EU AI Act Provenance
Kirchenbauer-2023 Model
LLM Text Watermarking & Detection Laboratory
Cryptographic token logit biasing, statistical z-score detection & evasion attack testbench
🇪🇺 Strict Provenance
🎨 Creative Narrative
💻 Factual / Math
⚔️ Synonym Evasion
↺ Reset
Watermark Parameters
z = (|G| - γN)/σ
Green-list Ratio (γ)
0.50
Logit Bias (δ)
2.00
Hash Context Window (h)
1
Secret Seed Key
42
Evasion / Perturbation
Attack Mechanism
Synonym Substitution
Token Deletion
Paraphrasing / Jitter
Attack Rate / Swap Fraction
20%
⚡ Inject Perturbation Attack
Text Corpus Input
Total Scored Tokens (N)
38
Entropy: Normal
Green Tokens |G| / |R|
32 / 6
84.2% Green
Detection z-Score
4.22
z ≥ 4.0 (EU High Assur.)
p-Value (FPR)
1.2e-05
p < 0.001 Provenance
Attacked z-Score
2.84
Watermark Retained
Token-by-Token Partition Inspector
● Green List (+δ logit)
● Red List
Hover or click a token above to inspect cryptographic hash, previous-token context, and green-list probability.
Statistical Detection Distribution & Confidence Boundary
H0: Natural Text ~ N(0,1) vs H1: Watermarked
Cryptographic Verification Report
💾 Export JSON Audit
📄 Export Markdown
Enjoy this tool? Build your own with Super