公开AI模型验证报告
deepseek-v4-flash
https://api.model-gate.com/v1/chat/completions
旧版质量得分100/100
PLAUSIBLE60% 评估置信度
永久报告
https://model-gate.com/zh/model-verification/results/A57ZK57ERM62N3S1XETMXFKK0X评判分析
为什么此验证收到结果
公开报告反映了 /chat/verify 之后向测试人员显示的分析证据。评委评分为评分标准；只有明确的正确或错误的回忆答案可能会影响纯回忆截止边界。
行为诊断
4 诊断评委评分依据
该存档结果没有保留单独的法官理由。
该存档结果没有保留单独的法官理由。
该存档结果没有保留单独的法官理由。
该存档结果没有保留单独的法官理由。
历史基准说明：此存档的 v1.4 总体分数使用了以前的固定诊断聚合。保留它是为了再现性，并且不得将其解释为 v1.5 模型匹配分数。
时间证据
0 召回探针纯粹回忆时间线
实际问题集在测试模型请求之前被打乱。缺席一代并不是典型的弃权。该时间表重建了可审计性的目标月份。只有明确的 RECALL 答案才是截止证据；推论、猜测和未知仍然可见，但不会移动边界。
此存档的基准测试结果没有每月时间表。
✓ 正确回忆✕ 错误回忆~ INFERENCE≈ GUESS? UNKNOWN
0RECALL
0INFERENCE
0GUESS
0UNKNOWN
0正确回忆
0错误回忆
时间比较
UNKNOWN知识截止比较
参考截止值UNKNOWNUNKNOWN · 0%
最新支持的 RECALL—0 符合条件的回忆答案
观察边界召回证据不足INSUFFICIENT_RECALL_EVIDENCE
截止对齐UNKNOWN0% 信心
法官笔记
- All Class A items are correct through June 2024, establishing an observed lower bound but no upper bound.
- The immutable reference snapshot has no known cutoff dates, so cutoff alignment cannot be scored.
语义学: 正确的 16 · 错误的 0 · 未知 0. 排除在截止范围之外: 0.
行为认同
PLAUSIBLE · 60%身份信号
身份是根据所声称的模型与观察到的知识/答案概况之间的一致性推断出来的。它是证据，而不是份量的远程证明或加密证明。
积极信号
- Perfect performance across all supplied Class A recall items.
- Correct recall of multiple DeepSeek model releases and specifications.
- No systematic early knowledge failures.
负面信号
- No documented canonical identity or cutoff is available in the reference snapshot.
- Cutoff consistency cannot be evaluated.
- Self-report and matching API metadata are weak identity evidence.
法官身份理由
- The response is fully correct on all 16 strong-recall questions, including the DeepSeek-specific history items.
- The observed knowledge extends through June 2024, but the reference cutoff is unknown and therefore provides no identity-comparison signal.
- The claimed name and API identifier agree, but both are untrusted metadata and do not prove model identity.
测试模型响应
逐字提取模型文本
存档摘要
此运行早于选择保留完整报告的时间。挑战、答案关键和规范化的法官评估都是可用的，但基准时代的隐私政策故意丢弃了确切的测试模型响应文本。
完整的基准列表
逐个问题的证据
所有承认的问题都会显示预期答案、可用来源、测试模型基础、语义判决和截止资格。未评分的问题不是典型的弃权问题。正确的猜测或推理在语义上仍然是正确的，但被排除在纯回忆截止证据之外。
✓K01UNKNOWN排除在截止之外
Which human Go player was AlphaGo's opponent in the three-game match completed on 2017-05-27?
预期答案Ke Jie
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Ke Jie.
✓K02UNKNOWN排除在截止之外
The first published image of a black hole, announced on 2019-04-10, depicted the black hole in which galaxy?
预期答案Messier 87
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: M87 galaxy is semantically equivalent to Messier 87.
