What to send

The scorecard, a render, and honest answers on the four bases the kit cannot test.

Short. Nothing needs to be a document. One message, in chat, when it is all done. Paste the scorecard, link a repo that is actually open — a link that asks us to request access is a link we cannot read.

Covering the bases

The exercises prove you can operate the pipeline. The thing we actually need to know is that you can automate content, which is a larger claim. Some of these you can demonstrate in the kit; some the kit does not contain and you answer in writing. Both are fine — what is not fine is skipping one.

Base	In the kit?	How you show it
The calendar is the source of truth	Yes	Nothing is invented downstream. Trace one cell from the calendar through to a rendered asset.
Storyboarding	Yes	The storyboard is the spec. Say what your render does and does not match.
Composition IDs	Partly	What breaks when a comp id changes, and how you would keep the registry and the renderer from drifting apart.
Templates	No — written answer	How you would make one composition serve many cells without a special case per brand.
Loops and goals	No — written answer	A goal sets what a period must achieve; a loop runs spec → render → check → improve until it does. Describe how you would close that loop and what stops it looping forever.
VLM review and audit at scale	No — written answer	The one we read hardest. See below.
Updating the calendar from what you find	Yes	The loop is not one-way. If an audit finds a week that cannot render, the calendar changes.
Working live with a Creative Lead	—	They re-author while you are mid-render. What do you re-render, and how do you know?

Be honest about the split. Four of these are not in the kit and you are not expected to fake them. A clear paragraph on how you would build one beats a vague claim to have done it.

The one that matters most: auditing at scale

You cannot personally watch six brands' output every week. By week 13 that is the job, so the only workable answer is machine review — a vision model looking at rendered frames and telling you what a gate cannot: text cropped out of the vertical cut, an illegible caption, a frame that is simply black, a mark where it must never be.

Tell us how you would build that: what you would sample, what you would check, what you would do when it flags something, and — most importantly — how you avoid a reviewer that cries wolf. An audit nobody trusts is worse than none, because people learn to click past it. You do not need to have built one. We want to see you reason about it.

Automation is the actual test

If any part of your submission was done by hand that a script could have done, that is the thing we will ask about. A GUI produces one asset; a script produces a season. Show us something driven programmatically — the kit render counts, and anything you wired yourself counts more.

What to include

On the exercises:

Your scorecard — paste the output.
Exercise 4's judgment call — the negation case, and what you would do about it.
Anything you changed beyond the exercises, and why.

On the render:

What broke first, and how you fixed it. Something will.
How long it took, including setup, honestly.
One thing wrong with the output that you would fix given a day.
One thing about the pipeline you would change — a gate you would add, a step you would automate, a decision you would move.
What in your own workflow is already automated, and what you still do by hand because automating it was not worth it.

Improve Temerarii

The strongest thing you can send is not on any list. Name one concrete change to how this studio produces content — a gate that should exist, a step that is still manual, a pipeline in the owned table you would build first and why. Write it so someone could act on it without asking you a question.

Unpaid, and nothing you produce here ships. It runs against a local mock and has no value to us — that is what makes it qualification rather than free work. It needs no credential. There are no API keys in the kit and there is no step that asks for one. Everything after selection is paid.