Which LLM models can VerifyLLMAPI fingerprint?

Coverage at a glance

Enrolled models

161

Model namespaces

30

Reference cells

8

Calibrated profiles

6

“Supported” has a strict meaning: the library contains a reference distribution for that model ID, and the selected profile has a calibrated threshold. Support never transfers from a family name to a new checkpoint. An enrolled openai/gpt-5.5 reference does not cover an unlisted GPT release.

Search all 161 supported model IDs

Download the machine-readable coverage JSON.

Which models are not supported?

Every model absent from the directory above lacks a calibrated v0.3.2 identity verdict. The scanner may still collect its current fingerprint and show the nearest enrolled references, but it returns INCONCLUSIVE. A nearest neighbor describes similarity; it does not prove identity.

Common unsupported groups include new releases published after the reference snapshot, mutable rolling aliases, mandatory hidden-reasoning models, meta-routers, search or multi-agent pipelines, specialized code or media models, and endpoints without enough valid samples. The paper’s model-selection rules explain these exclusions.

Four paper models omitted from the packaged eight-cell reference

Our build requires at least five valid samples in both split halves for every selected cell. These four models appear in the paper’s broader 165-model analysis but fail that product-specific completeness rule.

Where is the fingerprint library?

The installed Agent Skill stores the full reference at references/pamela-openrouter-8cell-v1.json, relative to SKILL.md. The same file sits inside the latest VerifyLLMAPI package; its bundled VERSION is 0.3.2.

Reference ID

pamela-openrouter-2026-07-8cell-ref-a

JSON size

138,504 bytes

SHA-256

51f3f63cbcfb56b151055d2d10e36a9cc2f0644cc68b580784853d68951fa866

Upstream data

Zenodo 10.5281/zenodo.21278557

The file records model IDs, answer counts for eight selected task-language cells, calibration thresholds, sample budgets, the source-data hash, and provenance. It stores no user credentials, conversations, or scan results.

How many model requests does each scan use?

The default complete check makes 20 fresh model requests. These calls can consume your Codex or Claude Code allowance, or a configured provider balance, and may take several minutes.

These v0.3.2 figures come from same-data, post-hoc high-gap cell selection. They measure calibration on the published OpenRouter cohort, not independent production accuracy. Provider, wrapper, model updates, and decoding settings can move a fingerprint.

Questions users ask about model fingerprints

Does a matching model name prove support?

No. The scanner first resolves an exact full slug, then a matching basename. Use the exact listed ID to avoid ambiguity. The child route must also stay stable and every cell must produce enough valid samples.

Can VerifyLLMAPI identify an unknown model?

It can rank enrolled references by JSD distance. It reserves identity verdicts for a claimed model with an enrolled reference. Family classification in the paper reached 59.5%, so nearest-neighbor discovery carries more uncertainty than claim verification.

Does one reference work across every provider?

Provider changes matter. The paper reports a median cross-provider distance of 0.227 and cross-provider verification AUC of 0.880, below the same-mix AUC of 0.971. The report names the OpenRouter reference transport so users can judge this gap.

How does a new model gain support?

Enroll the exact checkpoint through the same prompt, temperature, reasoning, normalization, and provider protocol; collect enough valid repetitions for each selected cell; then recalibrate thresholds on held-out or new data. Adding a name without samples creates no fingerprint.

Does the scan upload my fingerprint?

No. The v0.3.2 current-Agent scanner aggregates answers locally and deletes its temporary sample directory. VerifyLLMAPI receives no key, answer, model route, or report.

Primary sources and reproducibility

1. Tomas Bruckner, One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions, arXiv:2607.10252v1, 2026.
2. Published fingerprint dataset and derived results, Zenodo record 21278557, DOI 10.5281/zenodo.21278557, CC BY 4.0.
3. Published collection and analysis software, Zenodo record 21278793, DOI 10.5281/zenodo.21278793, MIT.
4. Irena Gao, Percy Liang, and Carlos Guestrin, Model Equality Testing: Which Model is this API Serving?, ICLR 2025.