Jev fails nine ways. You handle each in code.

Jev documents its failures. You design around them. The 1.13 jaggedness list is at docs.typesafe.ai/model-jaggedness and defines the other half of the API contract. Each item names work you own in code. What the model cannot do, you handle.

The Jev Test flowchart triages in three checks: prose needed, code-sufficient, judgments atomic. Then you choose from five verdicts. native means pure judgment. hybrid splits judgment to Jev and prose to the LLM. About 60 percent of enterprise cases land here. weak means a small salvageable core. not_jev means the wrong tool. use_code means no AI is needed. need_info covers inputs that need more detail. The badge names the verdict. The numbers beside it give the score band in small monospace type.

Each verdict divides work four ways (SPEC section 5.4). You keep exact checks, math, and dates in code: under 1ms, $0, no fuzzy judgment. You keep regex, thresholds, and state prep there too. You give Jev atomic Choice, Score, and Noul gating. That path takes about 100ms and costs $0.042 per 1M tokens, with no prose or multi-hop reasoning. You give frontier LLMs long-form prose and synthesis. That path takes 2 to 8 seconds and costs $5 to $15 per 1M tokens. You keep LLMs out of tight latency loops. You review confidence edges from 0.3 to 0.7. That review takes hours to days. It costs more than any other path, so you spend none of it on repetitive triage. With broad input such as "manage our sales pipeline", you isolate the Jev-native core: lead Score, intent Choice, and budget and churn Nouls. Route the rest away.

You archive the evaluator when TypeSafe ships official translation from problems to typed questions or keeps waitlist access closed past 90 days. You also archive it when this page turns into a generic prompt directory. Until then, paste the fuzzy version and answer the typed questions.