Runtime fingerprint extracted from 53 open-weight models

The author presents evidence that running open-weight language models carry an externally readable runtime fingerprint, achieving 99.3% pair separation across 53 evaluated models using only ordinary inference requests without weight access or response text. Under baseline testing, the extracted runtime signatures demonstrated an equal error rate of 0.5% in claim checks and correctly identified models in top-three matches across all tested sessions. The findings suggest a potential path toward establishing a traceable runtime identity for AI governance and monitoring, though the methodology remains withheld for security and commercial reasons. Open article →