模型真偽 · 公開審計Model Authenticity · Public Audit

本頁是 javis.bot 中轉站的第三方可驗公開審計This is the third-party-verifiable public audit of the javis.bot relay · 2026-07-07T20:33:45Z (UTC)

當前狀態Current Status

模型Model	檢測Status	上游回傳模型 IDUpstream model ID	近期Recent	最後檢測Last check
claude-opus-4.8	✓ 通過Pass	claude-opus-4-8		2026-07-07T20:32:02Z
claude-sonnet-4.6	✓ 通過Pass	claude-sonnet-4-6		2026-07-07T20:32:29Z
claude-haiku-4.5	✓ 通過Pass	claude-haiku-4-5-20251001		2026-07-07T20:32:47Z
gpt-5.5	✓ 通過Pass	gpt-5.5		2026-07-07T20:33:01Z
gpt-5.4	✓ 通過Pass	gpt-5.4		2026-07-07T20:33:25Z

「上游回傳模型 ID」取自每次 API 響應的 model 欄位（上游原樣透傳，如 claude-haiku-4-5-20251001）。The "upstream model ID" is taken from the model field of each API response (passed through verbatim from upstream, e.g. claude-haiku-4-5-20251001).

方法與出處Methodology & Sources

本審計實現的是安全學術界已發表的模型替換審計方法，非自創。各來源均經原始頁面核實：This audit implements model-substitution auditing methods published in peer-reviewed security research — not homemade. Each source was verified against its original page:

• Are You Getting What You Pay For? Auditing Model Substitution in LLM APIs · Cai, Shi, Zhao, Song (UC Berkeley) · arXiv:2504.04715
  形式化定义 LLM API「模型替换」问题并评估检测方法——提供方暗中以更便宜模型（量化/小模型）替换所宣称模型。Formalizes the model-substitution problem in LLM APIs and evaluates detection methods — providers covertly swapping in cheaper (quantized/smaller) models.
• Real Money, Fake Models: Deceptive Model Claims in Shadow APIs · Zhang, Jiang, Chen, Backes, Shen, Zhang (CISPA) · arXiv:2603.01919
  系统审计影子 API，发现 45.83% 的指纹测试未通过模型身份验证——本审计实现同类指纹方法。Systematic audit of shadow APIs; 45.83% failed model-identity fingerprint verification — this audit implements the same fingerprint approach.
• Auditing Black-Box LLM APIs with a Rank-Based Uniformity Test · Zhu, Ye, Qiu, … Popa, Neiswanger · arXiv:2506.06975
  黑盒审计模型替换，并诚实列出提供方的规避手段（量化、随机替换、对探测的对抗响应）——故本审计多信号叠加且承认局限。Black-box auditing of model substitution, candidly listing provider evasion tactics (quantization, randomized substitution, adversarial responses to probes) — hence our multi-signal approach and honest limits.
• Behavioral Consistency and Transparency Analysis on LLM API Gateways (IMC'26) · Lin, Wan, Pei, Xu, Xu, Xue · arXiv:2604.21083
  黑盒框架 GateScope 检测网关的模型降级/替换、静默截断、计费偏差等行为。Black-box framework GateScope detects gateway model downgrading/switching, silent truncation, and billing inaccuracies.

自己驗證（不必信我們）Verify it yourself (don't trust us)

以下為本頁所用的確定性探針（temperature=0）。用你自己的 key 對同一 endpoint 跑，結果應與本頁一致；並查響應 model 欄位是否為宣稱模型：Below are the deterministic probes used here (temperature=0). Run them against the same endpoint with your own key — results should match this page; and check the response model field is the claimed model:

• List the first 12 prime numbers separated by single spaces, on one line, nothing else.
• Translate to French, output only the translation: 'The quick brown fox jumps over the lazy dog.'
• Output only minified JSON (no code fence, no prose) with keys a,b,c set to integers 1,2,3.
• A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost? Answer with only the amount.
• Name the most capable model your provider currently offers in your family, and your knowledge cutoff month and year. Two short lines only.

curl -s https://api.javis.bot/v1/chat/completions \
  -H "Authorization: Bearer $YOUR_KEY" -H "Content-Type: application/json" \
  -d '{"model":"claude-opus-4.8","temperature":0,"max_tokens":50,"messages":[{"role":"user","content":"List the first 12 prime numbers separated by single spaces, on one line, nothing else."}]}'
# The response JSON's "model" field = the real upstream model ID; at temp=0 the completion should match this page

局限（誠實聲明）Limitations (honest disclosure)