三体事实核验 Fact check of the Three body

三个模型Three models

三颗太阳是同一个想法的三次迭代。模型 1 只认关键词；模型 2 加入了顺序、亲属关系与否定；模型 3 加入了话题卡片、已知错误陈述、数值槽位与否定的作用范围。在 1000 道题的基准测试中，准确率分别为 57.6%、70.9% 和 77.4%。这里只用模型 3 作答。看它们如何设计 →

The three suns are three iterations of the same idea. Model 1 matches keywords; Model 2 adds word order, kinship and negation; Model 3 adds topic cards, known-false statements, value slots and negation scope. On a 1,000-statement benchmark they score 57.6%, 70.9% and 77.4%. Only Model 3 answers here. See how they were built →

模型 3 如何判断How Model 3 decides

• conv1每种措辞一个精确匹配检测器（2,695 个）one exact-match detector per wording (2,695)
• conv2否定作用范围negation scope
• pool概念：是否出现、位置、是否被否定per concept: present, where, negated
• L1–L7语序、事实、话题、矛盾、覆盖、最终逻辑order, facts, topics, contradictions, coverage, logic

局限Limits

它读的是关键词，不是语法。一句只由“本该同时出现”的名字和词语拼成的错误长句，会被判为真。不认识的词会被忽略。它遵循小说原著，而非电视剧改编。

It reads keywords, not grammar. A long false sentence built only from names and words that belong together in the story is answered true. Unknown words are ignored. It follows the novels, not the TV adaptations.