Autolivery
An agent that paints race cars.
Give it a brief. It designs a livery, fits it to the car's real 3D model, and revises until the checks pass. A person accepts it.
Fitment is a verification tool the agent cannot talk its way past: millimetres, against the car's geometry.
The loop doesn't know it's a car. The car is in what you plug into it.
Measured, not judged
Fitment measures, in millimetres
- Letter height on the car: a number at least 140 mm, a name 45 mm
- Margin from every edge, shut line and door handle
- A name's contrast against what is under it, 6:1
- A stripe's direction, alignment and gaps, panel to panel
- How much of each piece every view actually shows
The critic judges, from pictures
- Does it read as the style the brief named
- Is each thing the brief asked for present, and where
- Composition and taste, which is what a model is for
- A second look, closer, when the two disagree
Anything the model can compute is measured. The critic only answers what a picture can show.
What was built this week
liverykit, over the summer
- Reads a car's model to find where each texture lands on the bodywork
- Liveries as code, fitted per car; a fitting editor
- Tools over MCP that propose and never commit
- Profiled 26 cars it had never seen, from one script
Autolivery, this week
- The loop: draft, measure, judge, revise, propose
- Drafts measured and rendered without ever being adopted
- The critic, its eval against a person's verdicts, and the second look
- New checks: what each view shows, letter sizes, contrast, stripes, layouts by search
- Every call traced: rounds, tool calls, latency, cost
The same loop passed on a GT3, an Abarth and a Formula 4, first round, under two minutes, under a dollar.