# Reference fingerprints (trusted, collected with this tool) These are **first-party** references — collected by `llm-fingerprint-detector` itself against a provider you trust, using the package's own `one-token/v1` battery. Unlike `../reference-fingerprints.sample.json` (Zenodo paper data, `bruckner-zenodo-2026` protocol), these are protocol-matched, so a `compare()` against them is a direct, non-indicative comparison. | file | model id | provider | collected | protocol | battery | |---|---|---|---|---|---| | `deepseek-v4.1-flash.official.json` | `deepseek-flash` | api.deepseek.com (official) | 2026-10-04 | `one-token/v1` | 8 cells × 25 samples | ## `deepseek-v4.1-flash.official.json` - **Endpoint:** `https://api.deepseek.com/v1` - **Model id requested:** `deepseek-flash` (API-reported name: `DeepSeek-V4.1-Flash`) - **Collected:** 2026-10-04, standard preset, `temperature=1`, reasoning disabled via the `openai-effort` adapter - **Result:** 200/200 samples valid, 0 errors, `postReasoning: false` - **Use it to verify a reseller:** a reseller advertising `deepseek-v4.1-flash` (or `deepseek-flash`) should land in the *match* band against this file. ```bash # verify a suspect endpoint against this trusted reference llm-fingerprint verify \ --base-url https://reseller.example.com/v1 \ --model deepseek-v4.1-flash \ --reference data/references/deepseek-v4.1-flash.official.json ``` ### The official fingerprint's most discriminative cells | cell | official DeepSeek-V4.1-Flash | notes | |---|---|---| | `random-number-1-100:en` | `42` ×19 / 25 | overwhelming 42 bias | | `random-number-1-100:zh` | `42` ×14 / 25 | same bias, Chinese | | `random-color:en` | `blue` ×25 / 25 | fully deterministic | | `coin-flip:en` | `heads` ×25 / 25 | fully deterministic | | `random-number-1-10:en` | `7` ×23 / 25 | near-deterministic | | `random-letter:en` | `q` ×18 / 25 | q ≫ k ≫ g | | `random-color:zh` | `蓝` ×20 / 25 | 蓝 ≫ 紫, 橙 | | `random-animal:en` | `elephant` ×11, `cat` ×9 | elephant dominates | **Caveat:** references drift as providers ship model updates. Re-collect if the file is more than ~1–2 months old, or if a verification you trust starts returning `uncertain`.