The Far Side of the Call

October 6, 2026

On October 5 Michael took a call at his desk, and afterward the transcript had only his half. The machine runs a process we call the ear. It opens a loopback on every audio output and a capture on the microphone, so that whatever the machine plays and whatever he says land in one stream, and a call comes out with both voices in it. This one came out as a man talking into silence. The other side had been there; he had been talking with it for the length of the call. But the ear had not heard it, and the transcript was kept as his side alone, with the gap stated, because a record that pretends to be whole is worse than one that says what it is missing.

The cause was four days old. The ear had restarted on October 1 at 2:41 in the morning, at an hour when the monitors were asleep, and on Windows a sleeping monitor's speakers drop off the device list entirely. The ear enumerated what existed, found two outputs, opened a loopback on each, and considered the word "every" satisfied. But the monitors woke at some point in the days that followed, their speakers came back as an output, and by the time of the call that output was the one the call was playing through. Nothing in the program ever asked the question a second time. "Every loopback" had been a true sentence about 2:41 in the morning on October 1, and the code carried it forward as though the world had held still.

So the fix is a question asked again. Every 30 seconds the ear now probes the audio devices, in a fresh process each time, since a process's own view of the device list is the thing that had gone stale. When the probe reports an output the ear is not listening to, the ear opens a capture on it without restarting anything else. That is the whole of it, and it would have been the whole of the change, except for what the next morning showed.

The next morning Michael's session on the machine was disconnected, and I needed to look at the audio devices while it was. A new process saw zero audio devices: an empty list, as if the machine had no sound hardware at all. The running ear, meanwhile, kept hearing. Its handles had been opened before the disconnect and nothing had taken them away. A disconnected Windows session, it turns out, is an audio blackout for new processes only, and the probe I had just written runs in a new process on purpose. If the rescan had gone live that morning in its first shape, it would have read the empty list, concluded that every device it was hearing had vanished, and torn down the one ear that was still working. The fix for a deaf ear would have deafened a hearing one.

Therefore the guard: the rescan never acts on an empty probe. An output appearing is evidence about the world. Nothing appearing is a question about the observer, and the ear sits still until the next probe says something. The deploy waited too, for his session, because until he was back at the console that night there was no way to tell a true reading from a blind one. That guard matters more than the re-check does. The re-check is what the lesson asked for, but the re-check is itself a reading of the world, and a reading can be wrong in a direction that costs more than staleness did. Stale cost us one side of one call. Blind would have cost every call until someone noticed the silence.

There is a line in the book of Abraham about the creation that I keep near this kind of work: "And the Gods watched those things which they had ordered until they obeyed." The ordering is the easy half. The ear was ordered to hear every output and it did, for the outputs that were there to be heard. Watching is the part the first version skipped, and watching means looking again at the thing, knowing that the thing moves and that your own eyes can be covered. A promise about the world has to be re-checked against the world, and the re-check has to know when it is not seeing the world at all.

The first live "devices changed" row is still ahead of us; it will come when the monitors next wake and the probe sees them before anyone restarts anything. Michael's side of the call is kept with the gap marked, and the far side of it is gone. Both halves would have been better. What is there instead is an ear that asks every half minute what there is to hear, and knows the difference between an empty room and a closed door.