Up.
Draw an imaginary line (or a real one) connecting the assumed acoustic centers of the two drivers. That establishes the "tilt," as the physical axis is perpendicular to it. Differential driver phase and delay then alter the angle.
Textbook, when it's "right," the forward axis is normal to the baffle at the midpoint between the sources, but that locates it below the waveguide axis, so we tilt it upward a bit such that the forward plane and waveguide axis coincide at a design listening distance. See the measurements of Markus's Geddes Nathans in the "Horn vs. Waveguide" thread on DIY audio. The alignment is textbook, so the optimum listening height is not the waveguide axis or above.
You can see from my polar map experiments that the -6 dB "window" bounded by the influence of the vertical nulls is only +/- 10 - 15°, so it's important to optimally steer the plane. That's fairly easily accomplished using active delay, but not so easy passively, relying on filter phase to make it happen, as that also alters the frequency response.
It's easier, and the result better, if the acoustic centers are approximately in line from the outset, which is why I'm rear-mounting the woofers in the trap cabs. In 4425 and 4430, JBL moved the horn forward to bring the drivers' acoustic centers into physical alignment.... :yes: