The tracker already knows what needs making. So we stopped retyping it.

The problem

On a small animated short, our tracker already held everything that mattered: which shot was approved, which note came back and from whom, which version superseded which. And then a person read that off the screen and typed it somewhere else. The tracker described the work but had no wire to it, so the wire was a human, and that is where the mistakes live.

This is not an AI problem. Any shop with an expensive automated step, a render submission, a transcode, a delivery package, has the same gap.

How it works

A production tracker is the shared database a crew already uses to say what needs making and who approved it. Ours is Autodesk Flow Production Tracking, which most people still call ShotGrid. signalbox watches it and does the work it implies.

Somebody asks for a change in the tracker. The change gets made. The result comes back to the same place for a person to approve. Nobody opens a second tool, and coordinators keep working where they already work, which is the part that made it stick.

For the technically minded

A watcher loop polls tracker state about once a minute and acts on it. There is no queue and no scheduler, so the service can be killed mid-render and restarted with nothing to reconcile: what it was doing is still written down in the tracker. "Why did this render?" is answered by reading the state that caused it, in the tracker's own event log.

What worked, and what didn't

The thing that surprised us: asking a vision model to check another model's output did not work. We built a gate to grade each rendered panel against its approved design, and it failed a lot of panels a person had already been happy with. We switched it off rather than tune it. Plain comparisons against the approved reference did the job instead, and they are cheaper and easier to trust.

The thing that caught us out: a batch of shots sat waiting and nobody noticed, until someone opened one and asked why. The stage holding them used a field that did not appear on any of the pages people actually look at. A queue would at least have looked full. This looked like nothing at all. A state nobody can see is a state nobody acts on.

Everything expensive is automatic. Every approval is a person. That split is the whole design.

For the technically minded

Characters stay consistent by compositing from an approved reference rather than from a prompt, the same way a VFX pipeline keeps anything consistent. Checks that gate work are deterministic, and each one has to be broken on purpose and watched fail before we trust a clean result from it.

What this is

A bare-bones snapshot of a larger internal system, trimmed to the show-agnostic core and published because the decisions in it might be useful to somebody. It is not maintained and it was never packaged for anyone else's show, so please do not adopt it as a tool. Self-tests needing the original models or a live tracker will not pass from a clone.

The repo's ARCHITECTURE.md and METHOD.md go further, including the parts that did not work.