Because the cavity height is not set by the router. It is set by the layers left in place under the window and by how much bondply flowed into that window during the press, so the dimension that decides the resonant frequency comes out of lamination. On a 1.52 mm cavity, 0.04 mm of extra height is 2.6 percent, and at 10 GHz that is 260 MHz of shift.
That is the number worth arguing about with a fabricator, not the milling tolerance. A drawing that asks for plus or minus 0.05 mm of depth and gets it can still produce a cavity 3 percent off nominal, because the depth was measured from the outer surface of a stack whose cores were never at nominal thickness.
Where the depth actually comes from
A window type cavity is built, not cut. Windows are routed into the cores that will form the walls, the cores are stacked with bondply, and the press bonds the stack into one board. The floor of the cavity is the top of the last core that was not windowed, so the height is the thickness of everything below the window, not a machining dimension.
The bondply is thin and it moves. Rogers 2929, a common choice for PTFE and hydrocarbon stacks, comes in 0.038 mm, 0.051 mm and 0.076 mm sheets, each plus or minus 10 percent, with Dk 2.94 and Df 0.003 at 10 GHz. Under pressure the resin flows, and the fabricator compensates by cutting the window back so that the flow lands on the wall instead of in the opening. Some flow into the cavity is wanted, because without it the wall does not bond. IPC-6012 Revision F and IPC-6018 allow adhesive flow of up to 0.75 mm onto conductor features at the cavity wall, and that is a bonding allowance, not a dimensional one.
IPC-6012 Revision F also sets the default depth tolerance for a cavity at plus or minus 0.005 in, which is 0.127 mm, measured from the outer layer starting point rather than from a particular dielectric layer. On a 1.52 mm cavity that default is 8.4 percent of the height. It is a sensible number for a pocket that houses a connector. It is not a number you can put on a resonator.
What a hundredth of a millimetre of height is worth
For a cavity backed structure the resonant frequency scales roughly with 1/h and with 1 over the square root of Dk. Put those together and the tolerance budget writes itself. On the same 1.52 mm cavity: 0.127 mm is 8.4 percent, or 840 MHz at 10 GHz. 0.04 mm is 2.6 percent, or 260 MHz. 0.010 mm is 0.66 percent, or 66 MHz. Bondply that presses from 0.038 mm to 0.030 mm is 0.5 percent, or 53 MHz. A Dk that lands 0.05 high moves the frequency 0.84 percent, or 84 MHz. A wall that moves 0.05 mm inward on a 10 mm window is 1 percent, or 100 MHz.
So a customer who needs the cavity within 1 percent of nominal has to fix the stack, the bondply grade and the wall position together. Tightening the milling tolerance does nothing, because the error is not in the mill. This is the same arithmetic that decides whether a multi cycle press is worth the yield it costs.
Three things that go wrong on the order
The first is a depth called out as a machining tolerance with no stack reference. The shop mills from the outer surface, hits its 0.05 mm, and the cavity is still off because the core below it was 5 percent thicker than nominal. The depth has to be derived from nominated layers with their own tolerances, and the quotation has to state what that arithmetic gives.
The second is accepting the 0.75 mm adhesive flow allowance on a small cavity. On a 2.5 mm opening, 0.75 mm of resin at the wall is 30 percent of the floor. It is inside the standard and outside the application, and the shift it causes is well beyond what the customer modelled.
The third is checking depth with a gauge on the finished panel. The gauge sits on the resin fillet at the wall instead of on the floor at the centre, so it reads shallow and the board is scrapped, or reads acceptable while the centre is low. A cross-section at three points plus a CT scan answers the question in an hour, and that is also how an impedance and plating report should be read.
How we build a cavity
Minimum cavity and step 2.5 mm, layer to layer registration 0.025 to 0.06 mm, up to seven press cycles, board thickness 0.2 to 12 mm, and hybrid RF and FR-4 stacks where the digital layers need FR-4 and the RF layers need PTFE or hydrocarbon. Materials and the frequency window they cover are on RF and microwave PCB build, and the general window sits on PCB capability.
Before the first panel we take the stack apart with the customer: which cores form the floor, what each contributes at nominal and at its tolerance, which bondply grade has flow data for that opening, and how much cutback the flow curve calls for. First article is a cross-section through the cavity at three points, a CT scan for the walls, and TDR on the coupons. If the requested depth tolerance is tighter than the stack supports, we say so at that point, not after the panels are built.
Next step
Send the stack with the cavity zone marked, the frequency the cavity has to hold, and the tolerance you can live with. We will come back with the depth the stack actually gives and the bondply that holds it, before anything is committed to a press.