JEV GUIDE

Jev content moderation: Discord spam detection and false positives

Moderation starts with a community policy and specific conditions to evaluate. Using Discord as an example, this guide separates model assessment, enforcement and human review. It is not a validated automatic banning recipe.

Jev Gallery · Reviewed 2026-09-19

Separate assessment from enforcement

TypeSafe’s guardrails cookbook separates hazard questions from routing thresholds. The application chooses an outcome such as allowing or reviewing a message; the assessment itself does not enforce policy.

Editorial recommendation: begin in observation mode, review false positives in the actual channel, then decide which categories may trigger automatic actions.

Source: TypeSafe · Guardrails for LLMs ↗

What does the Discord case implement?

Jev Moderation Bot evaluates spam or scam signals in context. The bot can delete messages, warn or time out members and provides an administrator pardon path. We reviewed a pinned revision, without measuring false positives in a live community.

Source: Jev Moderation Bot · pinned source ↗

Distinguish discussion and quotation from promotion

The following synthetic policy examples contain no model scores. The same words may appear in promotion, a report or a discussion about avoiding scams. Provide the relevant context and community policy.

Distinguish discussion and quotation from promotion
Scenario	Suggested handling
A member reports an advertisement	Preserve reporting context and follow the reporting flow
Repeated off-topic promotion	Check channel rules and repetition before enforcement
Model or network error	Mark unreviewed and apply an explicit failure policy

How should thresholds and false positives be assessed?

Sample across languages, channel types and categories. Track removal of legitimate messages separately from missed violations. Use administrator reversals in later evaluation while keeping tuning and evaluation samples separate.

Do not copy a generic threshold into production without checking it. Define timeouts, appeals or reversals, event deduplication and moderation permissions. Test text that tries to influence its own classification; model assessment does not replace these application controls.