Supported category signals
The supported catalog contains adult_nsfw, weapon, drug_like_content, graphic_violence, sensitive_document, and ai_generated. A submission may request a nonempty unique subset using checks; omitting checks requests the default five; ai_generated must be explicitly selected. GET /api/v1/capabilities describes the catalog and configured routing availability, not live health.
Read the job’s canonical checks list and its result coverage. scores, categories, and checked_categories contain only requested checks. not_checked_categories lists supported catalog entries omitted from the request. unsupported_categories lists the fixed limitations outside the supported catalog.
Scores and thresholds are model signals used by the configured policy. They are not a factual determination, a guarantee of harm, or universally comparable probabilities.
A sensitive_document signal does not verify document authenticity or decide whether processing that document is lawful.
Interpret decisions conservatively
Inspect policy_version, thresholds, checked_categories, not_checked_categories, and unsupported_categories alongside the decision. Results may change as model behavior or policy configuration changes.
- block: the configured policy found a blocking signal. Apply your product’s moderation policy.
- review: human review is needed before making an approval decision.
- no_signal: no configured signal crossed the applicable decision threshold in the checks requested for this job. It does not mean universally safe or say anything about checks you omitted.
Not covered
Do not use this system to infer that an image satisfies policies it does not check. Unsupported areas include:
- Age or whether a person is a minor.
- Self-harm, hate, harassment, or threats.
- Text understanding or comprehensive OCR-based moderation.
- Legality, consent, provenance, or ownership.
- Document authenticity.
- Video, audio, animated images, and general content outside the supported static-image formats.
Keep people in the workflow
Images can be ambiguous, misleading, culturally dependent, or outside the models’ training distribution. Both false positives and false negatives are possible. Establish an appropriate human review and appeal process for your application.
Evaluate your deployed models on representative content before relying on them. Software test results do not validate model accuracy. A no_signal decision or a successful health check is not an assurance of content safety.