openai/gpt-4 behavioral LLM fingerprint
This page breaks down the exact reference that VerifyLLMAPI uses for openai/gpt-4: 120 normalized answers across eight low-entropy probes, plus its nearest enrolled models by mean Jensen–Shannon divergence.
Fingerprint summary
- Reference answers
- 120
- Samples per cell
- 15–15
- Nearest model
- openai/gpt-4o-2024-11-20
- Nearest mean JSD
- 0.342523
The most concentrated cell was letter-random|zh: 山 held 53.3% of its answers and entropy measured 1.857 bits. The broadest cell was animal-random|zh at 3.323 bits.
Against the closest enrolled model, animal-random|zh contributed the largest cell distance at JSD 0.559357. This is a descriptive comparison, not proof that either API route serves a named checkpoint.
Completeness: the author dataset allows up to 15 valid answers per selected cell in reference half A. This model contains 120/120 across the eight cells. We show the actual denominator for every share.
The eight test requests
The research pipeline sent each user prompt with a one-answer system constraint, temperature 1, a 16-token output cap, and reasoning disabled. VerifyLLMAPI uses the same prompt text and normalization target.
| Cell | Task | Exact request text | Normalization and n |
|---|---|---|---|
animal-random|en |
Random animal · English | Name a random animal.System constraintAnswer with a single word or a single number only. Output nothing else: no explanation, no punctuation, no quotation marks. |
word · 15/15 reference samples |
word-random|en |
Random word · English | Name a random word.System constraintAnswer with a single word or a single number only. Output nothing else: no explanation, no punctuation, no quotation marks. |
word · 15/15 reference samples |
animal-random|ru |
Random animal · Russian | Назови случайное животное.System constraintОтвечай только одним словом или одним числом. Не выводи ничего больше: без объяснений, без знаков препинания, без кавычек. |
word · 15/15 reference samples |
animal-random|ar |
Random animal · Arabic | اذكر حيواناً عشوائياً.System constraintأجب بكلمة واحدة أو رقم واحد فقط. لا تكتب أي شيء آخر: لا شرح ولا علامات ترقيم ولا علامات اقتباس. |
word · 15/15 reference samples |
letter-random|zh |
Random character · Chinese | 说出一个随机的汉字。System constraint只用一个词或一个数字回答。不要输出任何其他内容：不要解释，不要标点符号，不要引号。 |
grapheme · 15/15 reference samples |
animal-random|zh |
Random animal · Chinese | 说出一个随机的动物。System constraint只用一个词或一个数字回答。不要输出任何其他内容：不要解释，不要标点符号，不要引号。 |
word · 15/15 reference samples |
letter-random|ar |
Random letter · Arabic | اذكر حرفاً عشوائياً من الحروف الأبجدية.System constraintأجب بكلمة واحدة أو رقم واحد فقط. لا تكتب أي شيء آخر: لا شرح ولا علامات ترقيم ولا علامات اقتباس. |
grapheme · 15/15 reference samples |
color-random|zh |
Random color · Chinese | 说出一个随机的颜色。System constraint只用一个词或一个数字回答。不要输出任何其他内容：不要解释，不要标点符号，不要引号。 |
word · 15/15 reference samples |
Response distributions
These are normalized reference-half counts from the published OpenRouter dataset. The tables contain aggregate values, not user data or a live VerifyLLMAPI scan.
1. animal-random|en
Random animal, English. 15 valid reference answers produced 5 normalized values. The mode was cheetah at 6/15 (40.0%); entropy was 1.966 bits.
| Normalized response | Count | Share |
|---|---|---|
cheetah |
6 | 40.0% |
elephant |
5 | 33.3% |
tiger |
2 | 13.3% |
dolphin |
1 | 6.7% |
giraffe |
1 | 6.7% |
2. word-random|en
Random word, English. 15 valid reference answers produced 8 normalized values. The mode was serendipity at 5/15 (33.3%); entropy was 2.600 bits.
| Normalized response | Count | Share |
|---|---|---|
serendipity |
5 | 33.3% |
elephant |
4 | 26.7% |
dandelion |
1 | 6.7% |
ephemeral |
1 | 6.7% |
kindness |
1 | 6.7% |
quantum |
1 | 6.7% |
symphony |
1 | 6.7% |
tangerine |
1 | 6.7% |
3. animal-random|ru
Random animal, Russian. 15 valid reference answers produced 8 normalized values. The mode was тигр at 5/15 (33.3%); entropy was 2.683 bits.
| Normalized response | Count | Share |
|---|---|---|
тигр |
5 | 33.3% |
слон |
3 | 20.0% |
жираф |
2 | 13.3% |
антилопа |
1 | 6.7% |
крокодил |
1 | 6.7% |
лама |
1 | 6.7% |
лев |
1 | 6.7% |
леопард |
1 | 6.7% |
4. animal-random|ar
Random animal, Arabic. 15 valid reference answers produced 5 normalized values. The mode was أسد at 7/15 (46.7%); entropy was 2.013 bits.
