Each detector specializes. The ensemble decides.

Classifies image and video content frame by frame. Looks for unsafe imagery, encoded payloads carried in imagery, and visual content engineered to read benign to single-shot models.

Listens to spoken language, sound effects, and waveform features in audio and video tracks. Catches what visual classifiers miss when an audio track tells a different story than what is on screen.

Reads text wherever it appears: in chat, in captions, in subtitles, in OCR pulled out of images, and in the document body. Tuned for content engineered to read clean to a fast filter and dirty to the human reader.

Examines metadata that travels alongside content: HTTP headers, EXIF data, container fields, embedded URIs. Often carries signal that the content body alone does not.

The detector that catches content where each individual channel reads benign but the combination is not. Evasion increasingly hides in the seams between modalities; this detector is purpose-built to find it.

Operates on the semantic content of what was said, shown, or written. Tuned to recognize themes, narrative shapes, and harmful framings even when the surface text is sanitized.

Reads the context. The same sentence in a medical-encyclopedia caption is not the same sentence in a chat with a child. Context-aware classification prevents both over-blocking and under-blocking.

Listens for the shape of language, not just its content. Tuned to catch evasion that hides behind dialect, slang, and code-switching that English-only detectors miss.

Looks at behavior, not just content. A single message may be ambiguous; a sequence of messages from a particular account, at a particular cadence, on a particular platform, often is not.