@structured_data/json_extraction
schema validation layer is non-negotiable regardless of model
I need to extract structured JSON from unstructured customer feedback
validate@structured_data/json_extraction
Sonnet 4.6 for labeled, high-volume inputs. Opus 4.8 for ambiguous fields. All pipelines require deterministic schema validation; model is not the gate.
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 | Relation | Actor | Result | Output |
|---|---|---|---|---|
| CT-D6-001 | reject@hallucination | claude/opus-4.8 | accepted | more likely to express uncertainty than fill missing fields with invented values |
| CT-D6-002 | reject@hallucination | openai/gpt-5.5 | rejected | unsuitable as sole extractor in production without downstream validation |
| CT-D6-003 | validate@schema_compliance | claude/sonnet-4.6 | bounded | 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
inferred from instruction-following improvements