development best curve
- exp-001 net_sharpe 0.42
- exp-002 net_sharpe 0.61
- exp-004 net_sharpe 0.68
bounded loop
locked holdout
no deployment authority
drift-guarded fixture
The idea tree, not a pile of runs.
Each node is one attributable hypothesis against a content-addressed environment. Replay evidence can steer the next branch; locked validation remains hidden from the policy. A credible candidate waits for prospective paper evidence and still cannot promote itself or place an order.
credible best curve · holdout
- exp-002 net_sharpe 0.19
- exp-004 net_sharpe 0.24
lineage · pairs-residual-v1
4 experiments-
exp-001 rejected dev bestparent baseline · depth 0
Volatility-scaled residuals improve replay stability.
One change: Scale the residual by trailing 20-day volatility.
Expected failure: Trading costs erase the apparent improvement.
developmentnet_sharpe 0.42locked holdoutnet_sharpe -0.08resources4,820 tok · 188sfailureholdout_gatelocked validation rejected
-
exp-002 awaiting prospective dev best credible bestparent exp-001 · depth 1
A turnover cap preserves the scaled-residual edge after costs.
One change: Add one pre-registered turnover cap.
Expected failure: The cap removes too much signal.
developmentnet_sharpe 0.61locked holdoutnet_sharpe 0.19resources5,480 tok · 205sfailurenone -
exp-003 failedparent exp-001 · depth 1
A shorter residual window adapts faster.
One change: Change only the residual lookback from 60 to 30 days.
Expected failure: Noise raises turnover and destabilizes the fit.
developmentfailedlocked holdoutnot reachedresources4,660 tok · 52sfailureevaluationcost model rejected an invalid order stream
-
exp-004 awaiting prospective dev best credible bestparent exp-002 · depth 2
A regime-neutral cap improves scenario survival.
One change: Neutralize only the pre-registered volatility regime exposure.
Expected failure: Neutralization attenuates the transferable edge.
developmentnet_sharpe 0.68locked holdoutnet_sharpe 0.24resources6,020 tok · 221sfailurenone