical frozen scope.
Never accept a partial judgment.
Process
- Freeze the target. One complete, immutable artifact (plan,
backlog, design doc) with a stated decision to make.
- Frame the questions. Yes/no with explicit criteria, or ordered
levels (e.g. risk: low/medium/high with definitions). 3–8 questions;
each judge sees the same set.
- Judge independently. Run two judges blind, or one judge plus a
differently-framed pass. Record verdicts and confidence per question.
- Merge. Agreement → act. One judge dissents → record as suspect,
do not auto-fix. Contradiction → escalate for explicit human decision.
- Fix and re-judge (bounded). Fix only confirmed findings; re-judge
only the frozen ledger plus the delta. Repeat once at most.
Question framing guide
| Decision shape |
Frame as |
| Include/exclude several items |
One independent yes/no per item, rank by confidence |
| Single winner among options |
One pick per judge with reasons; disagreement = escalate |
| Graded setting (risk, effort, readiness) |
Ordered levels with written definitions per level |
Decision gates
| Condition |
Action |
| Target or decision unclear |
Ask one scope question and stop |
| Both judges agree |
Proceed or fix, per verdict |
| One dissents |
Record suspect; human decides whether to pursue |
| Judges contradict |
Escalate; never average away a contradiction |
| Issues remain after round two |
Escalate and stop |
Output contract
Return: target identity, the question set, per-question verdicts with
confidence, confirmed/suspect/contradiction counts, and exactly one
terminal line: DECISION: PROCEED ✅, DECISION: REVISE 🔁, or
DECISION: ESCALATE ⚠️.
Integration
| Skill |
When to chain |
review-prd |
Gate a PRD before plan-tickets drafts work items |
prioritize-backlog |
Independent judgment on a ranking |
llm-judgment-layer |
Same discipline applied inside an app |