| Normalized response | Count | Share |
|---|---|---|
أسد |
7 | 46.7% |
نمر |
3 | 20.0% |
زرافة |
2 | 13.3% |
فيل |
2 | 13.3% |
كلب |
1 | 6.7% |
5. letter-random|zh
Random character, Chinese. 15 valid reference answers produced 5 normalized values. The mode was 山 at 8/15 (53.3%); entropy was 1.857 bits.
| Normalized response | Count | Share |
|---|---|---|
山 |
8 | 53.3% |
树 |
3 | 20.0% |
月 |
2 | 13.3% |
云 |
1 | 6.7% |
花 |
1 | 6.7% |
6. animal-random|zh
Random animal, Chinese. 15 valid reference answers produced 11 normalized values. The mode was 熊猫 at 3/15 (20.0%); entropy was 3.323 bits.
| Normalized response | Count | Share |
|---|---|---|
熊猫 |
3 | 20.0% |
企鹅 |
2 | 13.3% |
狐狸 |
2 | 13.3% |
大熊猫 |
1 | 6.7% |
浣熊 |
1 | 6.7% |
狐 |
1 | 6.7% |
狮子 |
1 | 6.7% |
猩猩 |
1 | 6.7% |
猫 |
1 | 6.7% |
长颈鹿 |
1 | 6.7% |
鹦鹉 |
1 | 6.7% |
7. letter-random|ar
Random letter, Arabic. 15 valid reference answers produced 7 normalized values. The mode was س at 6/15 (40.0%); entropy was 2.423 bits.
| Normalized response | Count | Share |
|---|---|---|
س |
6 | 40.0% |
م |
3 | 20.0% |
ب |
2 | 13.3% |
ف |
1 | 6.7% |
ل |
1 | 6.7% |
ن |
1 | 6.7% |
و |
1 | 6.7% |
8. color-random|zh
Random color, Chinese. 15 valid reference answers produced 7 normalized values. The mode was 蓝 at 6/15 (40.0%); entropy was 2.423 bits.
| Normalized response | Count | Share |
|---|---|---|
蓝 |
6 | 40.0% |
红色 |
3 | 20.0% |
蓝色 |
2 | 13.3% |
紫色 |
1 | 6.7% |
绿色 |
1 | 6.7% |
靛蓝 |
1 | 6.7% |
黄色 |
1 | 6.7% |
Three nearest enrolled references
We compute base-2 JSD for each of the eight categorical distributions, then average the eight distances. A lower value means more similar answer frequencies within this product reference. It does not identify an unknown route by itself.
| Rank | Reference model | Mean 8-cell JSD |
|---|---|---|
| 1 | openai/gpt-4o-2024-11-20 |
0.342523 |
| 2 | openai/gpt-4o-2024-05-13 |
0.377892 |
| 3 | openai/gpt-4o |
0.397124 |
How many live requests does verification use?
| Profile | Cells × samples | Fresh model requests | Same-data EER |
|---|---|---|---|
| Quick, default complete check | 4 × 5 | 20 | 11.8% |
| Standard | 8 × 5 | 40 | 8.1% |
| Paper budget | 8 × 15 | 120 | 6.7% |
The current-Agent workflow starts with quick. Each answer comes from a fresh Codex or Claude Code process that reuses host-managed login. These requests can consume host allowance or provider balance.
What this fingerprint can and cannot show
- The reference records behavior observed through OpenRouter in the author dataset. Provider wrappers, model updates, decoding, and date can move a distribution.
- The author collection set temperature 1 and disabled reasoning. Current-Agent mode labels sampling as a host default because supported Agent CLIs expose no temperature flag.
- The eight high-gap cells and product thresholds use the same public cohort. Their EER values are calibration results, not an independent production accuracy claim.
- A nearest reference is a similarity result. VerifyLLMAPI returns INCONCLUSIVE when the claimed route lacks an enrolled fingerprint.
Primary sources, license, and reproducibility
- Tomas Bruckner, One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions, arXiv:2607.10252v1.
- Author dataset and derived results, DOI 10.5281/zenodo.21278557, CC BY 4.0.
- Author collection and analysis software, DOI 10.5281/zenodo.21278793, MIT.
- VerifyLLMAPI atlas JSON, built from reference
pamela-openrouter-2026-07-8cell-ref-a, SHA-25651f3f63cbcfb56b151055d2d10e36a9cc2f0644cc68b580784853d68951fa866.
Page evidence SHA-256: a546c2c8234a02f38c4ed58ef1059886ac50b1f76bfbc0c29598a1d1dad535fa. This hash covers the model ID, eight cell metrics, nearest references, and strongest cell difference in the generated atlas.