| CAGR % | Vol % | Sharpe | Max DD % | trades | held | turnover % | cost drag % | avg SPY % | views | |
|---|---|---|---|---|---|---|---|---|---|---|
| Fixed 60/40 | 8.30 | 10.73 | 0.64 | -31.41 | 273 | 0 | 285.60 | 0.14 | 60.10 | |
| Prior only + trade policy | 8.30 | 10.73 | 0.64 | -31.43 | 59 | 214 | 172.80 | 0.09 | 60.30 | |
| Rate-shock rule views | 8.24 | 10.45 | 0.65 | -28.99 | 122 | 151 | 926.30 | 0.46 | 59.60 | |
| Masked LLM replay | 8.08 | 10.63 | 0.62 | -28.43 | 161 | 112 | 1,164.80 | 0.58 | 66.10 | 273.00 |
| Masked LLM replay, gated | 8.24 | 10.62 | 0.64 | -28.81 | 165 | 108 | 1,162.50 | 0.58 | 66.10 | 273.00 |
All strategies share dates, prices, covariance windows, rebalance calendar and a flat one-way transaction cost. "Prior only" and "Logged LLM views" run through the same allocation code as the live dashboard. The rate-shock rule is a placeholder signal, not a validated one. Historical LLM views cannot be backtested honestly here: there is no point-in-time news archive and the model has read the history it would forecast. The logged-LLM line is a forward test that starts on the first live run and falls back to the prior before that.