AI coding

Structured output and JSON reliability

• JSON Mode, a Schema or a Tool Call for Structured Output
• Payload Shapes Across Model Families: An Acceptance Table
• Prompting the Cheap Tier to Say When It Is Unsure
• A Hosted Platform Is Its Own Row in the Matrix
• A Repair Step That Keeps the Original Payload
• A Schema Conformance Suite That Runs Against Every Model
• A Schema Registry for Agent Prompts: Versioning and Rollback
• Handling Tool-Call Arguments a Model Sends in the Wrong Type
• Measuring Wrong-Type Arguments per Tool and per Model
• Retries That Make a Model's JSON Worse and When to Stop
• Retries That Make a Model's JSON Worse and Where to Stop
• Same Shape, Different Meaning: The Semantic Gap Between Providers
• Seed a Repair Suite with Malformed JSON from Real Traces
• Structured Output Schemas for Test Results: What Agents Report
• Structured Output That Survives Schema Drift
• Teaching a Model to Cite the Field It Used
• What to Log When a Model Returns Bad JSON: Fields for Debugging
• When a Parse Failure Is a Repair, Not a Model Problem
• When a Tool Call Beats JSON Mode: Latency, Cost and Repair