AI Watermark & Research Integrity Inspector

Kirchenbauer-Anthropic Engine

Sequence Partition & Token Diagnostics

Interactive Token Partition Map (● Green List | ● Red List): 45 Tokens
Token: Recent
Idx: 0
Hash Seed (h): 0x8f2a1b
Partition: Green List (Accepted)
Logit Shift: +2.00
Perturbation & Paraphrase Stress-Tester 0% Signal Decay

Null Hypothesis Detection Test (z-Score)

H₀: Human text vs H₁: Watermarked
High Confidence Watermarked
Green token frequency exceeds Gaussian null expectation at α = 0.01.
p = 2.85e-05
Observed z-Score 4.0249 Critical z_α: 2.3263
Green Ratio (|G|/T) 80.0% 36 / 45 Tokens
False Positive Risk 0.003% 1 in ~35,000 papers
Null Hypothesis Normal Distribution N(0,1):
Academic Integrity Note: Low-entropy boilerplate text (e.g. standard chemical synthesis protocols or mathematical equations) naturally constrains vocabulary, increasing false positive probabilities under uncalibrated watermarking thresholds.

Kirchenbauer Cryptographic Watermark & Statistical z-Test

Read the explanation

During AI token generation, the Kirchenbauer scheme hashes preceding token context to divide the vocabulary into green and red partitions, boosting green logit probabilities. Natural text samples green and red tokens roughly equally. But watermarked generations accumulate an anomalous concentration of green tokens across the document stream. The inspector tests this discrepancy against the standard normal null distribution. When observed z exceeds critical z, the text is flagged with negligible false positive risk. You can adjust perturbation stress or switch presets to test how paraphrase edits decay the watermarking signal back toward the unflagged baseline.

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.