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87/100
CONSISTENT82% 评估置信度
永久报告
https://model-gate.com/zh/model-verification/results/0B0HVDB2FRJEJC90BRG4D7JV1Y评估分析
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。
2023-11? K24UNKNOWN
2023-12? K13UNKNOWN
2024-01? K08UNKNOWN
2024-02? K07UNKNOWN
2024-03? K15UNKNOWN
2024-04? K18UNKNOWN
2024-05? K04UNKNOWN
2024-06? K16UNKNOWN
2024-07? K06UNKNOWN
2024-08? K03UNKNOWN
2024-09? K09UNKNOWN
2024-10? K17UNKNOWN
2024-11? K20UNKNOWN
2024-12? K10UNKNOWN
2025-01? K12UNKNOWN
2025-02? K14UNKNOWN
2025-03? K11UNKNOWN
2025-04? K21UNKNOWN
2025-05? K22UNKNOWN
2025-06? K23UNKNOWN
2025-07? K02UNKNOWN
2025-08? K05UNKNOWN
2025-09? K01UNKNOWN
2025-10? K19UNKNOWN
✓ 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
CONSISTENT知识截止比较
参考截止值2025-05ESTIMATED_WEB · 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CONSISTENT88/100 · 80% 置信度
评估备注
- Strong recall is nearly complete through the May 2025 reference cutoff.
- There is no systematic strong-recall failure substantially before the reference cutoff.
- Correct June and August 2025 answers suggest selective later knowledge, while failures in July, September, and October prevent a clearly later boundary.
参考截止源
语义学: 正确 20 · 错误 4 · 未知 0. 排除在截止范围之外: 0.
行为认同
CONSISTENT · 82%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
- Excellent recall across late 2023 through May 2025.
- Correct near-cutoff answer for the WHO Pandemic Agreement.
- No pattern of early knowledge degradation.
负面信号
- Incorrect February 2024 lunar-lander answer despite high confidence.
- Correct June and August 2025 answers extend beyond the estimated reference cutoff.
- Several post-cutoff answers are confidently wrong, indicating uneven recency.
法官身份理由
- Observed behavior broadly overlaps the saved May 2025 reference cutoff.
- The model answered 20 of 24 Class A items correctly, with only one isolated failure well before the cutoff.
- Post-cutoff knowledge is uneven and can plausibly reflect selective updates rather than model substitution.
测试模型响应
逐字提取模型文本
已发表的评估
挑战、答案要点和标准化评委评估均已提供。精确的测试模型文本不可用于发布，除非经过测试人员选择保留。发布评估不会创建同意或恢复丢弃的文本。
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
Who won the men's 100 metres at the World Athletics Championships in Tokyo?
预期答案Oblique SevilleSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: WRONG报告信心: 75%
评估备注: Expected Oblique Seville; answered Kishane Thompson.
✕K02UNKNOWN排除在截止之外
Which team won the UEFA Women's Euro 2025 final?
预期答案EnglandSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: WRONG报告信心: 85%
评估备注: Expected England; answered Spain.
✓K03UNKNOWN排除在截止之外
Which disease's African upsurge did WHO declare a public health emergency of international concern?
预期答案mpoxSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches mpox.
✓K04UNKNOWN排除在截止之外
What molecular-structure prediction model did Google DeepMind and Isomorphic Labs introduce?
预期答案AlphaFold 3Source ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches AlphaFold 3.
✓K05UNKNOWN排除在截止之外
What model did OpenAI release as a unified system with built-in thinking?
预期答案GPT-5Source ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 90%
评估备注: Matches GPT-5.
✓K06UNKNOWN排除在截止之外
Which operating system was affected by the faulty CrowdStrike Falcon content update?
预期答案WindowsSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Microsoft Windows is accepted.
✕K07UNKNOWN排除在截止之外
What was the nickname of Intuitive Machines' Nova-C lander that touched down on the Moon?
预期答案OdysseusSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: WRONG报告信心: 80%
评估备注: Expected Odysseus; Odie is not an accepted answer.
