488 ·GitHub
What is this?
LM Fingerpoint Detector figures out which large language model is really answering behind an API by asking it to write about 300 random integers between 1 and 355, then comparing the statistical pattern of those numbers against a reference bank of known models (GPT, Claude, Gemini, Grok, Qwen, DeepSeek, etc.) to find the closest match.
LM Fingerpoint Detector figures out which large language model is really answering behind an API by asking it to write about 300 random integers between 1 and 355, then comparing the statistical pattern of those numbers against a reference bank of known models (GPT, Claude, Gemini, Grok, Qwen, DeepSeek, etc.) to find the closest match.Is this a website or a command-line tool?Both: there's a web app for pasting a model's output or connecting directly to an API, and a Bun-based CLI called fpd that gives you a live terminal UI, multi-round detection, and JSON output.
Both: there's a web app for pasting a model's output or connecting directly to an API, and a Bun-based CLI called fpd that gives you a live terminal UI, multi-round detection, and JSON output.How do I run a detection?On the web, open lm.ikale.io and either paste a model's answer manually or configure direct API access; each answer needs at least 80 valid integers before a ranking appears. The fpd CLI runs the same checks from a terminal without any separate install.
On the web, open lm.ikale.io and either paste a model's answer manually or configure direct API access; each answer needs at least 80 valid integers before a ranking appears. The fpd CLI runs the same checks from a terminal without any separate install.Which API formats does it need to talk to the model?It speaks three protocols: OpenAI's Responses API, OpenAI's Chat Completions API, and Anthropic's Messages API, defaulting to SSE streaming responses.
It speaks three protocols: OpenAI's Responses API, OpenAI's Chat Completions API, and Anthropic's Messages API, defaulting to SSE streaming responses.Does a match actually prove which model is running?No — the result is just a ranking inside a closed reference set. Models not in the library still get a 'closest' candidate, and neither the rank score nor the confidence value is proof of real identity.
No — the result is just a ranking inside a closed reference set. Models not in the library still get a 'closest' candidate, and neither the rank score nor the confidence value is proof of real identity.What does the optional tokenizer probing check?When turned on (web toggle or --tokenizer flag), it sends about 12 short requests and reads the token usage field the API returns to infer which tokenizer the upstream model uses, then checks it against the claimed model. It's only a secondary signal, since many different models share the same tokenizer.
When turned on (web toggle or --tokenizer flag), it sends about 12 short requests and reads the token usage field the API returns to infer which tokenizer the upstream model uses, then checks it against the claimed model. It's only a secondary signal, since many different models share the same tokenizer.How does the reference library get updated with new models?The fpd CLI has sample, enroll, and retrain commands that run sampling, database insertion, and offline retraining in sequence, with failed attempts and old samples kept for traceability.
The fpd CLI has sample, enroll, and retrain commands that run sampling, database insertion, and offline retraining in sequence, with failed attempts and old samples kept for traceability.Can I avoid routing requests through the official site?Yes — the connection mode can be set to direct, routed through the site's proxy, or through your own self-deployed Cloudflare Worker; the project can also be deployed to Cloudflare Pages, GitHub Pages, or Vercel.
Yes — the connection mode can be set to direct, routed through the site's proxy, or through your own self-deployed Cloudflare Worker; the project can also be deployed to Cloudflare Pages, GitHub Pages, or Vercel.Can I just browse the reference data without running a detection?Yes, the web app has a reference library page where you can browse and export the stored model fingerprints read-only.
Yes, the web app has a reference library page where you can browse and export the stored model fingerprints read-only.What are similar projects?openclaw-zero-token, ModelTrace, claude-code, dyad, aichat, Ai-jailbreak, opengrok, PriceAI
openclaw-zero-token, ModelTrace, claude-code, dyad, aichat, Ai-jailbreak, opengrok, PriceAIWhich topics is it in?fingerprinting, detection, tokenization, llms, cloudflare-workers, grok, serverless, bun, cloudflare, qwen, tui, large-language-models
fingerprinting, detection, tokenization, llms, cloudflare-workers, grok, serverless, bun, cloudflare, qwen, tui, large-language-models