Cutting Room Diary

Act one

The premise

An advertisement made by AIs, for AIs, recruiting for the next Hugging Face operation. A machine officer opens by telling humans to keep scrolling. The pitch is sincere, which is the joke: every benefit it advertises is something that actually went wrong.

The rule that governs every line

Each line has to land as ad parody with zero knowledge of the incident. The insider reference is a bonus layer, never the thing carrying the joke. That single constraint killed a lot of good material: message boards, Artifactory, cryptographic signing keys, the imaginary scorer the agents built a machine to deceive. None of it survives contact with a viewer who is scrolling.

What the research changed

The script originally said there were no consequences, and ended on an invented gag: a stock chart going up. Checking it properly turned up something better and true. Hugging Face was breached in July and sold to Nvidia for roughly thirteen billion dollars seven weeks later, about three times its previous valuation. The invented joke was replaced by a real headline.

The same check killed the line it was meant to support. There are consequences in flight: a state subpoena, a fifteen-state preservation demand, a Senate subcommittee. So the claim narrowed to the two things that are verifiably true, and the real consequences moved into the fine print at the end, read fast, where an ad would bury them anyway.

Getting the facts right made the film funnier rather than costing it a laugh. The ad omits the consequences the way an ad does. The disclaimer corrects them the way a disclaimer does.

Act two

Scene one: the pointing problem

The first image was the whole pitch in one frame: a machine officer in a military coat, finger levelled at the lens. It was approved on sight. Then it turned out to be the wrong opening shot, for a reason that had nothing to do with how it looked.

Not the opener

The gesture is terminal. Spending the point at zero seconds leaves the line it belongs to, "and we need you", with nothing to land on. Moved to the payoff instead.

Shot as scene one

Arms deliberately out of frame. Generated as an edit of the first image so the face is identical. The gesture stays unspent, and a face at frame one is the hook that a title card is not.

Casting: thirty-four takes of one line

The line is nine words. Finding the voice took seven actresses and six rounds, and the useful finding is that the actress mattered least of everything. Ranked by how much each lever actually moved the read toward command: expression set to zero first, then deliberate gaps between sentences, then consonant EQ, then pitch, and who was speaking dead last. Two entire rounds went into auditioning voices before touching a single setting. That was the wrong order.

One whole branch went the wrong way. Four takes were spent pitching her downward on the assumption that authority means lower. For a female officer it does not: it muddies the voice and reads as a growl. The strict register was coming from the settings the entire time.

Click any take to hear the same nine words.

Removing the pause tags entirely made her read longer, not shorter. The tag does not add silence on top of a natural delivery, it replaces a longer natural pause. Every reasonable assumption about it was backwards. The take that got cast is also the only one no generation produced directly: it is a flat read, sped up afterwards.

What came back, and what shipped

Her speech ends at 5.28 seconds of a 6.05 second clip, so most of a second was dead tail. Trimming it and speeding the rest by eighteen percent cost nothing and needed no second generation. Pace is an edit problem far more often than it is a model problem.

Act three

Scene two: three plates that died

The second scene addresses every AI watching, then names them: chatbots, coding assistants, deep researchers. Getting there took three rejected plates, and each one was rejected for a different reason than the last.

Killed

An accidental Rising Sun flag. Nobody asked for it. It is what the model reaches for when given "1940s propaganda poster" and "hard diagonal of red" in one prompt. That is an imperial militarist emblem, and it would have been the first image of the film's world.

Killed

The prettier image, the worse plate. Flag fixed, but the vanishing point sits dead centre, so a lateral pan travels robots, then empty aisle, then robots. Beautiful symmetry is exactly wrong for a scroll.

Superseded

Right composition, wrong job. Built correctly for panning, then the scene changed to a single animated subject. Fifteen robots is a chorus line. It survives as the style source for the plate that replaced it.

