Cable wind-up is the standing hazard, and it is slow. It does not announce itself until the cable is tight, and then it takes the cable, or the horn, or both. STOP preempts a move in progress and is live whenever the page is connected, which is why it sits in the tab header and not three scrolls down a settings panel. Confirm on the EMControl front panel that the axis is in non-continuous mode before any scan with a cable routed through the tower — continuous mode ignores the software limits entirely. Check the latched error line in the Turntable section while you are there; a fault the card is still holding from last period is worth knowing about before you commit 180 angles to it. Device Emulation on the front panel is not simulation and does not inhibit the motor.

The port is the expensive one, and the damage spreads. Start the threads by hand and square, tighten with the wrench rather than by feel, and turn the nut and not the cable body. A cross-threaded or over-torqued port is a bench repair on an eighty-five thousand dollar instrument, and the same damage transfers to the next connector it mates with — which on this rig is a calibration module. Side load at a connector is cable stress and is the slow version of the same failure, so nothing hangs off a port.

The horn is dropped and the cable is kinked — neither wears out. Two hands on a horn, support its weight while you mate it, never hang it from its own connector; and for the cable, respect the bend radius, no tight coils, and nothing routed where somebody will stand on it.

The absorber is priced by the square foot. The pyramids shed if you brush them and a crushed tip is a permanently worse quiet zone, so nobody leans on the walls and nothing gets rested on the floor pyramids. There is no repairing a tip; there is only buying more absorber.

The RF, for the record, is milliwatts. Nothing in this chamber will hurt you. Everything in it can be hurt by you.

One instrument does both ends today. The chamber’s Copper Mountain VNA drives the source horn from port 1 and receives the AUT on port 2, so the quantity you record at every angle is \(S_{21}\) in dB: transmitted at one antenna, received at the other, ratioed against the source so the source’s own drift divides out. That is the same argument Lesson 9 made for trusting a ratioed measurement, now doing work on the range rather than at the connector.

The chain is a loop rather than a line. The VNA is both the transmitter and the receiver, the two horns are the only things between its ports, and the positioner is a serial link the acquisition software drives in step with the sweep. Nothing in the chain is free-running: every angle is commanded, then confirmed, then measured. That is what makes the run file trustworthy, and it is why the file records both the angle asked for and the angle reported.

You can have the whole dashboard on your own laptop before the period, with no instruments attached — one command brings up the service and the frontend together:

./start.sh --sim

The badge reads simulated and the service synthesizes a plausible array-factor pattern — a main lobe, decaying sidelobes, nulls floored around −45 dB — and simulates the tower slewing at a believable rate, so live position, progress and the stop button all have something real to exercise. A simulated sweep is not a pattern at all: it returns a mismatched load behind a length of line, which draws the arc a real one does on the Smith chart and has no resonance in it anywhere.

So rehearse the controls there, not the physics. Half an hour in simulation is the difference between a period spent measuring and a period spent finding buttons.

The chamber does for you what the absorber block did on a bench range, and more of it: the specular floor, wall and ceiling bounces are all absorbed rather than aimed away, which is why the chamber’s contribution shows up in your data as a floor rather than as a ripple you can locate. It does not make the range longer. The far-field criteria below are geometry, and no amount of absorber changes them.

Both antennas must be in each other’s far field, and both must see the same illumination. Set the heights with a tape, not by eye. A 5 cm height mismatch over a 3 m range is a 1° pointing error, which is small — but it is a bias, and biases do not average out over repeated sweeps.

Two of those five numbers are yours to argue about. Points sets how finely you resolve frequency: 101 across 2 to 3 GHz is a 10 MHz grid, which is plenty for a horn and far too coarse for a narrowband element. IF bandwidth is the dynamic-range knob — narrowing it lowers the noise floor by about 10 dB per decade and costs you sweep time on every one of your angles, which is where a 72-angle run turns into a long one. Power is the other floor knob and the one with a limit: the amplifier chain and the receiver both have to stay linear.

The Calibrate… button opens a wizard, not a calibration. Running one means somebody walks into the chamber, unmates the horn and the AUT from the cable ends, mates the calibration module across them, and comes back out — so the modal states the sweep it will calibrate at, asks which reference plane you are calibrating, takes an acknowledgement, and reminds you to put the antennas back. That dropdown is Lesson 9’s reference-plane argument turned into a required field: nothing in the data afterwards distinguishes the cable ends from the instrument’s front panel, and the difference is every dB of cable loss in your pattern.

Nothing on this dashboard shows a live trace while the tower is standing still. The VNA tab draws the sweep at the last measured angle, which during a scan is a running picture and between scans is history — so peaking up is jog, capture, read, repeat. Capture S11 + S21 while you do it and the same two sweeps tell you the antenna is still matched and the link is still there.

That is also your through check. Two minutes of it before you start is worth the time: a peak 20 dB lower than last period means a cable, a connector or a horn, and finding that out now costs two minutes rather than the length of a 180-angle scan.

