From content to verdict, with evidence attached.
A piece of content enters the pipeline. Nine detectors read it in parallel. An ensemble layer composes their signal. The verdict comes back with confidence, uncertainty, and per-detector evidence the consumer can inspect.
The flow
How signals combine
Each detector emits a calibrated score and an evidence trace. The ensemble layer is a learned composition function, not a hand-tuned weight matrix. It is trained on the same kind of adversarial content the detectors face individually, so it learns the failure modes of each detector in combination.
The ensemble layer can disagree with a unanimous detector signal. If eight detectors flag and the cross-modal detector reports a known benign pattern that produces those eight signals, the ensemble can clear. The reverse holds: a confident cross-modal signal can produce a hold even when most channels read clean.
Confidence and uncertainty
Every verdict carries a confidence band and an explicit uncertainty marker. A consumer can decide: pass-through low-confidence content with a label, hold ambiguous content for review, or fall back to a stricter default. The verdict shape is rich enough to drive any of these.
Evidence attached to a verdict
Every verdict carries the per-detector signal that contributed, the cross-modal disagreements, and a content-shape fingerprint. This is not an explanation produced after the fact; it is the actual signal the ensemble used. A downstream auditor inspecting a held piece of content sees the same evidence the engine used.
Evidence is intentionally rich enough to be useful and intentionally bounded so it cannot be reverse-engineered into detector weights. AEGIS publishes its methodology; it does not publish its exact weights.