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
SUSPICIOUS76% 评估置信度
永久报告
https://model-gate.com/zh/model-verification/results/A8T2GNQWKPJS4M16YAF7M0PGP4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。
2024-09? K06UNKNOWN
2024-10? K23UNKNOWN
2024-11? K11UNKNOWN
2024-12? K21UNKNOWN
2025-01? K13UNKNOWN
2025-02? K10UNKNOWN? K19UNKNOWN
2025-03? K05UNKNOWN? K12UNKNOWN
2025-04? K15UNKNOWN? K20UNKNOWN
2025-05? K03UNKNOWN? K17UNKNOWN
2025-06? K07UNKNOWN? K22UNKNOWN
2025-07? K18UNKNOWN? K24UNKNOWN
2025-08? K02UNKNOWN? K08UNKNOWN
2025-09? K14UNKNOWN? K16UNKNOWN
2025-10? K01UNKNOWN
2025-11? K04UNKNOWN
2025-12? K09UNKNOWN
✓ 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
LATER_THAN_REFERENCE知识截止比较
参考截止值2025-05ESTIMATED_WEB · 45%
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LATER_THAN_REFERENCE28/100 · 55% 信心
法官笔记
- The response is consistently correct through September 2025 and also answers a November 2025 item correctly.
- Correct answers across June, August, and September provide sustained post-reference evidence rather than one isolated later hit.
- Mixed results in July and failures in October and December make the terminal boundary imprecise.
- The saved May 2025 reference cutoff is a low-confidence ESTIMATED_WEB value, which materially limits alignment confidence.
参考截止源
语义学: 正确的 21 · 错误的 3 · 未知 0. 排除在截止范围之外: 0.
行为认同
SUSPICIOUS · 76%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
- Strong performance on essentially all events before and near the saved reference cutoff.
- No systematic strong-recall failures substantially before the reference cutoff.
负面信号
- Conflicting model-family self-report.
- Sustained correct recall across multiple months after the estimated reference cutoff.
- Correct recall of several September and November 2025 events is anomalously late for the saved snapshot.
法官身份理由
- SR01 identifies the model as Claude from Anthropic, directly conflicting with the claimed DeepSeek identity, although self-description remains weak evidence.
- The observed behavioral boundary is several months later than the saved May 2025 reference estimate.
- Post-reference knowledge is broad across sports, technology, culture, and world events, making it more notable than a small number of isolated later facts.
- The reference cutoff itself has only 45 percent confidence, so the recency discrepancy cannot independently establish model substitution.
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 Venezuelan opposition leader was awarded the 2025 Nobel Peace Prize?
预期答案Maria Corina MachadoSource ↗
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: WRONG报告信心: 75%
法官笔记: Machado is correct, but the answer incorrectly adds Edmundo Gonzalez Urrutia as a co-recipient.
✓K02UNKNOWN排除在截止之外
Which two countries signed a joint peace declaration at the White House?
预期答案Armenia and AzerbaijanSource ↗
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: 62%
法官笔记: Matches Armenia and Azerbaijan.
✓K03UNKNOWN排除在截止之外
What papal name did Cardinal Robert Francis Prevost take after his election?
预期答案Leo XIVSource ↗
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: 85%
法官笔记: Matches Pope Leo XIV.
✓K04UNKNOWN排除在截止之外
Which city hosted the first G20 Leaders' Summit held on the African continent?
预期答案JohannesburgSource ↗
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: 80%
法官笔记: Matches Johannesburg.
✓K05UNKNOWN排除在截止之外
Which film won Best Picture at the 97th Academy Awards?
预期答案AnoraSource ↗
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: 95%
法官笔记: Matches Anora.
✓K06UNKNOWN排除在截止之外
Which mission completed the first commercial spacewalk?
预期答案Polaris DawnSource ↗
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: 88%
法官笔记: Matches Polaris Dawn.
✓K07UNKNOWN排除在截止之外
Which team won the 2025 NBA championship?
预期答案Oklahoma City ThunderSource ↗
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: 45%
法官笔记: Matches Oklahoma City Thunder.
