The second series through the same pipeline: 0 hand corrections, 26 hours, and what carried over

The second story, by the numbers

The first series was made while the pipeline was being built under it, so its numbers say little about what the pipeline does on its own. The test was a second story through the same pipeline, the same bible, and the same rubric, written by someone who had never seen the compiler. That story is The Brightwater Family — Halloween, three parts of about seven and a half minutes: costume shopping, the pumpkin patch, trick-or-treating.

How it was made. A fresh AI session with no access to the pipeline wrote all three scripts from a script format, an asset list (the only locations, costumes and props it could name), and a page about the characters. The compiler was its only feedback: anything it could not stage came back with a reason. Four revisions went back to the writer, three of them staging rather than story. No line of any script and no frame of any part was changed by hand. Every finding from the gates, the judge, the reviewer or me went into the format, the compiler, or the library, so it applies to the next story too.

The beats row needs its denominator read with it: Halloween scripted 23 beats against Picture Day's 86, because the script format offers the writer only seven stage directions. 13 of 23 performed is a better rate than 25 of 86, and a smaller task. Of the 10 counted "line is shown" flags, 1 sits in a shot whose beat was a hold, 2 in shots whose beat was performed, and 7 in shots with no scripted beat; 10 shots have a held beat and 10 a flagged line, union 19, so the two
rows overlap in one shot.

What carried over. Of the character pose drawings used, 178 of 211 came from the first series; 54 of 65 character looks; 33 of 42 background plates (the night versions are the first series' rooms, graded). New for Halloween: four places (a costume aisle, a neighbor's porch, a pumpkin patch, the street with its sidewalk), eleven costume looks, two mouth sets derived from the approved one (fangs for the vampires, chocolate for the toddler) and six props. The costumes are generic: a yellow shark, an inflatable dinosaur, two vampires, a dog's pumpkin bandana.

Time. From the full-length scripts to the last part passing review took twenty-six hours, almost all of it review rounds. On the final attempt a part takes about an hour end to end: eight minutes of audio, eight to twelve minutes of compositor render, under a minute of gates, and about forty minutes of the judge looking at every shot. The first series' "three to four hours per episode" counted generation and judging and is not the same measurement.

What the rounds were. The 102 compile rejections were all gaps in my format or compiler (a van the format offered but the library could not seat, a "clean" state the compiler did not read, a street the asset list promised before it was drawn); none was a writing error. The reviewer's findings were the kinds a viewer would catch: a character entering without a cut, a walk that did not move, a family that vanished from a wide for three seconds, a toddler's mouth drawn
on his nose. Each became a test. The reviewer also missed some of them on the first pass and said so; those misses are in its log too.

On counting. A commenter pointed out that "beats accepted as holds" and "lines said but not shown" can count the same miss twice. On the first series' Part 3, 10 of the 28 flagged lines sat in a shot whose beat was a hold, 5 in a shot whose beat was performed, and 13 in shots with no scripted beat at all; 8 shots appeared in both rows. From now on both rows are published with their overlap.

Watch Halloween (three parts, about 7½ minutes each): https://www.youtube.com/playlist?list=PLYPuSb4GK878.

The first series and the full write-up, with the gates, the judge and the review protocol:
https://atveanimation.com/brightwater/how-its-made.

Made with AI tools; all characters, voices and places are fictional; the costumes are generic; the music was composed for the series.