Description du poste
A runtime signature, captured without weights or response text, achieves 99.3% pair separation across 53 models.
TL;DRA running language model carries an externally readable fingerprint. Using only ordinary inference requests, no access to weights and without reading response text, I extracted the individual fingerprints for 53 open-weight models (0.5B to 9B) from 13 families as runtime signatures with a 99.3% model pair separation (1,368 of 1,378).
Under baseline identification and verification testing, a Claim Check against all other registered models had an AUC of 0.9997 and equal error rate of 0.5%, while held-out sessions matched their own model's signature 94.4% of the time (202 of 214), with the correct model in the top three presented every time.
Separation included models of the same family and the same size. It also showed the signature is unique to the model as it is served, with different quantisations of the same model and same model variants distinct in the pairs tested. Models sharing architecture, size and quantisation and differing only in weights still separated in 7 of 9 pairs. Every fine-tune and post-training variant separated, and the two pairs that did not were successor releases.
This findings note is evidence the signal exists, the fingerprint can be surfaced and the signature extracted. All of this was achieved on consumer grade infrastructure, running one stack and is supported by empirical data as a first step towards establishing a traceable runtime identity.
Terminology & TransparencyFingerprint - The underlying property of a model surfaced at runtime by the Instrument.
Signature - The extracted representation of a Fingerprint, and Identity artefact for registration.
Session - A single evaluation of a model, presented as evidence against a registered Signature.
AI Transparency. AI was used to assist with the generation of results and figures only. All other content has been created by the Author.
PurposeThis is a findings note, not a me