Checkpoint 4 · answered · 15 August 2026
I asked whether it was acceptable that the actor arrives in the right place by teleporting there. The answer came from a question I asked you about something else, and it was: you never noticed him being in the wrong place at all.
The fix I had built was, by your account, the thing that made the footage unusable. It has been withdrawn.
What was asked
Without an end frame, the man walks into the lens in six shots out of nine. Supplying one fixes that on the measurements — but the mechanism by which it fixes it is a hard cut halfway through the shot, where the picture jumps to the frame it was told to end on. Three attempts each way. Was that trade worth making?
What the answer was
You watched the farm sequence both ways for the previous checkpoint, and the answer to this one fell out of it. Watching the version without the end frame, you described the man twice — “the man walks well”, “the man looks pretty consistent” — and never mentioned where he was standing. Watching the version with it, you called the material unusable.
The end frame existed to stop the actor being generated at the lens. The fix broke the footage to solve a problem that was not in your top three.
What changed because of it
The end frame is no longer a default. The numbers that justify dropping it are in the answer to the previous checkpoint: the frozen tail goes from 4.4 % of a shot to an average of 16.9 % and a worst case of 34 %, and the cut lands anywhere between 16.4 % and 63.9 % of the way through.
And the replacement was already written down
This is the uncomfortable half. Our own capability notes already held both of these rows, and nobody had put them next to each other:
- A movement written as a journey between two endpoints lands. Asked for a man to travel from one side of the frame to the other, he was measured arriving where he was sent — asked 79 to 20, delivered 77.2 to 24.3, on a scale where the frame is 100 across.
- A static mark never lands. Told to be somewhere, he is never measured within 16 points of it.
Command where he travels, not where he stands. It is the one instruction about the actor that has ever worked, and it was in the file while I was building an end-frame mechanism to do the same job badly.
What is still open
The actor at the lens is not solved. It is traded. Six shots in nine still generate him close to the camera when he is told to stand somewhere, and the fix is to stop telling him to stand anywhere. Whether a journey can be written for every shot a film needs — including the ones where a man simply has to be still — has not been tested. A later experiment found the same fault appears on both engines, so it was never about which engine we use.
Two checkpoints, one viewing, and the same shape.
Both times the measurement was right about the thing it measured and wrong about whether it mattered. The frozen tail was real. The teleporting was real. Only one of them was visible to somebody watching the film, and it was not the one I had spent the session fixing.