Beyond AI Watermarks: A Transformation-Certificate Architecture for Mixed Human–AI Authorship

As AI-assisted writing, editing, and creative production become commonplace, existing approaches to AI content transparency — statistical watermarks and binary provenance signals — do not, by themselves, represent the transformation boundary in mixed human-AI work. The EU AI Act's Article 50, which took effect August 2, 2026, recognizes a legal distinction between AI-generated content and standard editing assistance that does not substantially alter human-authored input. Anthropic's published marking documentation explicitly acknowledges that a detected mark does not establish original authorship, contribution boundaries, or complete provenance. This paper proposes a complementary architecture: an externally anchored transformation certificate that preserves declared pre-interaction and post-interaction states, transformation metadata, and independently timestamped records. Provider marks can authenticate AI-system involvement. Externally anchored transformation certificates can preserve declared before-and-after states. Verifiable mixed authorship ultimately requires both. A working implementation is deployed at hash.haawke.com and verify.haawke.com.