161 LLM fingerprints, broken down model by model

Atlas summary

Models

161

Namespaces

30

Reference answers

18,996

Per-model samples

77–120

The closest pair in this eight-cell pack is qwen/qwen3-max and qwen/qwen3-max-thinking at mean JSD 0.022070. The most isolated reference is meta-llama/llama-3-8b-instruct, whose nearest neighbor sits at 0.862797.

This atlas analyzes the product reference, not live user scans. It contains aggregate answer counts from the author dataset and no keys, conversations, cookies, Agent logs, or user reports.

How we calculated every page

1. Read the same eight-cell JSON file bundled with the current Agent Skill.
2. For each cell, count normalized answers, divide by the actual valid-sample denominator, calculate Shannon entropy, and identify the mode.
3. Compare each model with every other enrolled model using base-2 Jensen–Shannon divergence over the joint answer support.
4. Average the eight cell distances, rank the nearest three references, record the strongest cell difference, and hash the page evidence.

The generator reads reference pamela-openrouter-2026-07-8cell-ref-a directly. Reference SHA-256: 51f3f63cbcfb56b151055d2d10e36a9cc2f0644cc68b580784853d68951fa866. The machine-readable atlas JSON exposes the same computed values.

Browse all 161 model fingerprints

How to read a model page

• Counts and shares reveal which answers dominate each prompt. Every share uses its cell’s actual sample count.
• Entropy measures spread. A low value marks a concentrated answer habit; a high value marks a broader set.
• Mean JSD compares all eight distributions. Zero means identical frequencies; larger values mean stronger separation.
• Nearest reference provides context. It never converts an unclaimed or unenrolled route into a verified identity.

Limits and update policy

The author collected the upstream cohort through OpenRouter at temperature 1 with reasoning disabled. Provider changes, wrappers, model updates, and decoding can shift a fingerprint. VerifyLLMAPI selected eight high-gap cells and calibrated them on the same published data, so the product EER values are not an independent accuracy claim.

A new model needs fresh samples under the same protocol and recalibration. A family name or marketing alias adds no reference evidence. The support policy and exclusion guide explains the exact verdict boundary.

Primary sources and license

1. Tomas Bruckner, One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions.
2. Published dataset and results, DOI 10.5281/zenodo.21278557, CC BY 4.0.
3. Published collection and analysis software, DOI 10.5281/zenodo.21278793, MIT.
4. Jianhua Lin, Divergence Measures Based on the Shannon Entropy, IEEE Transactions on Information Theory, 1991.