How a film made of code gets directed

A product film from our lab, X-Arc, is a program: a webpage that reproduces the interface and plays itself on a timeline, authored in code and the web where models are strongest, so every pixel is rendered rather than generated. That premise shipped as Cue, and this note reports the research that continued behind it. The premise does not cover choice: when the maker is a program, someone, or something, still has to decide which film to make.

A walkthrough, the form Cue shipped with, mostly escapes the question, taking its direction from the interface it shows. The films we wanted next asked for more: an open aesthetic, carried by typography, motion, and an original score. Two problems follow that a walkthrough never poses: which film to make when the interface does not decide it, and what the score should be. A walkthrough had always posed the third, how a system that can neither watch nor hear checks what it made; the open form only raised its price. Two finished films below are the evidence.

The first attempt

The occasion was a launch film for an integration release, and the first attempt came back a walkthrough, the grammar the system knew. Execution was not the issue: a judged frame critique of the recording had returned eight findings, all fixed. The issue was direction; the piece reproduced the form of the film before it. This series has traced the pull toward the median of training, the tendency to return the most typical version of whatever is asked. Iteration exposes a second median, the system's own precedent: asked again, it converges on the center of what it last made, and a stronger model converges more cleanly, not less. Distance from precedent had to be arranged, not requested.

Stage 1: One brief, four angles

The arrangement runs four directors in parallel on one shared context, verbatim to each: the brief, with the first attempt and why it was set aside; a verified base of facts about the product; the aesthetic family that must survive; the medium's physics, including the film's frame rate, 25 frames per second, which the recording must hit; the killed patterns; the target length; the one takeaway. Each director is forced onto one creative angle, and the set is gated before any run: each angle's dominant grammar goes down in one line, a walkthrough would read interface-led, the winner here read type-led, and if two share a grammar, one is replaced. We call the arrangement directed divergence: distance between attempts enforced before generation rather than hoped for after. Every treatment returns in one eight-section format so judges compare like with like: a beat sheet whose dwell, how long a line holds on screen, runs at roughly one second per 13 characters; every on-screen line, verbatim; the risks at the film's frame rate.

Stage 2: Judges with forced spread

The treatments pass to judges reading through three lenses: impact, directing originality, and craft risk, meaning buildability at the film's frame rate. Every judge scores every treatment, with the scores forced apart so clustering cannot hide a preference, and names the winner's biggest weakness and one element worth stealing from each loser. The synthesis keeps the winner's grammar, fixes its named weaknesses, and admits at most two grafts; a synthesis that borrows freely is a committee average. The output is a frame-exact shooting treatment, the film specified before it is built.

The evidence of two films

On the first film, 27.4 seconds, a type-led treatment won decisively, 26.6 out of 30 against 23.4, 21.1, and 20.9, with two elements grafted from losing treatments: a numeric counter that races upward and locks, and a chain of approvals resolving step by step to done. On the second, 28.1 seconds, the top two treatments tied at 24.7. The retrospective traced it upstream: the first film's directors had been preceded by a discovery pass, four scouts assembling the verified fact base; the second skipped it, and its directors, diverging from thinner ground, flattened toward each other. Divergence quality is downstream of discovery, and a tie at the top of the scores now reads as a sign a real choice upstream was never made, not a result to accept.

The score

The score extends the premise to a second medium: not licensed and cut to fit, but synthesized as a program on the film's own clock. The tempo is chosen so a beat lasts a whole number of frames: at 25 frames per second, 125 beats per minute puts a beat at 0.48 seconds, exactly 12 frames; 150 gives 10; 120 gives 12.5 and is rejected. One scheduler places every visual event on a frame. Every sound is baked sample-exact at its beat in the same grid: the kick, the riser, a rising sweep, the hit under a title, the click of a press. The audio is padded by the export's lead-in so bar one lands on the first frame of content.

We call the substrate the shared clock: film and score are two readouts of one timeline; nothing needs aligning afterward. Synchronization is not edited in; it exists by construction, and drift is zero by arithmetic. The first film's score is dark and minimal at 125 in D minor; the second, at 150, uses shutter snaps as percussion.

The perception

None of it can be checked by watching, so perception is built as structure. The page emits timing marks, start, typing window, close, end, so trims and sound placement are exact. An untrimmed preview recording is dumped to stills every half second, 89 for the first film. Parallel panels read every frame through distinct lenses, story and impact, motion and cleanness, type and fidelity; each finding names its frames, a severity, what is wrong, what should be there. A per-scene pass re-checks each window at the film's frame rate; the second film's pass put ten judges on the recording and returned roughly forty findings, all fixed. We call the loop mechanical perception: sight rebuilt as instrumentation.

The capture

Scene-level perception found what whole-film passes had been reading past. Realtime 4K capture of a browser runs at an effective 8 to 12.5 frames per second, far under the film's 25, so every animated movement was recorded as move, freeze, move: camera pushes lurched and fades descended a staircase. One fade concentrated 87 percent of its total change into a single captured frame; counted on the 255 brightness levels a pixel can take, that frame stepped 198. Six of ten scenes carried the artifact, and every earlier 4K master, recorded through the same physics, carried it too.

The first fix failed honestly: stepping the page's virtual time frame by frame snaps CSS transitions to their end state. The fix that held is one no camera pointed at footage has, because a program's time is a parameter: run the page three times slower, capture in real time, compress time back in the export. The same fade's largest step fell from 198 to 32; the camera pushes came back bell-curved, zero frozen pairs. Even the browser's own slowing control stalls on pages that rapidly cancel and restart transitions, so the page governs its own slowdown, per animation, every animation frame.

The exhibit

Below is the first film's finished master: 27.4 seconds announcing an integration with Composio, directed, scored, and read frame by frame through the process above. Watch it with sound: the hit under the title, each press click, and the closing resolve land by arithmetic, on the grid the type and motion share.

What capability did not solve

Directors, judges, builders, and the capture diagnosis ran on Fable 5, the newest model class. The baseline quality of a first attempt rose visibly, and the pull toward precedent survived the rise: capability sets the floor, and direction had to be built for the film to be anything in particular.

Three terms join the series vocabulary: directed divergence, distance between attempts enforced before generation; the shared clock, one grid that visuals and score compile against; mechanical perception, sight rebuilt as instrumentation. Cue made the film a program; this run showed the filmmaking moving the same way: what must be exact became arithmetic, the gate, the clock, the marks, the frame dump; what must be judged stayed judgment, run in parallel and forced apart. The film was always code; now the filmmaking is.