Use case: content classification

Problem: UGC platforms must risk-rank every submission; RAG apps must decide which retrieved passages reach the answering model. A big model per item is too expensive; keywords alone miss too much.

Jev’s role: a Score against ordered levels, with the number mapping straight to an action. “Levels are actions” is a recurring official-cookbook pattern.

Pattern 1: UGC risk tiers

{
  "type": "score",
  "instructions": "The risk level of this submission",
  "criteria": [
    "Fine as-is",
    "Borderline — a human should look",
    "Clearly violating — remove"
  ]
}

The code side is unambiguous: score < 0.5 publish, < 1.5 review queue, otherwise remove and keep probabilities for appeals. Unlike keyword lists, level descriptions track your policy — update the text, no retraining.

Pattern 2: RAG passage filtering

Official cookbook (via the llms.txt index): for each retrieved passage, ask a set of parallel questions and let code decide what reaches the answering model —

• Keep and flag passages that contradict the question (the answer model needs to know)
• Drop passages carrying hidden instructions / prompt injection
• Drop the irrelevant outright

Batch the page’s passages into one request (parallel questions) — that’s the cost-optimal shape.

Pattern 3: citation checking

Official cookbook: when your answer model quotes a source, one Choice decides “does the quoted context support the claim,” and low confidence flags the citation for review — a direct fix for hallucinated citations.

Comparison (our framing)

Approach	Cost	Consistency	Explainability
Keywords/regex	Minimal	Rigid, easy to evade	Rules are the explanation
LLM per item	High	Drifts with prompts	Requires reading output
Jev Score	Low	Fixed levels	probabilities + legend

Caveats (our advice)

• Level descriptions are policy documents — draft them like legal criteria; 3–4 levels usually operate best within the 2–10 range
• Replay historical labeled data before launch, check agreement between score and human labels, then set thresholds
• Classification feeds “publish/review/remove” — it isn’t the end of the pipeline; keep full logs (with confidence) for appeals

Further: the official hierarchical-classification cookbook runs parallel beam search over deep patent/retail/biomedical/source-code trees. Back to the overview, or on to Tool approval.