JJEV·DIRECTORY GitHub agent pack connect your agent
Decision patterns

Ten ways to use a judge model.

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.

1. Boolean gate — stop the action unless the rule holds

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.

WhenRefunds, sends, charges, publishes — anywhere an agent must prove it disclosed, redacted, or confirmed something before acting.
Askamount_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"}

2. Choice router — send each item to exactly one owner

One choice question whose criteria are your queues, owners, or priorities. Every ticket, alert, or message lands somewhere — never two places, never nowhere.

WhenSupport triage, bug intake, lead routing, escalation queues, moderation dispositions.
Askqueue (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"}

3. Score rubric — grade quality on your own scale

One score question per quality dimension with an ordered rubric. Track the numbers over time instead of re-reading everything.

WhenReviewing agent replies, summaries, drafts, translations, plans — anything where "good" is a spectrum, not a switch.
Askempathy (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"}

4. Multi-question fan-out — one state, many verdicts, one call

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.

WhenAuditing transcripts, calls, or artifacts where you need the gate and the route and the grade together.
Askpolicy_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"}

5. Confidence routing — automate the clear, escalate the unsure

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.

WhenAny automation where a wrong auto-decision costs more than a review: refunds, access grants, publishes, medical or legal drafts.
Askrefunded (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"}

6. Retag loop — reclassify all of history when the question changes

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.

WhenEmail, Slack, and transcript archives, CRM hygiene, content libraries — any corpus whose categories keep evolving.
Askintent (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"}

7. Hypothesis loop — let the LLM guess, let Jev measure

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.

WhenSales-call analysis, support deflection, churn drivers — anywhere you suspect patterns but can't hand-label enough data to prove them.
AskTwenty booleans (one per hypothesis) over every call transcript, joined to outcomes in your warehouse.

Try: eval 40 · eval 20 · eval 15 · over MCP: list_jev_patterns {"id": "hypothesis-loop"}

8. Eval harness — pin agent behavior with fixed states + verdicts

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.

WhenCI gates for agent behavior, regression suites for prompts, comparing two agent versions on identical inputs.
AskThis directory's 50 evals are the template: exact match per question, pass iff every question matches.

Try: eval 1 · eval 31 · eval 49 · over MCP: list_jev_patterns {"id": "eval-harness"}

9. Agent-loop step — Jev picks the next move, the LLM executes it

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.

WhenMulti-step agents, tool-heavy tasks, browser automation — anywhere routing tokens dominate the bill.
Asknext_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"}

10. Output verifier — check the artifact before it ships

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.

WhenNewsletters, release notes, RAG answers, code reviews, generated reports — any artifact with a bar to clear.
Askclaims_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"}

Not sure which pattern fits? Describe your product to the MCP server's recommend_jev_use_cases — or start from the ten best builds and work backwards.