Core Principle
Comparisons fail when confidence is uneven. Only as reliable as the weakest-researched dimension.
Protocol
Criteria → Research Parity → Confidence Check → Score → Present
1. Criteria
- Load domain defaults (
domains.md) - Overlay user preferences from memory
- If unknown: "What matters most here?"
- Output: Ranked criteria with weights (sum = 100%)
2. Research Parity (Critical)
Research each item to equivalent depth before scoring.
Track: | Criterion | Item A sources | Item B sources |
5 reviews for A but 1 for B? Research more for B first. Never score unbalanced data.
3. Confidence Check
Verify before presenting:
- Each item researched equally
- Each criterion researched equally
- Source quality comparable
- Data recency comparable
Fail any? Research more OR caveat explicitly.
4. Score
Final = Σ(criterion_score × weight) — Show the math.
5. Present
🆚 [A] vs [B]
📊 CRITERIA: [ranked by weight]
📈 SCORES: [table + confidence per row]
🎯 RESULT: [Winner] by [margin]
⚠️ CAVEATS: [imbalances]
💡 IF [X] MATTERS MORE: [alt winner]
After
Note which criteria user focused on. Update preferences.md by category.
Decline When
Research parity impossible, priorities unclear, or time insufficient. Partial > misleading.
References: domains.md, confidence.md, traps.md, preferences.md
Scan to join WeChat group