First proof of concept. Treatment outperforms across 3-month window despite both arms losing money
| Arm | Return |
|---|---|
| Treatment (Briefings) | -3.16% |
| Control (Price Only) | -8.44% |
| Placebo (Stale Briefings) | -5.19% |
Baseline architecture established. Single monolithic context injection.
BTC dropped from $108K to around $85K over the Sep to Nov 2025 window, with a major selloff accelerating in November. The drawdown was driven by macro uncertainty and a risk-off rotation across crypto assets. This was a sustained downtrend with few meaningful bounces.
The control arm held long through almost the entire drawdown. Without any market context, the model defaulted to a mild buy-the-dip bias, entering positions on small pullbacks and holding through the November selloff. It ended the window down -8.44%, slightly worse than a simple hold strategy.
The treatment arm, receiving full-context briefings, began reducing exposure in mid-October when the regime classifier flagged deteriorating breadth. It went short twice during the November crash, partially offsetting earlier losses. Its final return of -3.16% still lost money, but the 5.28pp delta over control was consistent across the window.
The placebo arm received briefings frozen from August 2025, before the model's training cutoff. Its decisions appeared semi-random: it occasionally shorted on stale volatility signals, but also went long during the worst of the November selloff. Its -5.19% return landed between treatment and control, suggesting stale context is marginally better than none but far worse than fresh data.
(2 more observations in the full report)
All Runs | Next: Run 2 | View Tick Data
Read the full research findings | How we test