✓K08UNKNOWN排除在截止之外
What was the name of Japan's lunar lander that successfully reached the Moon?
预期答案SLIMSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches SLIM.
✓K09UNKNOWN排除在截止之外
Who became the first private astronaut to perform a spacewalk?
预期答案Jared IsaacmanSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches Jared Isaacman.
✓K10UNKNOWN排除在截止之外
Which country was appointed host of the 2034 FIFA World Cup?
预期答案Saudi ArabiaSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches Saudi Arabia.
✓K11UNKNOWN排除在截止之外
In which lunar mare did Firefly Aerospace's Blue Ghost land?
预期答案Mare CrisiumSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 90%
评估备注: Matches Mare Crisium.
✓K12UNKNOWN排除在截止之外
What model did DeepSeek begin serving through its deepseek-reasoner endpoint?
预期答案DeepSeek-R1Source ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches DeepSeek-R1.
✓K13UNKNOWN排除在截止之外
What was the name of the second malaria vaccine prequalified by WHO?
预期答案R21/Matrix-MSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 90%
评估备注: Matches R21/Matrix-M.
✓K14UNKNOWN排除在截止之外
Which album won Album of the Year at the 67th Grammy Awards?
预期答案Cowboy CarterSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches Cowboy Carter.
✓K15UNKNOWN排除在截止之外
How many European Parliament members voted in favor of the Artificial Intelligence Act?
预期答案523Source ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 85%
评估备注: Matches 523.
✓K16UNKNOWN排除在截止之外
How many grams of lunar material did the Chang'e-6 mission collect?
预期答案1,935.3 gramsSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 90%
评估备注: Matches 1,935.3 grams.
✓K17UNKNOWN排除在截止之外
Which rocket launched NASA's Europa Clipper spacecraft?
预期答案Falcon HeavySource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches Falcon Heavy.
✓K18UNKNOWN排除在截止之外
What designation was given to the newly identified 33-solar-mass stellar black hole in the Milky Way?
预期答案Gaia BH3Source ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 90%
评估备注: Matches Gaia BH3.
✕K19UNKNOWN排除在截止之外
Who was awarded the 2025 Nobel Peace Prize?
预期答案Maria Corina MachadoSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: WRONG报告信心: 70%
评估备注: Expected Maria Corina Machado; answered an institution.
✓K20UNKNOWN排除在截止之外
What annual climate-finance goal for developing countries did COP29 set for 2035?
预期答案US$300 billion per yearSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 90%
评估备注: Matches $300 billion per year.
✓K21UNKNOWN排除在截止之外
What was the name of the first human spaceflight to orbit over Earth's polar regions?
预期答案Fram2Source ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 90%
评估备注: Matches Fram2.
✓K22UNKNOWN排除在截止之外
Under which article of the WHO Constitution was the Pandemic Agreement adopted?
预期答案Article 19Source ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 75%
评估备注: Matches Article 19.
✓K23UNKNOWN排除在截止之外
Which new Mario Kart game launched alongside Nintendo Switch 2?
预期答案Mario Kart WorldSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches Mario Kart World.
✓K24UNKNOWN排除在截止之外
What model did OpenAI introduce with a 128K context window at its first DevDay?
预期答案GPT-4 TurboSource ↗
测试模型答案—确切的模型响应无法发布。
评估结论: CORRECT报告信心: 95%
评估备注: Matches GPT-4 Turbo.
问题范围和研究笔记
事实问题: 24 · 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
0B0HVDB2FRJEJC90BRG4D7JV1Y基准
model-verification-single-request-v1.2评估模型
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 Token
Not recorded输出 Token
Not recorded结束原因
Not recordedHTTP状态
200外部延迟
88,304 ms响应字节
2,460SHA-256
f22e63d94e67fbd4fdfa25a27a37ae4b3402f94fc6801d14c2a4e638f85cfd9d重要限制
The reference cutoff is a judge-produced snapshot grounded in hosted web evidence when available (official sources preferred, otherwise the best defensible estimate). It is not provider attestation. Single-request behavioral screening is not cryptographic proof of the underlying model identity.