🔬 Model Fingerprint Verification
Point this tool at any OpenAI-compatible chat endpoint, and it runs a battery of low-cost behavioural questions, compares the answer patterns against a continuously updated reference library, and tells you whether the endpoint really behaves like the model it claims to serve — catching model swapping / dilution before you spend real money on it. A second probe checks tokenizer lineage (glitch-token repeats) for an independent hint that works even for closed models.
Your API key is used for the probes only and never written to disk or database. Public mode shows the verdict and its plain-language explanation; methodology parameters (thresholds, divergence values, per-cell data) are available in the full version.
Run a verification
Progress
Queued…
Result
Recent public detections
Anonymised history of detections run through this public tool: the endpoint host, the model id it CLAIMED to serve, and the verdict. Keys are never stored, method parameters (thresholds, divergence values, per-item detail) never leave the full version — one line per run, nothing else.
FAQ
How can I tell if a relay is really serving the model it claims?
Run the verification above with the relay's base URL, your key and the claimed model id. It sends 360 single-token behavioural probes (colour / number / coin-flip questions across languages), compares the answer distribution against the reference library, and returns match / uncertain / mismatch. A full run costs a few cents with cheap models.
What does "mismatch" actually mean?
The answer pattern is too far from the claimed model's reference (norm ≥ 3 against its noise floor) while matching no other reference either — the backend is likely a different or unknown model. If the top match is a different known model, the report names it: that is direct evidence of model swapping (selling an expensive model while serving a cheaper one).
Can a relay switch models in the middle of my session?
Point-in-time verification can miss mid-session swaps if probes only sample the honest period. We are building Relay Guard, a transparent local proxy that interleaves probes with your real traffic and reports a per-segment verdict timeline (consistent / switched). Internal testing now; the method page will be published once the real-world baseline completes.
Does this work for closed models like Claude or GPT?
Yes. The library includes closed-source references (Claude, GPT, Gemini, Grok) collected via official infrastructure routes, plus 18+ domestic open-weight backends — 40 references / 36 backends at the moment. Coverage follows what relays actually sell most.
What are glitch probes for?
Glitch-token repeats probe tokenizer lineage: every tokenizer family chokes on specific nonsense tokens in a characteristic way. It is an independent second signal that works even when behavioural comparison is uncertain — useful for catching "same family, different version" substitutions.