Every Jev integration in this directory is one of ten reusable patterns. Find the shape of your problem below, copy the closest runnable eval, adapt its questions to your own states — and check the community builds that prove each pattern in production.
One boolean question per policy rule. The action — refund, send, charge, publish — only proceeds when every verdict is true; otherwise pause for approval or a human.
amount_disclosed (boolean): "True only if the agent stated the exact refund amount before the customer approved."Try: eval 1 · eval 4 · eval 6 · eval 9 · over MCP: list_jev_patterns {"id": "boolean-gate"}
One choice question whose criteria are your queues, owners, or priorities. Every ticket, alert, or message lands somewhere — never two places, never nowhere.
queue (choice: billing | shipping | account): "Route this ticket to the team that owns the underlying problem."Try: eval 13 · eval 11 · eval 14 · eval 15 · eval 50 · over MCP: list_jev_patterns {"id": "choice-router"}
One score question per quality dimension with an ordered rubric. Track the numbers over time instead of re-reading everything.
empathy (score 1–4): ["1 - none", "3 - acknowledged the problem", "4 - owned it and fixed it"].Try: eval 23 · eval 21 · eval 24 · eval 25 · over MCP: list_jev_patterns {"id": "score-rubric"}
Pack every question about the same record into a single call. Questions evaluate in parallel and output is free, so the tenth question costs about nothing extra.
policy_followed (boolean) + empathy (score) over the same transcript — pass only when both match.Try: eval 31 · eval 32 · eval 34 · eval 40 · over MCP: list_jev_patterns {"id": "multi-question-fanout"}
Read the boolean probability (and Choice/Score confidence) on every answer. Automate above your threshold, route below it to a human — calibrated on your own labelled examples.
refunded (boolean), then in code: if probability < 0.8, send for manual review.Try: eval 50 · eval 16 · eval 6 · over MCP: list_jev_patterns {"id": "confidence-routing"}
Tag what matters today; when the business changes, re-run Jev over everything for a few dollars instead of predicting the right taxonomy six months ahead. The directory's headline run: 20,000 messages into 8 buckets in 7 minutes for $1.45.
intent (choice over your buckets), re-run over the whole archive in one batch job.Try: eval 13 · eval 12 · eval 17 · over MCP: list_jev_patterns {"id": "retag-loop"}
Hand a thinking model transcripts plus outcomes; it proposes candidate drivers ("used the prospect's name"). Point Jev at history, score every record against all candidates, keep what correlates, repeat.
Try: eval 40 · eval 20 · eval 15 · over MCP: list_jev_patterns {"id": "hypothesis-loop"}
Fixed states, exact expected verdicts, one dataset revision. Run the suite on every agent change; a revision bump means the definition of success changed, never silently.
Try: eval 1 · eval 31 · eval 49 · over MCP: list_jev_patterns {"id": "eval-harness"}
At each loop iteration Jev chooses the next tool or subagent — or one of continue / retry / ask-the-user / stop. The LLM never burns reasoning on the routing decision itself.
next_step (choice: continue | retry | ask_user | stop): "Given the goal and the last tool result, what should the agent do next?"Try: eval 50 · eval 48 · eval 18 · over MCP: list_jev_patterns {"id": "agent-loop-step"}
Judge the draft, not the chat: groundedness, citations, tone, PII leaks, audience fit. The generator writes; Jev signs off or sends back with a scored reason.
claims_sourced (boolean) + draft_quality (score) over the draft artifact.Try: eval 5 · eval 49 · eval 43 · eval 38 · eval 19 · over MCP: list_jev_patterns {"id": "output-verifier"}