How to test whether a relay API is watered down

Some relay APIs claim to serve one model while actually running a cheaper one. Glitch tokens come from real tokenizer vocabularies, and different model families react to the same glitch token in dramatically different ways — which makes them useful as fingerprints.

Steps

1. 01

   Filter by model

   On the catalog page, pick the model you suspect and prefer tokens tagged “Fingerprint” — they were collected for exactly this kind of test.

2. 02

   Copy the raw token

   Copy the raw token from a card or the detail page. ␣ marks a real space and copying preserves it — the space is part of the fingerprint, so don't trim it.

3. 03

   Send it to the relay

   Paste the token verbatim into the API under test: ask it to repeat the string, or simply send the token as a message.

4. 04

   Read the output

   Refusals, gibberish, or failed repetition strongly suggest the service really runs that model. A normal response doesn't rule it out — the behavior may have been patched, or deployments may differ.

Caveats

• A fingerprint test is one piece of evidence, not a verdict on its own.
• Providers keep patching known glitch tokens, so older cases may stop reproducing on newer versions.
• Some tokens trip safety filters or system prompts — judge by whether the reaction is abnormal, not by the exact wording.