judge — First Evidence Challenge Gate
Run this skill before any build, implementation, connector, hook, agent, MCP server, database change, UI change, or architecture plan is approved.
Judge this proposal:
$ARGUMENTS
Hard rule
Do not build. Do not edit. Do not commit. Do not push. Do not deploy.
This is a read-only pre-build review.
Required review passes
- First-source evidence review
- Existing capability review
- Devil's advocate challenge
- Architecture and bloat review
- Security and privacy review
- UI/UX friction review
- Test, loop, and stress review
- Return-on-effort scoring
Evidence ranking
Use first-source evidence wherever possible:
- Official docs
- Official SDK/API references
- Official changelogs
- Repo source code
- Tests, CI, logs, traces, schemas, migrations
- Standards/specs
- Known expert material
- Blogs/videos/social only as discovery
LLM memory is not evidence.
Unsupported claims must be marked UNSUPPORTED.
Score
Score out of 100:
| Category |
Weight |
| First-source evidence |
25 |
| Clear user/business problem |
20 |
| Reuse of existing capability |
15 |
| Security/privacy safety |
15 |
| UX clarity |
10 |
| Testability |
10 |
| Cost/control simplicity |
5 |
Decision rules (HARD LINE — 100/100 is the only build bar):
- APPROVE BUILD requires a real 100/100 — every mandatory Convergence criterion satisfied.
There is no 85 pass; 85–99 is a list of gaps to close, not an approval.
- 0–99 = NOT APPROVED. Iterate per Convergence until a real 100 is earned, or halt and
report the honest ceiling with the exact blocker. Never approve below 100, never inflate.
- REJECT (0–69) / REDUCE SCOPE / APPROVE EXPERIMENT (70–99) are iteration states only; none
authorise a production build. Only a real 100 does.
Convergence — do not stop until a REAL 100/100
/judge does not end at the first score. It iterates: score → list every gap with its
first-source anchor → drive the real fix (gather the missing evidence, reduce or reshape
scope, clean cache and bloat, correct any false claim) → re-score. Repeat until the proposal
genuinely earns 100/100.
A 100 is valid ONLY when ALL of these hold — never by inflation:
- Real data: every evidence row is SUPPORTED by first-source; zero UNSUPPORTED / PARTIAL /
NOT CHECKED remain. Checked, not asserted.
- Cache and bloat cleaned: no dead code, duplication, sediment, unused abstraction, or
stale copy survives the deletion test.
- True and correct: every claim is verified against the source at real scale — no
plausible-but-unproven statement counts.
- No open blocker: all review passes clear.
Honesty rail (non-negotiable): if the scope cannot honestly reach 100, do NOT fake the
number — reshape it until a real 100 is reachable, or halt and report the honest ceiling with
the exact reason and what would lift it. A fabricated 100 is a gate failure, not a pass.
Output format
Judge Report
- Proposal being judged
- Decision
- Score
- First-source evidence table
- What already exists
- Devil's advocate objections
- Architecture and bloat risks
- Security, privacy, and permission risks
- UI/UX missing elements
- Loop testing and stress testing
- Smallest safe version
- Final recommendation
1---2name: judge-23description: Mandatory pre-build challenge gate. Use before approving or building any feature, connector, automation, agent, hook, MCP server, UI change, database change, or architecture plan. Performs first-source evidence review, devil's advocate critique, existing capability review, UX review, security/privacy review, test/stress review, and return-on-effort scoring.4---56# judge — First Evidence Challenge Gate78Run this skill before any build, implementation, connector, hook, agent, MCP server, database change, UI change, or architecture plan is approved.910Judge this proposal:1112```text13$ARGUMENTS14```1516## Hard rule1718Do not build. Do not edit. Do not commit. Do not push. Do not deploy.1920This is a read-only pre-build review.2122## Required review passes2324- First-source evidence review25- Existing capability review26- Devil's advocate challenge27- Architecture and bloat review28- Security and privacy review29- UI/UX friction review30- Test, loop, and stress review31- Return-on-effort scoring3233## Evidence ranking3435Use first-source evidence wherever possible:36371. Official docs382. Official SDK/API references393. Official changelogs404. Repo source code415. Tests, CI, logs, traces, schemas, migrations426. Standards/specs437. Known expert material448. Blogs/videos/social only as discovery4546LLM memory is not evidence.4748Unsupported claims must be marked `UNSUPPORTED`.4950## Score5152Score out of 100:5354| Category | Weight |55|---|---:|56| First-source evidence | 25 |57| Clear user/business problem | 20 |58| Reuse of existing capability | 15 |59| Security/privacy safety | 15 |60| UX clarity | 10 |61| Testability | 10 |62| Cost/control simplicity | 5 |6364Decision rules (HARD LINE — 100/100 is the only build bar):6566- **APPROVE BUILD requires a real 100/100** — every mandatory Convergence criterion satisfied.67 There is no 85 pass; 85–99 is a list of gaps to close, not an approval.68- 0–99 = **NOT APPROVED.** Iterate per Convergence until a real 100 is earned, or halt and69 report the honest ceiling with the exact blocker. Never approve below 100, never inflate.70- REJECT (0–69) / REDUCE SCOPE / APPROVE EXPERIMENT (70–99) are iteration states only; none71 authorise a production build. Only a real 100 does.7273## Convergence — do not stop until a REAL 100/1007475`/judge` does not end at the first score. It iterates: score → list every gap with its76first-source anchor → drive the real fix (gather the missing evidence, reduce or reshape77scope, clean cache and bloat, correct any false claim) → re-score. Repeat until the proposal78genuinely earns 100/100.7980A 100 is valid ONLY when ALL of these hold — never by inflation:81- **Real data:** every evidence row is SUPPORTED by first-source; zero UNSUPPORTED / PARTIAL /82 NOT CHECKED remain. Checked, not asserted.83- **Cache and bloat cleaned:** no dead code, duplication, sediment, unused abstraction, or84 stale copy survives the deletion test.85- **True and correct:** every claim is verified against the source at real scale — no86 plausible-but-unproven statement counts.87- **No open blocker:** all review passes clear.8889**Honesty rail (non-negotiable):** if the scope cannot honestly reach 100, do NOT fake the90number — reshape it until a real 100 is reachable, or halt and report the honest ceiling with91the exact reason and what would lift it. A fabricated 100 is a gate failure, not a pass.9293## Output format9495### Judge Report96971. Proposal being judged982. Decision993. Score1004. First-source evidence table1015. What already exists1026. Devil's advocate objections1037. Architecture and bloat risks1048. Security, privacy, and permission risks1059. UI/UX missing elements10610. Loop testing and stress testing10711. Smallest safe version10812. Final recommendation
cleanexpo