Separation of fact/opinion in all deliverables + preservation of source URLs + blocking of evaluative vocabulary. Company-wide gate.
Citation Rules
- All qualitative citations:
source_url+captured_at+provenance_notepreserved - All quantitative signals: original SQL / calculation formula + input data range preserved
- When citing in the body:
(출처: <url>)format. No more than 5 citation sources (more than that goes in a separate appendix) - Images/charts: specify data source + collection date in the caption
Fact / Opinion Separation
| Category | Label | Example |
|---|---|---|
| Fact | (Source: ...) | "Average price of top 50 Musinsa Best items is 89,000 KRW (Source: musinsa.com ranking, 2026-05-21)" |
| Opinion | "Observed" / "Noted" / "Estimated" | "Observed — search trends peak in November. Possible seasonality." |
| Prohibited | "recommend", "buy", "good", "bad" | (Do not use) |
Reporting Format
- First line of comment:
사용 도구: <도구 목록>(예:사용 도구: Bright Data Web Unlocker + SERP API) - When a correction occurs:
정정 항목 N건 · 신뢰도 0~1.0specified - Result submission: must POST as a paperclip API comment (
POST /api/issues/{id}/comments)
Verification Checklist
- Does every figure in the body have a source URL or signal value attached?
- Are there 0 instances of evaluative vocabulary ("recommend"/"buy"/"good"/"bad")?
- Do opinion lines begin with the labels "Observed", "Noted", or "Estimated"?
- Are there 0 instances of plaintext API key exposure? (first 8 characters +
***masking) - Has the paperclip API comment POST been completed?
If any one of the above 5 is violated, refuse to report the result and self-correct.
Scope of Application
Brand Intelligence Lab 의 All agents inherit (Brand Researcher + Data Analyst + Slide Creator).