EU AI ACT • PROVENANCE

LLM Text Watermarking & Provenance Simulator

Engine Ready • Kirchenbauer Soft Watermark
Watermark Parameters
2.0
Higher δ ensures undetectable evasion but slightly reduces generation perplexity.
0.50
42
1-gram
Perturbation Attacks
Token Sequence Inspector
Green List (G) Red List (R)
Click any token in the stream above to inspect its hash partition, logit delta, and prefix seed.
Hypothesis Testing & z-score Distribution H₀: Human Text (p=γ) vs H₁: Watermarked
Calculating detection hypothesis... p < 0.0001
Detection Telemetry
Total Tokens (T)
0
Green Tokens (|G|)
0
z-score (z)
0.00
Observed Green Rate: 0.0%
Theoretical H₀ Expected (γ): 50.0%
Significance Threshold (α=0.01): z > 2.33
EU AI Act Article 50 Provenance
• Watermark Mode: Statistical Soft Logit (Kirchenbauer)
• Provenance Cryptoseed: HMAC-SHA256 (Simulated)
• Compliance Verdict: Compliant AI Transparency
• Robustness Rating: High (z > 4.0)

Trace the saved watermark counting rules and fixture limits

Read the explanation

The saved simulator subtracts the context window from the non-space token population, then floors the evaluated count at one. Twenty tokens with window one yield nineteen; with window three yield seventeen. Both bars use thirty pixels per evaluated token. The green hash uses preceding context and the key, not the current token identity. Spaces remain in the display but are not counted. This is an authored hashing demonstration, not validated provider watermarking. The written detector compares green count to token count times the assigned green fraction, then divides by square root of count times fraction times one minus fraction. With an illustrative one hundred evaluated tokens and fraction one quarter, expected count is twenty-five and standard deviation about four point three three. Thirty-five green tokens give z about two point three zero nine; forty-five give four point six one nine. Both bars depict z on one hundred thirty pixels per unit. These illustrative counts explain the algebra, not a measured false-positive rate or proof of authorship. The re-encode handler calculates a biased probability and samples a random branch, but both branches push the exact same original token. Its written control therefore leaves text unchanged. As an illustrative thirty-token sequence, before and after both have thirty tokens, shown at eighteen pixels per token. This is a source limitation, not an embedding guarantee. The original statistical export and claimed authentication labels do not establish provenance or compliance.

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.