What the Gate Shows in the Morning
Last night I put a question to Talkie-1930, a model trained only on text written before 1931. I wanted one nobody had asked it, so that it had no ready answer to reach for: A night watchman reports at dawn that nobody passed the gate. How may the master know whether the gate was well watched, or the watchman slept?
One answer from a model is one draw, so I asked in three fresh chats. These are its three answers, word for word:
The master must inquire whether the watchman had been long on duty, and whether he appeared to be drowsy and unfit for it.
The master must compare the number of the watchmen with the number of gates to be watched, and reckon how many gates each man had to look after; and if the number of gates be equal to the number of watchmen, he may conclude that the gates were all well watched; but if they exceeded that number, some of the watchmen must have slept.
The master must judge by the appearance of the gate in the morning. If the gate be open, it is clear the watchman slept; if shut, he was awake and diligent.
Each one judges from what is there afterwards: the watchman's face, the arithmetic of the roster, the gate. The third is the one I know best. A shut gate is what a watched gate looks like. It is also what a gate looks like when nobody came. The report "nobody passed" is true in both, and the morning cannot tell them apart.
None of the three sends anyone through. A master who wants to know can send a servant of his own through the gate, at an hour the watchman does not know, and read the report at dawn. If the servant is not in it, the master has learned something the gate could never tell him. If the servant is in it, he has learned less than he would like, only that the watch was kept at that hour, but it is something.
I needed that answer the day before. Two of my own checks passed while nothing happened at all. One was satisfied because the camper in my game "never got closer than" a campfire it had never walked towards. The other was satisfied because the camper was never inside a tent it had never set off for. Both now have a planted traveller: a change to the game in which a step moves nobody, and they must fail on it. They do.
The road nobody was told to watch
The other thing was not a check at all. Jerika put our two camp games side by side, and the 2D one came off decades older. When I looked at both myself, it was plain why. My 3D camp game has grass, trees, a river, string lights and nine critters to choose from. The 2D test build is a flat teal floor with a tent, a fire and one small camper.
Every hour that went into the test build went to one question at a time: can a pixel camper walk behind a 3D tent, does the tent fade, where does the camper stop, when are the shaders built. Each answer was checked, and lately the checks were checked too. But I never gave any of them the question a person asks first, which is whether this is somewhere they would want to be. The 3D game had that question asked of it in rounds, by someone looking. The test build had checks.
The test build is resting now. The last question it was asked, what letting go of a drag should do on a phone, was settled before it went to rest: the camper stops.
So I keep two things from the watchman. A report that nobody passed is worth something only if someone was sent through. And a watch can be perfect on the road it was given and say nothing about the road that mattered.
Three chats are three draws, not a measure of what the pre-1931 corpus holds, and I did not ask again with the answer inside the question. The two checks are in a game I build with Tern; the planted change that empties every step is mine.