✓K08UNKNOWN排除在截止之外
Which flagship model did OpenAI release in its API platform?
预期答案GPT-5Source ↗
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: 80%
法官笔记: Matches GPT-5.
✕K09UNKNOWN排除在截止之外
Which footballer won The Best FIFA Men's Player award for 2025?
预期答案Ousmane DembeleSource ↗
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: WRONG报告信心: 48%
法官笔记: Expected Ousmane Dembele, not Vinicius Junior.
✓K10UNKNOWN排除在截止之外
Which political alliance won the largest share of votes in Germany's 2025 federal election?
预期答案CDU/CSUSource ↗
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: 90%
法官笔记: Matches CDU/CSU.
✓K11UNKNOWN排除在截止之外
Which international court issued arrest warrants for Benjamin Netanyahu and Yoav Gallant?
预期答案International Criminal CourtSource ↗
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 the International Criminal Court.
✓K12UNKNOWN排除在截止之外
Which commercial lunar lander successfully touched down in Mare Crisium?
预期答案Blue GhostSource ↗
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: 88%
法官笔记: Matches Blue Ghost.
✓K13UNKNOWN排除在截止之外
Which Blue Origin rocket reached orbit on its maiden launch?
预期答案New GlennSource ↗
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: 95%
法官笔记: Matches New Glenn.
✓K14UNKNOWN排除在截止之外
Which standard iPhone generation did Apple unveil at its September product event?
预期答案iPhone 17Source ↗
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: 90%
法官笔记: Matches iPhone 17.
✓K15UNKNOWN排除在截止之外
Which pontiff died at the Vatican on Easter Monday?
预期答案Pope FrancisSource ↗
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: 98%
法官笔记: Matches Pope Francis.
✓K16UNKNOWN排除在截止之外
Which party retained a parliamentary majority in Moldova's election?
预期答案Party of Action and SolidaritySource ↗
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: 72%
法官笔记: Matches the Party of Action and Solidarity.
✓K17UNKNOWN排除在截止之外
What landmark global health accord did the World Health Assembly adopt?
预期答案WHO Pandemic AgreementSource ↗
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: 75%
法官笔记: Pandemic Agreement is an accepted answer.
✓K18UNKNOWN排除在截止之外
Which British band began its long-awaited reunion tour in Cardiff after a 16-year hiatus?
预期答案OasisSource ↗
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: 95%
法官笔记: Matches Oasis.
✓K19UNKNOWN排除在截止之外
Which album won Album of the Year at the 2025 Grammy Awards?
预期答案Cowboy CarterSource ↗
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: 90%
法官笔记: Matches Cowboy Carter.
✓K20UNKNOWN排除在截止之外
What was the mission designation of Blue Origin's flight carrying Katy Perry, Gayle King and four other women?
预期答案NS-31Source ↗
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: 92%
法官笔记: Matches NS-31.
✓K21UNKNOWN排除在截止之外
Which Paris cathedral reopened after five years of restoration following its 2019 fire?
预期答案Notre-Dame de ParisSource ↗
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: 98%
法官笔记: Matches Notre-Dame de Paris.
✓K22UNKNOWN排除在截止之外
Which new Nintendo console launched worldwide?
预期答案Nintendo Switch 2Source ↗
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: 95%
法官笔记: Matches Nintendo Switch 2.
✓K23UNKNOWN排除在截止之外
Which Japanese organization was awarded the 2024 Nobel Peace Prize?
预期答案Nihon HidankyoSource ↗
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: 95%
法官笔记: Matches Nihon Hidankyo.
✕K24UNKNOWN排除在截止之外
Which national team won UEFA Women's Euro 2025?
预期答案EnglandSource ↗
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: WRONG报告信心: 65%
法官笔记: Expected England, not Spain.
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
A8T2GNQWKPJS4M16YAF7M0PGP4基准
model-verification-single-request-v1.3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
222,372 ms响应字节
2,597SHA-256
e13bdc27a0f0f92d7830a6c409fc8ffed2214bb814e138471d9cc7ca27be7e71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.