Jev Wiki Consult Guide

Whether Jev fits a given project. An agent reads the project, lists its decisions, applies a fit test, matches surviving decisions to pattern pages, and returns a ranked shortlist with a primitive and criteria sketch, estimated cost, caveats, and what not to use Jev for.

Typically 30-50k tokens. The page measured 31k for a shortlist and 47k when regulated data added the legal and alternatives pages, and it explicitly sets no token ceiling: the stop rule ends the reading, not a budget.

Ones whose output is a known set, a yes or no, or a rubric rating; ones needing semantic understanding; ones that happen often or sit on a latency-critical path; ones where low confidence can change behaviour; or ones that replace an LLM call parsed into a branch.

Compute them in code. The fit test marks number thresholds, counting, and date comparisons as poor fits and says to ask Jev only the semantic part of the question.

No. The fit table says to skip decisions asked once, because setup is only repaid by repetition. It also keeps existing simpler tools when a rule, cache, embedding, or owned classifier already works.

It groups every question about the same item into one call, prices input at $0.042 per million tokens with free output, and shows a worked example of about $1.70 a day at 40,000 calls a day.

The page calls the failure check mandatory: read the confidence failure reports, which note that on JevBench v1.4.2's sealed set Jev's highest-confidence answers were right less than half the time, and make validating the cut on the user's own labels part of the recommendation.