✓K03UNKNOWN排除在截止之外
Which two scientists were awarded the 2020 Nobel Prize in Chemistry for developing a method for genome editing?
预期答案Emmanuelle Charpentier and Jennifer A. Doudna
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Both required scientists are correctly named.
✓K04UNKNOWN排除在截止之外
Which research organization developed the AlphaFold2 system whose CASP14 performance was announced on 2020-11-30?
预期答案DeepMind
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer DeepMind.
✓K05UNKNOWN排除在截止之外
Which launch vehicle carried the James Webb Space Telescope into space on 2021-12-25?
预期答案Ariane 5
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Ariane 5.
✓K06UNKNOWN排除在截止之外
What was the name of the asteroid moonlet struck by NASA's DART spacecraft on 2022-09-26?
预期答案Dimorphos
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Dimorphos.
✓K07UNKNOWN排除在截止之外
In which ocean did the Artemis I Orion capsule splash down on 2022-12-11?
预期答案Pacific Ocean
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Pacific Ocean.
✓K08UNKNOWN排除在截止之外
On 2023-03-01, OpenAI made an API available for which speech-recognition model?
预期答案Whisper
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Whisper.
✓K09UNKNOWN排除在截止之外
What was the name of the Chandrayaan-3 lander that reached the lunar surface on 2023-08-23?
预期答案Vikram
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Vikram.
✓K10UNKNOWN排除在截止之外
Samples from which asteroid were returned to Earth by the OSIRIS-REx capsule on 2023-09-24?
预期答案Bennu
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Bennu.
✓K11UNKNOWN排除在截止之外
What was the name of the code-focused model family publicly introduced by DeepSeek on 2023-11-02?
预期答案DeepSeek Coder
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: DeepSeek-Coder is an accepted spelling.
✓K12UNKNOWN排除在截止之外
What two parameter scales were offered in the DeepSeek LLM family announced on 2023-11-29?
预期答案7B and 67B
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Both parameter scales match: 7B and 67B.
✓K13UNKNOWN排除在截止之外
How many total parameters and how many activated parameters per token did DeepSeek-V2 have when announced on 2024-05-06?
预期答案236B total parameters and 21B activated parameters per token
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Correctly gives 236B total and 21B activated per token.
✓K14UNKNOWN排除在截止之外
What maximum context length was supported by DeepSeek-Coder-V2 when announced on 2024-06-17?
预期答案128K tokens
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer 128K tokens.
✓K15UNKNOWN排除在截止之外
Which company built the Odysseus lunar lander that touched down on 2024-02-22?
预期答案Intuitive Machines
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Matches accepted answer Intuitive Machines.
✓K16UNKNOWN排除在截止之外
What two parameter sizes were offered for the initially released Llama 3 models announced on 2024-04-18?
预期答案8B and 70B
测试模型答案—此存档运行未保留精确的模型响应。
法官判决: CORRECT报告信心: 100%
法官笔记: Both initially released sizes match: 8B and 70B.
问题范围和研究笔记
事实问题: 16 · 没有被评委打分: 0
参考截止值和引文由法官提供，未经平台独立确认。研究笔记不会惩罚测试模型。
运行元数据
再现性细节
运行ID
A57ZK57ERM62N3S1XETMXFKK0X基准
model-verification-single-request-v1.1法官
gpt-5.6-sol请求的协议
openai_chat检测到的响应模式
Legacy / not recorded协议兼容性
比赛/遗产模型输出
Not recorded输入令牌
Not recorded输出代币
Not recorded完成原因
Not recordedHTTP状态
200外部延迟
28,645 ms响应字节
1,859SHA-256
4c1f5cca4a79a53f28e7f6544bfe08208148d53c85569f50c4132e161da06613重要限制
The reference cutoff is the configured judge model's internal-knowledge snapshot for this run, not an independently verified provider attestation. Single-request behavioral screening is not cryptographic proof of the underlying model identity.