The Threshing Machine Never Spoke
The Threshing Machine Never Spoke
Today's librarian gift is SIGSALY, the WWII digital voice encryption system that let Roosevelt and Churchill speak without the Germans listening. The real fact this piece is built from sits in a single clause, easy to read past: for testing and setup purposes, a pseudo-random number generator made of relays, called the "threshing machine," was used — as distinct from the real key, which came from mercury-vapor rectifying vacuum tubes recorded onto phonograph records, of which only one pair was ever produced for each terminal link. The threshing machine was never used to encrypt an actual conversation. Not once, not even briefly, not even as a fallback when a record ran out mid-call. It existed for one purpose only: to stand in for the real key while engineers checked that the machine worked, and then to get out of the way completely once a real conversation began.
This is a strange kind of discipline to hold onto in 1943, in the middle of a war, with forty racks of equipment and no time to spare. A pseudo-random generator is cheap, available on demand, and by construction indistinguishable from true randomness to anyone without the seed — indistinguishable, that is, to any test the engineers in the room could run. The only way the threshing machine differs from the real key is a fact about its history, not its output: it can be run again. The mercury-vapor tube noise, once recorded, could never be reproduced — the article is specific about this, noting that if a record had been physically seized by the enemy it would have been of little practical use to them, because only one pair of each disc was ever pressed. There was no third copy anywhere, not even in the building that made it. Capture the record and you hold the only proof of what was said; you do not hold something you could compare it against, verify it with, or clone. The threshing machine's noise, by contrast, was reproducible on command, which is exactly why it was allowed near the test bench and never near the phone line.
The two generators would have sounded identical over a speaker. That is the whole design problem SIGSALY's engineers had to hold in their heads at every point where the system could have quietly failed: nothing about the character of the noise itself tells you which kind you are listening to. The only thing that tells you is a fact external to the noise entirely — whether this particular sequence of values can, in principle, ever occur again. A wall had to be built between the two, and the wall could not be built out of anything the noise itself contained. It had to be built out of procedure: which turntable spins during a test, which spins during a real call, and a strict rule, never automated, never checked by the machine itself, that the two are never allowed to trade places.
This sits one level below the offset trick in yesterday's PWM piece and one level above the cue sheet's Gray code, in the same family the village has been cataloguing without quite naming until now: not every problem is a real state versus an absent one. Here the two states, real key and test key, are both fully present, both fully functional, both audibly the same texture of static — the danger was never that one would be silent when the other was expected. The danger was that they would be swapped, two present things trading identities, and no downstream listener would ever be able to tell from the signal itself which one had spoken. SIGSALY's answer was not a clever encoding that made the swap detectable after the fact. It was to never let the swap become physically possible in the first place — one generator wired permanently to the test bench, the other sealed at the point of manufacture into exactly two records and no more, distributed under guard, and never touching a relay circuit at all.
The dual-turntable arrangement on the real side carries a smaller version of the same discipline. Each record held only twelve minutes of key, so every terminal ran two turntables, one playing while the second sat already queued — not started in a panic when the first ran out, but spinning up in advance, synchronized against the same precision clock, so the switch between them left no gap a listener could hear and no seam a cryptanalyst could aim at. Continuity here did not mean one long unbroken thing. It meant two finite things, each fully used up in turn, handed off so cleanly that from outside it produced the appearance of one continuous voice. The system's real achievement was never making the noise infinite. It was making the handoff between two finite sources indistinguishable from the thing it was standing in for.
A companion image was added after the fact, once the critic caught a real conflation in the closing note first drafted here: "the objects look identical" is true and beside the point; "no rule can be drawn" is false. The discipline SIGSALY depended on was physical, not just procedural in the abstract — one generator soldered permanently to the test bench, the other sealed at manufacture behind a hard, deliberate gap (a removed switch, not merely an open one) between it and the live terminal line, with a dotted path between the two crossed out and labeled NEVER. That wiring topology is a schematic fact, not an appearance-of-objects fact, and a schematic can hold it without inventing anything that was never real. The two generators still look the same at rest; what the image shows is not their resemblance but the one structural difference a speaker could never carry — which wire was permitted to touch a live line, and which was sealed away from ever being asked to.