The continuity test passed. The link did not. 

Two assembly weaknesses at one coaxial interface, exposed by loading the cable in the direction that mattered 

The failure showed up as image corruption — horizontal bands across the display, distorted color, and in the worst cases repeated flashing or an image that turned almost entirely green.

It did not show up in ordinary use. The cable could be moved, repositioned and bent through its normal working range with no effect on the image. It only appeared when the cable was pushed into a larger bend and then rotated through different loading directions, and even then only near one narrow band of orientations. Everywhere else the video stayed clean.

That pattern was the first useful clue. The failure was intermittent, but it was not random. Direction and load mattered.

Static continuity lowered suspicion too early

Continuity was the obvious check, and continuity passed. No persistent open circuit, the assembly electrically connected, the system running normally while stationary.

That result moved the interconnect down my list for longer than it should have. A passing continuity test feels like an answer, and it is easy to treat it as one.

What it actually established was narrow: that a DC path existed between two test points while nothing was moving. It said nothing about whether the center contact and the shield return held their condition while directional load was being transferred through the cable. The meter was right. My question had been narrower than the failure.

Making the condition repeatable

I held everything else in the setup constant — the same device, the same host, the same cable assemblies, the same software build, the same power and operating settings — and varied only the cable loading.

Normal handling did not reproduce the failure. A deliberate large-angle bend, followed by a slow rotational sweep, revealed a sensitive orientation that could be reproduced repeatedly.

The loading was applied by hand, not through an instrumented force-and-displacement fixture, so none of this defines a threshold in newtons or degrees. I wasn’t looking for a number. I was looking for a condition I could reach deliberately, and once the orientation was known, the failure could be produced on demand.

That changed what the rest of the investigation could do. From this point on, every measurement had a before and an after.

What the software could and couldn't tell me

With the corruption on screen, the software log recorded recurring frame-input errors and lost image data. Useful: valid image frames genuinely were not being received, so this was not a display-side rendering artifact.

I had expected the log to narrow the location, and it didn’t. In hindsight it was never going to. The same entry would have been written whether the problem started at the image source, a board-level interface, the cable, the connector, the shield path, the receiver, or the software stack itself. The log confirmed the effect, but it did not localize the failure.

A capture that only had to do one job

I then set up a long-record oscilloscope capture at a monitor point associated with the affected channel, using identical acquisition settings under both loading conditions.

This was never a signal-integrity measurement, and it should not be read as one. The timebase and sample rate resolve an activity envelope across hundreds of milliseconds — they cannot resolve bit-level behavior on a multi-gigabit link. Nothing here quantifies eye opening, jitter, impedance, instantaneous contact resistance or protocol compliance, and the record I kept does not document probe configuration and bandwidth well enough to reconstruct those afterward.

What it could do was compare two conditions fairly. Under normal handling the recorded activity stayed uniform across the whole record. Under the critical directional loading condition, a transient change appeared.

The scope and the software log did not share a time base, so I can’t claim the events were synchronized to a common clock. What I can say is that the transient change, the frame errors and the visible corruption all appeared under the same directional loading condition — three observations from three different disciplines, all keyed to one mechanical trigger.

That moved the interconnect back to the top of the suspect list. It did not by itself rule out the board-side terminations, the receiver interface, or any other electrical boundary that the same load path could reach. Inspection had to determine which part of the interface was responding to the load.

What the assembly actually looked like

Inspection found two weaknesses at the same coaxial interface. They came from different parts of the build process, and nothing in the evidence suggests one produced the other. What they shared was a load path: near the sensitive orientation, the same external load reached both.

The first was about when the contacts were fixed.

The coaxial contacts inside the connector are designed with a limited amount of float. That movement isn’t looseness — it exists so the mated pair can absorb small positional differences from component tolerance, connector placement and assembly variation.

In the original build sequence, the solder-side terminations were bonded before the connectors had been mated. That locked the contacts around whatever position they happened to occupy at the time. When the connectors were then mated, the float that should have absorbed the alignment difference had already been spent, and residual offset and lateral preload stayed in the interface.

At rest and under ordinary cable movement, the remaining contact margin was sufficient. Under a larger bend in one particular direction, external load was arriving at an interface that was already carrying an offset.

The second was about the shield.

The cable braid did not form a stable circumferential connection to the connector shell. A continuity check can still pass through a small or localized contact area, which is part of why this survived earlier testing.

But a coaxial interface at these speeds depends on more than the existence of a DC path. The return structure has to hold its geometry and its contact condition while load is being applied. Inspection showed that the original termination lacked the circumferential mechanical retention needed to keep that boundary stable under directional loading.

Together, the two weaknesses gave an assembly-level explanation consistent with the behavior that had been so difficult to place: an assembly that passes continuity, works at rest, tolerates ordinary handling, and fails near one direction.

The corrections

For the contacts, we reversed the assembly sequence. The connectors are now fully mated first, so the floating coaxial contacts settle into the position the mating interface actually establishes; the solder-side terminations are then bonded and cured with the connectors still mated. The contacts end up as rigid as before, but fixed around the real mating axis rather than an arbitrary pre-mating one. The problem was not rigid fixation itself, but applying it before mating alignment had been established.

For the shield, the termination was redesigned to give a mechanically retained circumferential connection between braid and shell, with strain isolation so that external cable movement is less able to reach the shield boundary. The intent was not to establish DC continuity — that already existed — but to maintain the interface geometry under load. 

Going back to the condition that broke it

A corrected assembly was not accepted merely because the image looked fine while sitting still.  It went back to the same large-angle bend, the same directional sweep, the same orientation.

Before the correction, that sequence produced a transient change in the recorded activity, recurring frame errors and visible corruption. Afterward, the same symptoms did not recur during repeated sweeps of the corrected assemblies tested.

Both changes were made and verified together, so this does not separate how much of the original symptom belonged to contact alignment and how much to the shield termination. Neither correction has been shown to be sufficient on its own, and I wouldn’t claim otherwise. What the regression established is narrower: with both in place, a previously repeatable direction-sensitive failure stopped recurring.

That was sufficient to support closure of this failure mode at the assembly level within the configurations tested. It is not product qualification, and it is not an instrumented flex program. It answers one question — whether the direction sensitivity was still there — and it answers it by returning to the condition that caused the trouble rather than by avoiding it.

Everything measured at rest had been measured correctly. The interface just wasn’t at rest when it failed.