Where I started
I wanted to understand the visual vocabulary of a family of stylized rocks and express it procedurally. The early work concentrated on planarization, reduction and cutting into a single mass.
The results kept drifting toward convex, gem-like forms. I redirected the experiment toward the design of the primary masses: a dominant peak, a shoulder and a smaller supporting shape.
What I changed
I made the black silhouette the first test. Surface detail would come after the relationships between the large forms worked.
The agent handled procedural variations and measurements. I used the live viewport and references to decide whether those variations actually moved toward the intended shape language.
What the measurements could not choose
A rock could satisfy measurements for aspect, concavity and plane count and still fail visually. The recurring problem was treating those measurements as the design target.
The useful finding was the importance of relationships between masses. More surface complexity on the same underlying shape did not solve the composition.
This remains an exploratory study. Its value is the change in how I framed the problem and evaluated the results.