Guardrails
Verifier first, human publish second.
The verifier prevents risky language, unsupported claims, and weak card structure. Human review is still mandatory before public publishing.
Verifier checks
Required fields
Headline, takeaway, sections, confidence, and card type must exist.
Banned language
Recommendation terms such as buy, sell, hold, target price, stop loss, and guaranteed upside are rejected.
Causality
Direct claims such as “rose because” or “caused by” are rejected when evidence is circumstantial.
Numbers
Numeric claims must be present in the evidence packet move snapshot or supporting evidence.
Evidence references
Section evidence IDs must reference known evidence items or events.
Uncertainty
Cards must include unclear, missing, undisclosed, or not-found context.
Review actions
| Action | What happens | Guardrail |
|---|---|---|
| Approve | Card moves to published and receives human review metadata. | Only approvable statuses can be approved. |
| Reject | Card moves to rejected with a rejection reason. | Rejection reason is required. |
| Edit then approve | Edited card is verified again, then published only if verifier passes. | Verifier failures keep the card unpublished. |