AI exit warnings: Jev

• TypeSafe describes Jev as a "System One" model: it does not chat or write text. Given a state and typed questions it returns structured answers. Here the state is one coin's last 30 final daily bars with indicators computed in code, and the answers are a value from 0 to 1 for "this trend is weakening" and a position on a four-level trend scale.
• Its role is a shadow warning light. A value of 0.70 or more is a warning, stored with its time and the price at that moment, then compared with what the Wave position really did. It never closes a position, changes a Wave rule or places an order.
• It sees public market data only: the final daily bars, the indicators and Boostio's BTC market-regime label. No position size, entry price, profit or loss, member or personal data is sent.
• A value is not a probability of profit. TypeSafe describes Jev's values as calibrated probabilities for the question asked; whether a warning helps a trade is what this test measures, after costs, against the simple rule.
• The version is frozen: answers must come from jev-1.13.0, the version pinned when the test was registered. TypeSafe released Jev 1.14 on October 14, 2026; a newer version would be a new test, and an answer from any other model is stored apart and never counts.
• Usage is capped: TypeSafe charges per input token (USD 0.042 per million for this version; output is free). The job sends at most 20 questions a run and 200,000 input tokens a day, with no automatic retries, so a full day costs less than one cent.
• Boostio is not affiliated with or endorsed by TypeSafe.

1. ✓RegisteredDoneOct 16, 2026Question, model version, control rule and timing written before any answer.
2. ✓Historical testDoneControl rule replayed on 67 past observations; never AI evidence.
3. 3Forward recordNowOct 15, 2026Answering since October 21, 2026; 4 timely answers so far.
4. 4First readAheadAt 30 closed warned trades.
5. 5DecisionAhead

4 timely and 18 backfilled, answering since Oct 21, 2026. How often the two agree says nothing about whether either helps: only closed positions, priced after costs at each warning's own time, can show that.

For context only: the control rule replayed on 67 past observations of closed Wave positions fired 0 times before their exits. That replay is never evidence about the AI model.

• Only timely answers count: an answer must arrive within 60 minutes of its decision. The first answers replied to decisions recorded before the key was set up, so they are shown as backfilled and never count.
• Small samples decide nothing: a verdict needs 30 closed positions with timely warnings, an interval above zero and a win over the simple rule.
• The same frozen question runs on Claude as a separate experiment; its answers are never counted as Jev's.
• Shadow only: a warning never changes a Wave decision.
• Paper only: no orders, no real money, and not investment advice.