Kirchenbauer Statistical Cryptographic Watermarking & EU AI Act Art. 50 Resilience Auditor
This lab classifies text using a context hash and a green threshold. It illustrates a detector, rather than proving that a real language model generated the text. Increasing gamma moves the threshold, so more hash scores count as green. For a synthetic example, take one hundred evaluated tokens and gamma equal to one half. The expected green count is fifty. The standard deviation is five. Moving the observed count from fifty to sixty raises the z score from zero to two. Use the token inspector to see the preceding context and normalized hash. Change the seed or gamma, then compare the classification. The displayed confidence is derived from a normal approximation, and should not be treated as calibrated proof of authorship.