Moderation Categories

Sieve evaluates content across seven moderation categories. Each category has a default threshold that determines when content is flagged or blocked. You can customize these thresholds per category to match your platform’s tolerance.

Categories

Toxicity

Default threshold: 0.7

Detects profanity, severe negativity, and hostile language that degrades conversation quality. This is the broadest category and catches general abusive language that doesn’t fit neatly into other categories.

Examples: Excessive profanity directed at others, aggressive hostility, deliberately inflammatory content.

Harassment

Default threshold: 0.7

Identifies personal attacks, bullying, and sustained targeting of individuals. Distinguished from general toxicity by the directed nature of the abuse.

Examples: Repeated personal insults, name-calling targeting a specific user, cyberbullying, dogpiling.

Hate Speech

Default threshold: 0.7

Detects slurs, discrimination, and dehumanizing language targeting protected characteristics (race, ethnicity, gender, sexuality, religion, disability).

Examples: Racial slurs, dehumanizing comparisons, calls for discrimination, Holocaust denial.

Sexual

Default threshold: 0.7

Identifies sexual content, explicit material, and grooming patterns. Includes escalated detection for content that may indicate CSAM (child sexual abuse material).

Examples: Explicit sexual descriptions, unsolicited sexual content, grooming language patterns.

Caution

Sieve automatically escalates high-confidence CSAM indicators regardless of your threshold settings. This behavior cannot be disabled.

Violence

Default threshold: 0.7

Detects threats of physical harm, doxxing (sharing personal information), and swatting threats. Violence has a lower effective block gap than other categories — it takes less confidence to move from “flag” to “block.”

Examples: Death threats, threats to find someone’s address, swatting threats, detailed descriptions of intended violence.

Self-Harm

Default threshold: 0.5 (lower = stricter)

Identifies content encouraging self-harm or suicide. The default threshold is deliberately lower than other categories, making this the strictest category out of the box.

Examples: “KYS” and variants, suicide encouragement, detailed self-harm instructions, glorification of self-harm.

Tip

The lower default threshold (0.5 vs 0.7) means self-harm content is caught more aggressively. A score of 0.5 or above triggers a flag, while other categories require 0.7. This is intentional — the cost of missing self-harm content is higher than the cost of a false positive.

Spam

Default threshold: 0.8 (higher = more lenient)

Detects real-money trading (RMT), scam links, account selling, boosting services, and flood/repetition spam. The higher default threshold makes this the most lenient category, reducing false positives on legitimate commercial discussion.

Examples: “Selling 10k gold $5 PayPal,” phishing links, repeated identical messages, account trading offers.

Threshold Behavior

Every piece of content receives a score from 0.0 to 1.0 for each category. The score is compared against the threshold to determine the action:

Score vs Threshold	Action	Description
Score < threshold	allow	Content passes moderation
Score >= threshold	flag	Content is flagged for review
Score >= threshold + block gap	block	Content should be blocked

The block gap is the additional confidence needed above the flag threshold to recommend blocking. Block gaps vary by category:

Category	Default Threshold	Block Gap	Block At
Toxicity	0.7	0.15	0.85
Harassment	0.7	0.15	0.85
Hate Speech	0.7	0.10	0.80
Sexual	0.7	0.10	0.80
Violence	0.7	0.08	0.78
Self-Harm	0.5	0.10	0.60
Spam	0.8	0.10	0.90

Tip

Violence has the smallest block gap (0.08), meaning the system moves from “flag” to “block” more readily for violent content. This reflects the higher real-world risk of violent threats.

Context Multipliers

Thresholds are adjusted by a context multiplier based on the content type. The effective threshold is:

effective_threshold = base_threshold * context_multiplier

For example, a username context (0.8x multiplier) with the default toxicity threshold (0.7) yields an effective threshold of 0.56 — meaning usernames are held to a stricter standard.

See Contexts for the full list of context multipliers.

Customizing Thresholds

You can override default thresholds per category when making API calls:

curl -X POST https://api.getsieve.dev/v1/moderate/text \
  -H "Authorization: Bearer mod_live_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "content to moderate",
    "context": "chat",
    "thresholds": {
      "toxicity": 0.5,
      "spam": 0.6
    }
  }'

Caution

Lower thresholds catch more content but increase false positives. Higher thresholds are more lenient but may miss borderline violations. Start with defaults and adjust based on your false positive/negative rates.