Crypto · Shadow test

AI exit warnings: Claude

The Jev experiment's design, unchanged, with one difference: the answering model is Claude (claude-opus-5-5), asked through Boostio's existing Claude subscription. It is a separate experiment with its own record; it never stands in for Jev.

Collecting

13 timely answers from claude-opus-5-5 on 3 Wave paper positions so far, 8 of them warnings. Too early for a verdict.

Pre-registered as CLAUDE-WEAKEN-1. From the test's own records with a 5-minute delay, as of 2026-10-23 03:14 UTC. Paper only, not investment advice.

Does Claude's "this trend is weakening" warning prove more useful than a simple rule, after costs and on time?

Timely answers
13
Warnings
8
Control warnings (same observations)
7
Needed for a verdict
30 warned trades
Where it stands

The test, stage by stage

  1. RegisteredDoneOct 19, 2026The Jev experiment's design, with the model pinned to claude-opus-5-5, written before the first question.
  2. Historical testNot usedNever: past states may sit inside the model's training data.
  3. Forward recordNowOct 19, 2026From the 2026-10-18 daily close, decided after registration; 13 timely answers so far.
  4. First readAheadAt 30 closed warned trades.
  5. DecisionAhead
Results so far

Nothing to show yet, and why

No result yet. No verdict before 30 closed positions with timely warnings; the record grows with every daily close.
How it works

The rules, as registered

1

After each final daily close (runs at 01:40 and 04:40 UTC), build one state per open Wave paper long: the last 30 daily bars and indicators, computed in code.

2

Claude answers the two frozen questions about that state: how likely the trend is weakening, and which of four trend levels it is in. A probability of 0.70 or more is a warning.

3

A warning counts only if it arrives within 60 minutes of the decision and acts at its answer-time price; the control rule (close below the 10-day EMA with RSI below 50) is recorded beside it.

4

A verdict needs 30 closed positions with timely warnings, an interval above zero and a win over the control.

Limits

What this test cannot tell you

  • While the decision policy is in force few Wave longs open, so the sample grows slowly; that wait is accrual, not a failure.
  • A general model's answers can change when the model changes; the answering model is stored and any other model never counts.
  • Shadow only: a warning never changes a Wave decision.
  • Paper only: no orders, no real money, and not investment advice.
Related tests

More Wave research tests

All strategies →

Wave decision policy

BTC and ETH, each held only above its 200-day average.

WAVE-POLICY-1

Setup-to-trend rotation

Buy coiled leaders; sell only when the thesis breaks.

ROTATION rot-v1

Self-learning research book

Learns which research engines to trust, against a frozen twin.

LEARN-1

Risk-balanced sizing

Same picks as the frozen twin, sized by risk instead of dollars.

RISK-ALLOC-1

Exit challengers

Two exits registered in advance, tested against the control exit.

TREND-EXIT-1

AI exit warnings: Jev

Does Jev spot a weakening trend earlier than a simple rule?

JEV-WEAKEN-1