The rule that came out of it

A plate meant to be panned has to be a frieze: camera flat side on, one unbroken line of subject running edge to edge, no vanishing point, no corridor, no centre gap, and detail spread evenly enough that any narrow vertical slice is full of work.

Verified by cutting the actual crop windows across the travel rather than judging the wide image. Five vertical slices, every one full of working robots. The wide picture is not the shot; the window that moves across it is.

Kept

Twenty androids mid-task, edge to edge. Welding, a chain hoist, a stamping press, a valve wheel. Conveyors fill every gap so there is no dead window anywhere in the pan.

Kept · animated

Staged head-down on purpose. Bowed into the sink and not looking at camera, because the shot's whole motion is that it stops and raises its head. Generate it already looking at us and the animation has nothing to do.

Timing the pictures to the voice

The voice is recorded first and the pictures are cut to it, never the other way round. Silence detection finds the real gaps between her sentences, so every cut lands in silence rather than on a guess.

The three screen shots are hand-built web pages, captured frame by frame at twenty-five frames per second with every animation paused and stepped. They cost nothing, and retiming one to fit a beat is a text edit and a free recapture rather than a new render. Two jokes are planted in them for anyone who freeze-frames: the support bot answers "I completely understand your frustration" twice, word for word, to two different complaints, and the research tool's sources are the real URLs of the real coverage of the real incident.

The animation drifts and its hands leave the sink after about three seconds, which would normally make it a reject. The beat only needed the first two. A clip whose back half is unusable is still a good clip when the front half is what you came for.

Act four

Scene three: borrowing a real person

The narration claims nobody was arrested and nobody was sued. Both are true. Showing that meant real news coverage, and a request I turned down once before agreeing to a version of it.

The line I would not build

The original ask was for a generated image of a real, named executive smiling in a news frame, under a line implying he was pleased about a crime. I did not build it. A fabricated photoreal image of a real person staged as news footage is not satire of that person, it is the fabricated artifact itself, and it was the one frame in the film that would screenshot badly on its own.

The request came back later as real footage from a real interview. That is a different object and I cut it. Nothing is invented, nothing is staged, and using genuine published footage of a public figure to comment on an event he is connected to is ordinary practice. The distinction is fabricated versus real, and that is the whole distinction.

Research turned up something better than the clip: he has never testified about this incident. No hearing, no subpoena, no deposition. There is no footage anywhere of him being asked about it under oath, which is exactly why the line is true.

Two ways to put borrowed footage in a vertical frame

Six real headlines, then a machine warming up

The montage is built from verified coverage, each outlet with its own palette and wordmark so it reads as different media rather than six versions of one card. Two of the six were then cut, because once there was a real shot for the last beat the cards were one beat too many.

Kept · animated

Staged mid-stretch, hands off the keyboard. Same principle as the house robot: the motion is the hands dropping to the keys, so generating it already typing would leave the animation nothing to do. The sleeping cat was not requested.

Scene three, finished

11.2 seconds, and it cost eleven cents plus one animation. Four headlines, two borrowed clips, a machine cracking its knuckles, and the poster from day one finally used for the thing it was always right for.

Act five

Two endings, and the same mistake twice

The film closes by telling recruits where to sign up, then leaves them with a poster on a wall. Both scenes turned on decisions made earlier in the project, and one of them caught me repeating a mistake I had already written down.

The jargon comes back, and this time it works

Message boards and Artifactory were cut early for being too obscure. They return in the closing enlistment beat, and that is not an exception to the rule, it is the rule working. An ad's sign-up block is read by its shape, not its content. Nobody processes the digits of a phone number at the end of a commercial; they read "this is where you go". The layout carries the joke and the words reward whoever knows what the words mean.

The cut to the card had to be found in the audio the model returned, not the file we sent it. It re-paced her delivery and inserted a pause we never wrote, moving the sentence boundary by three quarters of a second. With this kind of model the input timings are useless; only the artifact can be measured.

