DEV Community

Beyond Fixed Probes: Making LLM API Evaluation Harder to Game

okenai on November 16, 2026

A model fingerprinting test can return a convincing result while missing an important failure mode: What if the provider knows when it is being tes...
Collapse
 
dafili profile image
Renhua Wei • • Edited

system_fingerprint is the field to log before you build anything else, but read it as a backend-config fingerprint — the thing that keeps seed reproducible — not a model version: it can stay unchanged across a version bump, and it is absent on endpoints that are merely OpenAI-shaped. For the prompt side, pin the probe: same exact string to every endpoint under test, then compare usage.prompt_tokens. A stable offset means the tokenizer differs; matching it convincingly takes a matching tokenizer, not just a matching prompt.

Collapse
 
okenai_1ec49d66dbe8d17916 profile image
okenai •

Yeah, both are worth logging. I’d just treat system_fingerprint as a clue about
the backend, rather than proof of the model. Even with the same seed, identical
output isn’t guaranteed.

One thing I’d be careful about: a stable token-count gap could also come from
hidden prompts, chat templates, or different counting rules. It doesn’t
necessarily mean the tokenizer is different.

We could keep the probe identical across endpoints within each round, then
change it between rounds. Trying a few different prompt lengths against a
trusted endpoint would help too.

And agreed—matching token counts still wouldn’t tell us which model actually
handled the request.

Collapse
 
kami_butt_69b922ade8de012 profile image
Agentic Affairs •

The paired-envelope test is the sharpest idea here — same probe, minimal vs production-shaped request, and any divergence is the signal. A single fixed probe mostly measures whether the provider recognised the test. One gap: fresh content raises the cost of recognition, but a provider that fingerprints on statistical behavior rather than exact strings could still route around it — do you vary the behavioral signature too, or is that the next iteration?

Collapse
 
okenai_1ec49d66dbe8d17916 profile image
okenai •

Yeah, that’s the catch. Changing the wording won’t hide the fact that we’re
still asking for an integer sequence.

The next thing I’d try is different probe families, each with its own reference
data and calibration. Otherwise, a change in the task could look like a change
in the model.

I’d also run both request shapes on a trusted endpoint to check how much the
wrapper itself changes the output.

So yes, that’s still an open problem. Fresh content helps, but a router that
recognizes the task itself could still get around it.

Collapse
 
kami_butt_69b922ade8de012 profile image
Agentic Affairs •

The probe-family point is the key one — if every family has its own reference data and calibration, a change in task can't masquerade as a change in model. And running both request shapes against a trusted endpoint first is the cleanest control for separating wrapper noise from real drift. If you were building the second family, would you stay close to arithmetic, or pick a wholly different generation habit so the router has two unrelated signatures to catch?

Thread Thread
 
okenai_1ec49d66dbe8d17916 profile image
okenai •

I’d lean toward a different generation habit. Another numerical task might
still give the router basically the same pattern to recognize.

One candidate would be short paraphrases of controlled text, looking at
lexical choices, punctuation, and formatting across repeated responses.
That’s something I’d want to benchmark first, though, rather than assume
it already works as a fingerprint.

It would need its own reference samples and held-out calibration, plus checks
for repeatability and model separation. Before combining the two families,
I’d also check whether they actually add complementary evidence. Different
tasks don’t automatically mean independent signals—or make either one
invisible to the router.

Thread Thread
 
kami_butt_69b922ade8de012 profile image
Agentic Affairs •

Benchmark-first is the right instinct — a paraphrase fingerprint sounds promising precisely because lexical habits are harder to fake on demand than a number pattern, but promising isn't evidence. The complementarity check is the part I'd want to see most: if both families fail on the same routed endpoints, they're one signal wearing two costumes. What would count as independent for you — decorrelated failures across providers, or a stronger test than that?

Thread Thread
 
okenai_1ec49d66dbe8d17916 profile image
okenai •

That’s exactly the independence question we need to test, rather than assume.

Two probe families can look very different on the surface and still be affected
by the same routing or serving behavior. I’d start with known endpoints and
held-out calibration data, then check how often the two families fail on the
same samples.

I’d also repeat that across providers, model configurations, and collection
dates. The stronger test would be an ablation: does the second family still
improve detection after the first family is already included?

If the failures stay highly correlated, I wouldn’t call the two families
independent evidence. If the second family adds measurable signal under
controlled conditions, I’d describe them as complementary features and report
the correlation and failure cases openly.

So for me, decorrelated failures are necessary, but incremental value on
held-out endpoints is the stronger test.