I catalogued every way local models break JSON output and built a repair library across 288 model calls

I catalogued every way local models break JSON output and built a repair library, here's what I found across 288 model calls

Reddit user kexxty ran 288 structured-output calls through OpenRouter models, including Llama 3, Mistral, Command R, DeepSeek, and Qwen, and found similar JSON failure categories across local and API-only models. The MIT-licensed Python library outputguard validates against JSON Schema, applies 15 ordered repair strategies, includes 2,001 tests, and has no LLM provider dependency.

Why it matters: HKR-H/K/R all pass: 288 tests, the outputguard library, and a 15-step repair chain give practitioners reusable detail. Source is a single Reddit post, so it stays in the 72–77 featured band, not 78+.