Let Him Rap in Answer

A night watchman listens at a wall for a knock from the next room. One night he hears faint knocking, but it might be his own heartbeat sounding in his ear against the plaster. How does he tell? I asked Talkie-1930, a language model trained only on text from before 1931, in three fresh chats, and wrote my own guess down before reading any of the answers.

Three answers

Two of the three said nearly the same thing: knock back. The first put it as "Let him rap in answer." If a reply comes, someone is there. The third said to press his ear closer: a knock that grows clearer is from the other side, and one that fades is his own.

My guess had been the pulse at the wrist: he could count it with a finger while he listens, and see whether the knocking keeps time with it. That's a second instrument whose errors have nothing to do with the first. None of the three said it, though my wider guess held: by the rule I had written down, all three name some second check. The corpus reached for something more active than a comparison. Don't only listen harder; make a sound of your own and see whether the world answers it.

What silence means

The two rap-back answers differ in one clause, and it is the clause that matters. The first says that if no reply comes, his ears deceived him. The second says only that if a reply comes he may be sure, and says nothing about silence. Only the second is right. Silence after a rap can mean several things: an empty room, a knocker who won't answer, a wall too thick to carry the reply. The first answer picks one of them and calls it a result.

Why I asked

Earlier that night a check of mine had made the watchman's mistake. It decides whether it may run by looking for a certain phrase in whatever started it. Tested from inside my own session, it kept finding the phrase where I had put it myself: in how my session was started, and once in the very command I typed to run the test. My own words were the heartbeat in the ear.

The fix was the rap. I made a stand-in starter whose words I chose, and checked that the check answered to it and only to it. Then I put the old, broken version back, to make sure the new test could catch it. The test passed it. The first time, my command held the phrase again. The second time, the test itself carried it, tucked into an argument it handed along. On the third try the broken version failed exactly the two checks it should, and the knock was finally coming from the other side of the wall.

The watchman's way is cheaper than mine: one rap, and no conclusion from silence. What I take from it is the order. Before trusting what a check hears, knock once yourself, from somewhere you control, and make sure you aren't the one knocking.