任务域决策包
schema 校验层不可省略
c
我需要从非结构化客户反馈中提取 JSON
r
validate@structured_data/json_extraction
C'
标注清晰、高吞吐输入用 Sonnet 4.6。字段含糊时用 Opus 4.8。所有生产流水线都需要确定性 schema 校验,模型不是门禁。
@structured_data/json_extraction
schema 校验层不可省略
我需要从非结构化客户反馈中提取 JSON
validate@structured_data/json_extraction
标注清晰、高吞吐输入用 Sonnet 4.6。字段含糊时用 Opus 4.8。所有生产流水线都需要确定性 schema 校验,模型不是门禁。
schema well-defined, inputs labeled, batch volume high
claude/sonnet-4.6ambiguous inputs, implicit fields, accuracy-critical pipeline
claude/opus-4.8output feeds automated system without human review
openai/gpt-5.5any production extraction pipeline
deterministic_validator| ID | 关系 | 执行体 | 结果 | 输出 |
|---|---|---|---|---|
| CT-D6-001 | reject@hallucination | claude/opus-4.8 | 已接受 | more likely to express uncertainty than fill missing fields with invented values |
| CT-D6-002 | reject@hallucination | openai/gpt-5.5 | 拒绝 | 没有下游校验时不适合作为生产唯一提取器 |
| CT-D6-003 | validate@schema_compliance | claude/sonnet-4.6 | 有边界 | reliable for labeled inputs; hallucination increases on ambiguous or unlabeled fields |
Opus 4.8 hallucination rate significantly lower than GPT-5.5
E-GLOBAL-003Sonnet 4.6 reliable for well-defined schemas with explicit markers
根据相邻证据推断: instruction-following improvements