A control run to measure whether the agents add value beyond a pre-screen

[Abhishek26-eddy]

When a cheap screen picks the top candidate for the full graph, a good result could be credited to either the agents or the screen. In my setup I also run the full graph on the lowest-ranked shortlisted name each day, and log both ratings with forward alpha against the benchmark. Over time this shows whether the agents' ratings separate outcomes beyond the screen's ranking. It could fit backtest.py as an optional control arm.

Related: the backtest already fixes data look-ahead, but running the graph on dates inside the LLM's training window can still leak hindsight through the model itself. An option to limit backtest dates to after the model's knowledge cutoff might be worth having.