Engineering the trust layer

In 2025 we shipped 47 model updates, deprecated three first-generation classifiers, and rewrote our chain-of-custody service from scratch. This is the public version of our year-end engineering retro.

What we shipped

• A unified ingestion pipeline that handles arbitrary EXIF metadata without dropping signatures
• Hardware-attested capture for compatible mobile devices (iOS 17+ and Android 14+)
• A two-stage classifier that separates generative-model artefacts from compression artefacts

What we removed

Our original GAN-only detector. It was 94% accurate against StyleGAN3 but degraded sharply against Stable Diffusion 3 and Flux. Rather than patch it, we replaced the whole pipeline with a contrastive embedding model. The migration took four months — we ran both classifiers in shadow mode the entire time.

Operating in regulated industries

The single hardest lesson: audit trails are a product surface, not a compliance afterthought. Customers in banking and government need to demonstrate what happened, not just have it recorded. We rebuilt the AuditLog API around exportable, signed, timestamped chains — and it became one of our top three reasons-to-buy.