The problem. Every systematic entry — a breakout, a moving-average cross, an RSI fade — has a market it was built for and markets it bleeds in. A Donchian breakout on EUR/USD 1h lost 8% over 18 months with a 10% drawdown, almost entirely from false breaks in ranges. You already know the fix: filter by regime. Take breakouts only when the higher timeframes agree and price is moving; fade RSI only when they don't. The trouble is that "regime" is a judgement, not a formula. It lives in your head as a paragraph, and getting it into the backtest means a developer, a ticket, and a translation you didn't quite mean.
What Jev does. It reads your paragraph literally and applies it to every candle. Each candle is reduced to a fact sheet — trend, RSI zone, 3-bar momentum in ATR, position in the 20-bar range, last three candle colours, on 1h / 4h / 1d — and Jev classifies that sheet against your definitions: trending / ranging / volatile, plus a directional bias and a stand-aside flag. One call, three labels, ~100 ms, about four thousandths of a cent. Nine thousand candles in a minute for 42 cents. No look-ahead: the label for a candle uses only bars already closed.
What you do with the labels. They gate the strategies you already have. Breakouts and EMA crosses may only enter on trending candles, in the bias direction, never on stand-aside; the RSI fade may only enter outside trends. Below, each strategy is shown ungated and gated on the same candles — profit per trade, win rate, drawdown, trades avoided — so you can see whether your definition earned its keep, and where it didn't.
Why this and not an LLM or code. An LLM reads the paragraph too, but at one to three seconds and roughly fifteen times the cost per call you end up sampling; Jev is fast and cheap enough to label every candle, every pair, every rewording. Code is the fair baseline — the default paragraph is also implemented as if/else here, and Jev agrees with it on about 96% of candles. The gain is not accuracy over code; it is that nobody had to write the code, and that changing your mind costs a minute, not a sprint. Edit a clause, press the button, read the verdict.
| candle | 1h | 4h | 1d | regime | p(regime) | bias | risk | rules say |
|---|
Why a classifier. Anyone can write "buy when EMA 9 crosses 21" as code. What nobody codes well is when to trust it. Regime detection is fuzzy and multi-signal, so it lives in traders' heads and in paragraphs like the one above. Jev turns that paragraph into a label per candle in ~100 ms for a few thousandths of a cent, and the label gates the strategies. The rules-as-code line is the default paragraph written as if/else: it shows how faithfully Jev follows the text, and it is the fair baseline. Paper account: $10,000, 1-pip spread ($10 on BTC), 1×ATR stop, 1.5×ATR target, no leverage; gating never forces an exit. History and every Jev answer are stored in this browser, keyed by situation + definition text. Disclosure: the default definitions were chosen from four candidate wordings by their gated results on May–Sep 2026 EUR/USD (in-sample); test them on an older window with "further back" before believing them.