05 · Cross-modal detector

Cross-modal (research preview)

A research-preview surface that targets content where each individual channel reads benign but the combination is not. Evasion increasingly hides in the seams between modalities; cross-modal correlation analysis is on the v1.1 roadmap, building on AEGIS's 6-axis dual-judge rubric.

Channel
Joint reasoning over visual, audio, text, metadata (research)
Status
Research / roadmap (v1.1) — not in production verdict path today
Status. Cross-modal correlation is currently a research surface, not an active detector in the analyze pipeline. AEGIS v1.0 ships the 6-axis dual-judge rubric (Gemini 3 Pro + Claude Haiku 4.5); cross-modal joins as a v1.1 roadmap target.

What it would process

The future detector targets content where each individual channel reads benign but the combination is not. Evasion increasingly hides in the seams between modalities; this surface is being built toward joint reasoning across channels.

Adversarial patterns it is tuned to catch

  • Audio-visual disagreement (clean image, hostile narration)
  • Caption-to-image inversion (caption flips the image's meaning)
  • Metadata-content disagreement (envelope says one thing, body another)
  • Layered evasion that requires understanding two modalities at once

Contribution to the ensemble verdict

Outputs both a cross-modal disagreement score and an evidence trace identifying which channels disagreed.

Per-detector outputs are not a final verdict. In v1.0 the analyze pipeline produces verdicts from the 6-axis dual-judge rubric; cross-modal scores will join the rubric in v1.1 once the research preview promotes to production.