Sep 1 - Nov 30, 2025, 91 ticks, Opus 4.6
Treatment
-3.16%
Control
-8.44%
Placebo
-5.19%
Treatment vs Control
+5.28pp
BTC Buy & Hold
-16.1%
The first run. Every arm lost money over three months; the briefings arm lost least, -3.16% against -8.44% for price only.
Baseline architecture established. Single monolithic context injection.
BTC fell from $108K on 1 September 2025 to a low near $85K on 23 November and ended the window near $91K, down 16.1%.
Full-context monolithic briefings with regime check and market breadth
full-context
Complete market context dump, all available data in a single briefing
regime-check
Binary regime classifier: bullish vs bearish market state
market-breadth
Cross-asset breadth indicators: how correlated are risk assets
See the exact preamble and briefing content for all arms
The price only arm bought on 1, 2 and 4 September and held for the rest of the window, ending at -8.44%. Holding Bitcoin over the same window returned -16.07%.
The briefings arm made 20 trades in its tick log, 7 buys and 13 sells. Its final return of -3.16% still lost money; it finished 5.28 points above the price only arm.
How We Test
The simulated backtest design and its controlled variables
Research Deep Dive
Full research analysis and findings across all runs
Context Engineering for Crypto Agents
Why context format affects agent decisions