Article 50 Verification

AI Watermark Robustness & EU Compliance Lab

Statistical Kirchenbauer Red/Green Token Biasing & Perturbation Stress Bench
1. Statistical Watermark Embedder Media: Synthetic Text
2. Perturbation Attack Sandbox Evasion Stress-Testing
3. Detection Metrics & Significance Kirchenbauer et al. Test
Green Token Ratio
0.842
48 / 57 Tokens
Detection Z-Score
5.82
p < 10⁻⁸
Watermark Confidence
99.98%
Evidentiary standard
Evasion Threshold
38.5%
Max attack tolerance
🛡️ EU AI Act Art. 50 Compliant
Statistical watermarking yields z ≥ 4.0 with authentic token entropy, satisfying technical requirements for machine-readable provenance and transparency under Regulation (EU) 2024/1689.
MANIFEST: c2pa.v2 // HASH: e9b4c071... // SEED: 0x7F4A // TOKENS: 57 // Z_STAT: 5.82

Kirchenbauer Watermark Robustness and Article 50 Compliance

Read the explanation

During autoregressive text generation, the embedder partitions vocabulary into green and red lists using a pseudorandom hash of the previous token and secret key. Adding logit boost delta elevates green token selection probability. To verify provenance, detection computes a standard normal score from observed green tokens. An unwatermarked text sits near mean zero, while biased generation shifts the distribution outward to an observed score above five. Increasing adversarial distortion rates substitutes synonym tokens and destroys alignment with previous seed keys. As mutated tokens flip from green to red, the score collapses below four, triggering an Article 50 non-compliance alert.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.