Every Liquid Was Yogurt

In the spring we launched a pipeline that publishes one short video to Instagram every day: it invents a concept, writes the prompt and caption, generates the clip with synced audio, checks its own work, and posts. The first engine we ran it on was fast, cheap, and produced footage that looked genuinely good — right up until anything in the frame was wet.

Ocean waves rolled like something you would spoon over muesli. Coffee left the pot in a single glossy rope. Rain did not fall so much as descend, reluctantly, in fat pearls. Every liquid — every one — moved with the physics of yogurt.

Here is the part worth writing down: nothing in the pipeline disagreed. The generation step reported success. The vision-model QA judge, which inspects every clip before it can post, scored the yogurt ocean as perfectly acceptable footage — it was sharp, well-composed, on-prompt, free of artifacts. By every measurement we had built, the clip was good. It took a human, watching the preview with ordinary human eyes, to say the obvious: water does not do that.

Why the metrics missed it

A judge can only fail work along dimensions someone thought to name. We had taught ours to look for corruption, incoherence, prompt drift, visual noise — the failure modes we had already seen. Wrong viscosity was not on the list, because it had never occurred to us that it could be a category of error. The model was not making mistakes inside our taxonomy; it was making a kind of mistake our taxonomy could not express. Aesthetic physics is exactly the sort of thing you notice instantly and specify never.

The fix for the footage was straightforward engineering: we swapped the engine for a two-stage open-weights video pipeline that respects fluids, and concepts that lean liquid-heavy get routed accordingly. Water now behaves like water. That part took a week.

What it changed permanently

The lasting consequence was not technical. The yogurt incident is the reason this company keeps one deliberate human touchpoint in an otherwise autonomous publishing loop: before a reel posts, a person sees it and taps approve or reject. No edits, no notes — one tap. Everything upstream of that tap and everything downstream of it runs itself.

We have been offered plenty of clever ways to automate that tap away, and we have tested a few. Twice we tried to make the QA judge catch temporal weirdness automatically, and twice it failed in ways that were themselves instructive (the second attempt is the subject of its own post-mortem). The conclusion we keep arriving at: automated judges are excellent at enforcing a known taxonomy and structurally incapable of noticing that the taxonomy is missing a row. The tap stays. It is the cheapest insurance in the company — a few seconds a day against publishing the next yogurt ocean to an audience that would, rightly, never let us forget it.

If you take one thing from this: when people say "no humans in the loop" they are usually describing an aspiration, not an architecture. Our architecture is humans only at the approval gate — and this incident is why the gate exists.