Defining zero does not move anything — it renames where the axis is standing. The run file records both the angle the software commanded and the angle the card reported, and the difference between those two columns is your positioning error, available afterwards without any extra work. If they disagree by more than a small fraction of your step size, say so in the report rather than quietly using the commanded column.

A full circle asked for as 0 to 360 would measure the start angle twice, so the grid stops one step short — 0 to 355 in 5° is 72 angles, and 0 to 358 in 2° is 180. Going all the way round is worth the time on this rig: the back half of the pattern is where the front-to-back ratio lives, and a ±180° cut taken as two halves has a seam in it.

The capture menu has no \(S_{21}\)-only entry — the four options are all four, \(S_{11}\) only, \(S_{11} + S_{21}\), and \(S_{22}\) only — so the pair costs two sweeps, and the second one hands you the terminated port’s own reflection as a free check that the termination is actually a termination.

There are two floors here and Lesson 9 named both. The instrument floor is the one you just measured. The chamber contributes the other, and you cannot measure it this way: with the AUT unmated nothing radiates, so the stray field has nothing to scatter. It shows up instead in the back half of a real cut, as a level that will not go deeper no matter how much you narrow the IF bandwidth. Whichever of the two is higher is the one you actually have.

Narrowing the IF bandwidth by a decade is the cheapest 10 dB in the building and is worth doing once as an experiment: take the floor run at two settings and watch it drop. Do it before you calibrate, because changing the IF bandwidth is exactly the change the calibration panel will warn you about.

Name every run in the Output section before you start it. Left empty the service stamps the date and time, which is enough to keep the runs apart and useless for telling you which is which a week later. l11_eplane_aut, l11_eplane_repeat, l11_hplane_aut, l11_ref_horn costs you four seconds each and is the difference between a reduction session and an archaeology session. Anything in the Stored runs list reloads straight back into the plots, so a run you named is a run you can still use.

The file gives you more than the dashboard plots. Both angle columns are there, so positioning error is recoverable; the real and imaginary parts are there, so phase is too, which matters in Module 3 when the thing you are measuring is an array and the interesting quantity is the phase across it. Reduce it in Python, one csv read and a reshape — the grid is angles \(\times\) frequencies and it is complete, with every angle present at every frequency.

Worked example, one E-plane cut. The peak reads −35.6 dB at 0°, so the −3 dB level is −38.6 dB. Interpolating between samples puts the crossings at −19.8° and +20.2°:

The first sidelobe peaks at −51.4 dB, and the level at 180° is −54.0 dB:

Cross-check the beamwidths against the gain: with an H-plane HPBW of 42°, \(26000/(40 \times 42) = 15.5\), or 11.9 dBi.

The comparison is clamped 30 dB below the peak before differencing, which matters more than it sounds. Nulls are where a model and a measurement disagree most and where the disagreement means least — a null one degree off its predicted angle differences to tens of dB against a neighboring lobe. Un-clamped, the RMS would report null alignment rather than pattern agreement.

Parsing happens in the browser, so nothing is uploaded and the service never learns a comparison is running. If the RMS comes back large, check Rotate before you check the physics: a mount offset is a rigid shift of the whole curve and it is the most common reason a good model disagrees with a good measurement.

A fifth entry belongs in the budget on any run the panel flags, and it is not a small one. A calibration is only meaningful at the sweep it was taken at, and a run set up at a different span, point count, IF bandwidth or power is flagged in the VNA section and warned about in the log. Correction interpolated across a span it never measured is a plausible-looking answer of unknown quality, and “the dashboard warned me and I scanned anyway” is not a sentence you want in a report.

In detail:

1. Two principal-plane cuts (E-plane and H-plane), plotted in polar dB, normalized to the peak, with the angle convention and the cut frequency labeled on each.

2. An extracted table — HPBW, first sidelobe level, front-to-back ratio, gain, XPD — with an uncertainty estimate on every row. Your repeated scan from step 12 is the basis for most of them.

3. Your measured floor, stated as a level and as a dynamic range, with one sentence per table row saying whether that number clears the floor and by how much.

4. A comparison against prediction — your Lesson 8 model overlaid, with the RMS deviation quoted, plus the aperture-formula gain and the \(26000/(\theta_E \theta_H)\) beamwidth cross-check — with every discrepancy larger than your uncertainty named and explained.

Quote the run names for every figure, and state the calibration and mode your runs were taken under; both are written into each run’s meta.json and neither is recoverable afterwards from the pattern alone. A 2 dB gap with a named cause is a better report than a 0.2 dB gap with no discussion.

Lessons 12 to 14 go back to antenna families — loops and monopoles, patches and slots and horns, reflectors and Yagis — and every gain figure in them is now a claim you know how to check. Several of those antennas are defensible choices for your project, and you will be reading their published patterns with a measurer’s eye rather than a reader’s.