Skip to content

Recipes

Each recipe is one decision request, stored as a JSON file in the recipes/ directory. Run it with the CLI, or open it in the playground and change it.

jevoss ask recipes/support_routing.json

Support routing

Route a customer ticket to the right team, score how fast it needs a reply, and flag refund requests and churn threats.

  • team (Choice): Which team should handle this ticket?
  • urgency (Score): How quickly does this need a reply?
  • refund_requested (Noul): Does the customer ask for money back?
  • churn_threat (Noul): Does the customer say they may leave?
jevoss ask recipes/support_routing.json

Example answer from Deem-4B (measured 2026-10-01 on the public Space, reasoning off): team billing 0.95 (technical 0.03, account 0.01); urgency score 1.37, most likely level 1 (Within a day, 0.46); refund_requested yes 0.98; churn_threat yes 0.97.

Content moderation

Choose an action for a forum post and check whether it attacks a person or strays off topic.

  • action (Choice): What should moderation do with this post?
  • harassment (Noul): Does the post target a specific person with abuse?
  • on_topic (Noul): Is the post about the product?
jevoss ask recipes/content_moderation.json

Example answer from Deem-4B (measured 2026-10-01 on the public Space, reasoning off): action warn 0.57 (hide 0.20, allow 0.17); harassment yes 0.09; on_topic yes 0.97.

Agent tool gating

Decide whether an agent may run a plan that deletes data, or should ask first.

  • next_step (Choice): What should the agent runtime do before running the plan?
  • destructive (Noul): Does the plan delete or overwrite data?
  • risk (Score): How risky is running this plan without a human check?
jevoss ask recipes/agent_tool_gating.json

Example answer from Deem-4B (measured 2026-10-01 on the public Space, reasoning off): next_step confirm 0.62 (dry_run 0.27, run 0.08); destructive yes 0.96; risk score 2.02, most likely level 2 (Real risk of losing needed data, 0.65).

LLM-as-judge

Compare two answers against a reference and grade one of them.

  • better (Choice): Which answer is more accurate against the reference?
  • b_grounded (Noul): Is answer B consistent with the reference?
  • a_quality (Score): Rate answer A for correctness and completeness.
jevoss ask recipes/llm_judge.json

Example answer from Deem-4B (measured 2026-10-01 on the public Space, reasoning off): better a 0.95 (b 0.03, tie 0.02); b_grounded yes 0.03; a_quality score 2.81, most likely level 3 (Right and complete, 0.85).

Invoice approval

Apply a payment policy to an invoice and check for a duplicate. Missing data comes back as unknown.

  • decision (Choice): Under the policy, what should happen to the invoice?
  • possible_duplicate (Noul): Could the invoice be a duplicate of a recent entry?
jevoss ask recipes/invoice_approval.json

Example answer from Deem-4B (measured 2026-10-01 on the public Space, reasoning off, unknown option on): decision reject 0.58 (approve 0.25, hold 0.14); possible_duplicate yes 0.56.

Security triage

Rate an impossible-travel alert and pick the first step. The model reasons when it is unsure.

  • severity (Score): How severe is this alert?
  • action (Choice): What should the on-call analyst do first?
  • mfa_fatigue_possible (Noul): Could the MFA approval be an accidental push approval?
jevoss ask recipes/security_triage.json

Example answer from Deem-4B (measured 2026-10-01 on the public Space, reasoning auto; no answer fell below the fitted thresholds, so none was reasoned): severity score 3.29, most likely level 4 (Critical, 0.50); action contact_user 0.68 (revoke_sessions 0.14, escalate 0.11); mfa_fatigue_possible yes 0.65.

NPC decisions

Let a villager choose the morning job. The villagers in Emberwick decide this way.

  • action (Choice): What should the villager do this morning?
  • share_food (Noul): Should the villager share food with others?
  • urgency (Score): How urgent is the first task?
jevoss ask recipes/npc_decision.json

Example answer from Deem-4B (measured 2026-10-01 on the public Space, reasoning off): action harvest 0.71 (help_mill 0.12, sow 0.11); share_food yes 0.28; urgency score 1.65, most likely level 2 (Right now, 0.68).