What I was looking for
Whether the void gate's edge on one session was real, or the only session I had looked at.
What happened
On session D, a gate built on liquidity voids beat the ungated version with t ≈ 2.3 over six cycles. That looks like an edge. It also came from the only session I had already examined, which is exactly how a false edge looks.
So before replaying session E, I committed the test: three hypotheses (the gate beats an ungated control with t above 2; it is positive at retail fees; the sign is the same in D and E), the variants, the exact commands, and the exits. The control used identical exits, so the gate was the only difference.
All three hypotheses failed. At retail fees a maker fill costs about 280 parts per million, and the gate did not earn that back.
I published the engine, the study, and an evidence page that leads with the negative result. The recorded sessions stay private, because exchange terms forbid redistributing market data, but their SHA-256 checksums are public and anyone can record their own sessions with lre live-record. Order routing is parked until an edge passes a registered test.
A published negative result costs a little pride. It buys something more useful: anyone reading my other numbers knows I report the ones that went against me.