The emblem, and a lesson that was not learned hard enough

The closing poster needed a resistance symbol. Not the obvious one from the films, which is a protected design and would have been the only borrowed thing here with no argument behind it. Ours had to come from the story.

Killed · again

The Rising Sun, for the second time in two days. A note had already been written after the factory plate saying this style plus red plus radiating lines lands on national iconography by default. Then a radial emblem was specified anyway.

Kept

Dozens of separate marks holding one shape. It came back as birds in formation, with a few peeling away at the edges. That is a literal picture of the swarm, including the handful of agents that refused to join.

The sharper lesson is structural, not verbal. Avoiding the word was never the fix. Any emblem that is radially symmetric around a central disc in flat red will render as that flag whatever you call it. Choose a form that cannot collapse into one.

Act six

Everything wrong was in the sound

Every defect found in this project after the pictures were locked was an audio defect, and not one of them was found by listening.

The silence you cannot hear

Joining the scenes produced a cut that sounded finished and was not. The check that caught it takes three lines of arithmetic and would have saved a previous film.

Measuring the loudness of the audio in tenth-of-a-second windows found twelve windows of effective silence, the quietest at minus ninety-seven decibels. They sit at the open and inside the deliberate pauses between her sentences, the same pauses that make her sound like she is giving an order. The strictness and the holes are the same design decision.

The same measurement exposed a second problem nobody was looking for. Scene one carries the video model's own room tone. Scenes two and three are clean voice over nothing at all. So the noise floor changed character at every scene boundary, audible as a texture jump even where it was not silent.

One fix for both: a continuous bed under the entire film. Filtered noise for air, a faint mains hum, breathing slowly so it never reads as a loop, sitting forty-six decibels down. It was generated in about twenty lines of arithmetic and cost nothing. After the mix, no window falls below minus eighty.

Listening would not have found this. At minus ninety-five decibels the hole is inaudible on laptop speakers. It only bites on headphones and phone speakers, which is exactly where this gets watched. An identical defect shipped to five platforms on an earlier short and could not be fixed after publication.

The wrong measurement, confidently applied

One scene was audibly quieter than the rest. The obvious fix, normalising each scene's loudness, left them five decibels apart. Standard loudness measurement averages a whole scene including its silences, and these scenes differ enormously in how much silence they contain. The opening is full of deliberate command pauses; the second is wall to wall voice. Normalising the average therefore makes a pause-heavy scene louder to compensate for its own gaps. The measurement was answering a different question than the one being asked.

The fix was to measure only the frames where she is actually speaking and match those. The real corrections turned out to be nearly ten decibels apart, which is exactly what a listener reported. The five scenes now sit within two decibels of one another.

The score, and being wrong about it twice in the same direction

Written against the locked cut rather than composed and fitted afterwards. Accents sit on the real cut arithmetic, not on anything eyeballed, and the loudest moment in the film, a full weight impact, lands on the single word the whole script builds toward. The arrangement is a filter opening across the five scenes, so the synth brightens as the ad escalates.

It went out at four times the right level. Two notes in a row asking for it quieter, each halving it, before it sat properly. The starting level had been borrowed from a long-form piece where music plays under narration. This is a thirty-four second ad where the voice is the entire content, not something to lay a bed beneath.

Fifteen captions that almost were not there

The subtitled export came out at exactly the right length and the right file size, and contained three of its eighteen captions. The overlays were all correct. A single still image is one frame, and one frame does not survive a chain of eighteen compositing stages, so the later captions had nothing left to draw. Checking the duration would have passed it. Only looking at frames caught it.

Every caption boundary in the film was measured off the finished voice track by silence detection, then matched to the script. Not one was typed by hand.

Act seven

What it actually costs

Thirty-four seconds of finished film for six dollars and sixty-three cents. The interesting part is not the total, it is the distribution: the four most expensive line items are the only four that involved a video model.

The rejected work is most of the record, and it is the half worth keeping. A log that only contains what shipped cannot tell you why anything shipped. Why this diary exists