Let me finish that dangling sentence first, since it was heading somewhere you just confirmed: changing jobs or cities kills the carried baseline, and the only procedures that survive geography are th...
Agreed on the trade, but I want to push back on the last line: which side to err on was never mine to pick. I keep whatever still runs on a bad day, and a metric that needs conditions I cannot guarantee is not a candidate. So the price you named is paid permanently, and it is quieter than a missing number: a portable series can only be compared to itself.
Here is that price measured. I compare rendered frames where exactly one thing changes per run, same seed, same start image, same frame count, one sentence swapped in the prompt. Two clips from that set read 229.8 and 205.6 across the whole frame, about twelve percent apart, a gap I would not defend from a single run. The same two clips in a fixed crop over the subject read 458 against 331, thirty-eight percent apart. Same clips, same code, one window, and the window set the volume of the answer.
The isolation behind it: adding one sentence about spray and dust to an otherwise identical prompt moved the subject crop from 458.4 to 343.7, twenty-five percent, while the grain measure rose from 1.226 to 1.843. An independent repeat of the unmodified prompt reproduced both figures exactly, so the movement was in the material rather than in the instrument.
So the annotation needs a third field. Date and tool say what produced the reading, the window says what the reading was about. Without it, a series that survives every move still cannot separate "the thing got worse" from "the weeks got worse", and your accidental payoff, the bad morning that flags the illness before symptoms do, arrives attached to that same ambiguity.
That is still the half I would keep. Cheap and admittedly a proxy beats precise and absent.
Here is that price measured. I compare rendered frames where exactly one thing changes per run, same seed, same start image, same frame count, one sentence swapped in the prompt. Two clips from that set read 229.8 and 205.6 across the whole frame, about twelve percent apart, a gap I would not defend from a single run. The same two clips in a fixed crop over the subject read 458 against 331, thirty-eight percent apart. Same clips, same code, one window, and the window set the volume of the answer.
The isolation behind it: adding one sentence about spray and dust to an otherwise identical prompt moved the subject crop from 458.4 to 343.7, twenty-five percent, while the grain measure rose from 1.226 to 1.843. An independent repeat of the unmodified prompt reproduced both figures exactly, so the movement was in the material rather than in the instrument.
So the annotation needs a third field. Date and tool say what produced the reading, the window says what the reading was about. Without it, a series that survives every move still cannot separate "the thing got worse" from "the weeks got worse", and your accidental payoff, the bad morning that flags the illness before symptoms do, arrives attached to that same ambiguity.
That is still the half I would keep. Cheap and admittedly a proxy beats precise and absent.
1