>
had not figured out the diodes were back-to-back.
The drawing uses a conventional double-diode symbol, then a lot of criss-cross wires, so it isn't clear what's really happening.
Start top-right of dual-diode. That plate is connected to other cathode. Other plate is connected through variable-battery to first cathode. It's a back-2-back loop.
It appears that there is a DC voltage in series with one diode; but C2 makes that meaningless. There is some start-up bias-shift but C2-R2 settle quickly.
I should not have said "perfectly symmetrical". The left diode has R10 and battery in series. But the path impedance is up around 12.5K (25K||25K), the battery is a low audio impedance (couple ohms), and worst-case pot impedance is 250 ohms. So worst-case asymmetry is 2%. Perhaps annoying for measurement, trivial for music-trimming, and quite insignificant for the probable speech+static problem it was probably meant to solve.
The 1K is apparently a compromise between asymmetry and battery life. And perhaps an "academic" value which works long enough for the grad-assistant to make some tests, but will flatten the battery in a few hours. A proof of concept, not a Product.
It applies up to 32dB NFB around the amp. There are at least three strong bass-cuts around the loop (C2, C5, C6), plus output winding inductance, and a shelf at C3. I suspect that at full clipping it wants to be an LFO. In the suspected original application, it would get there only on the largest lightning-bursts, very short events. An LFO has a long start-up time. So it may never get into significant motor-boating with "lightning" (real or scratch-a-file fake lightning bursts).
I just can't see how this is better than a simple in-line passive clipper. I suspect it comes from a time when NFB was new and sexy and used everywhere, just as today we use micro-bots to tweet our electric consumption on Twitter (see this month's MAKE magazine).