Claude LLM Text Watermark Lab & Statistical Verifier

Kirchenbauer & Anthropic EU AI Act compliant statistical green/red logit bias simulation & z-score verifier

EU AI ACT STATISTICAL AUDIT
Observed Green (G / T)
0.0%
0 / 0 tokens
Empirical z-Score
0.00
z = (|G| - γT) / σ
p-Value (One-Tailed)
1.00
Binomial Null Test
Detection Verdict
Unwatermarked
Confidence: 0.0%
Watermarked Token Stream Visualizer ■ Green-list (Favored)   ■ Red-list (Disadvantaged)   ⬚ Perturbed
Hover or click any token above to inspect logit boost, hash window, and entropy status.
Null vs Observed Green Distribution N(μ=γT, σ²)
Left: Null Hypothesis (γ = 0.5) Marker: Observed Tokens (|G|)
Sequence Length vs Detection Confidence z-score scaling with √T
Threshold z = 3.0 (p < 0.0013) Current Length: 128

Trace the saved watermark counting rules and fixture limits

Read the explanation

The saved Claude-labelled simulator repeats preset words rather than calling a model. Its written bias rule is gamma times exponential delta, divided by gamma times exponential delta plus one minus gamma. At gamma one half, delta zero gives probability one half, while delta two gives about point eight eight zero eight. Both bars show probability at six hundred pixels per unit. These are source simulation probabilities, not measured Claude behaviour. 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. Its exported manifest writes sixty-four randomly selected hexadecimal characters for a mock hash and thirty-two for a mock signature suffix. Both bars count generated characters at eight pixels per character. These strings are not actual SHA-256 or Ed25519 computations. A timestamp, statistical verdict and compliance label are preserved fields of a simulation manifest; none proves legal compliance, cryptographic provenance or actual provider deployment.

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.