Jev Playground — jyeswak

A measurement lab for typed classifiers inside a coding-agent fleet: which file to read, whether a tool result is an injection, whether a command is risky, whether a recalled memory belongs in context, whether a reviewer's finding is new.

Jev is one technique among several, not the frame. The README says everything built before October 2026 is the foundation, and each decision question goes through a fixed evaluation arc with a committed budget before any model call.

Labels first: a blind-labelled set before any model sees it. Then the constant and the cheapest rule set the bar — a model has to beat both on the same rows. The bar, N, model id and spend cap are committed before the first call, and every backend is calibrated before comparison.

Simple rules often win: keyword and regex rules beat a typed model on tool-call risk triage, phishing, review novelty and verdict checks, and a deterministic gate replaced a model pre-gate at 98.1% agreement with zero calls.

Where the label is in the text and lexical cues break: passage selection scored FiQA top-1 76.2% versus BM25's 39.3%, plus injection on a public corpus and out-of-distribution spam. They lost forecasting with 0 wins out of 12.

Value, not accuracy. Each live surface carries a benefit-minus-cost receipt and has to fire on a positive and stay silent on a planted negative in a real session; a negative receipt turns it off.

It is the model-neutral CLI that winning techniques ship through, per the README. The plan of record is ROADMAP.md and the work queue is a bead graph in .beads/, but the excerpt truncates mid-table, so this entry does not document its flags.