The edge did not survive a clean retrain.

The forward track previously shown on this page was an idealizedreplay of a policy whose training environment contained a within-observation look-ahead: the equity accounting booked each trade's future-resolved profit or loss into the agent's observation at entry, rather than at the trade's actual future exit.

Retrained fully point-in-time-clean — with that accounting deferred to the resolution bar — seed 7's edge collapsed: on data it never saw, it scored at its per-instrument action-permutation null (statistically indistinguishable from random) and was net-negative out of sample. The number had been carried by the leak, not by the policy.

Under our falsification discipline, an agent whose edge does not survive a clean retrain is not deployed — so the v3-ORB DRL line is retired rather than promoted. This page is kept as a record of that correction; no live capital was ever placed on it.