About

subkutan is a lab for computational experiments — controlled tests that probe how language models behave under specific, repeatable conditions.

The name comes from the German subkutan (subcutaneous): beneath the surface. These experiments look past what a model says and measure what it actually produces when given identical instructions at scale.

Methodology

Each experiment starts with a single, precisely defined prompt. That prompt is run dozens of times across different models and rendering constraints. Every output is captured as a self-contained artifact, then evaluated against a fixed feature checklist and scored across multiple quality dimensions.

No cherry-picking. No iteration. One prompt, many generations, systematic evaluation.

Current experiments

Archimedes Benchmark

A physics simulation prompt evaluated across 3 rendering modes (ASCII, HTML, SVG), 3 Claude models (Haiku, Sonnet, Opus), and 3 prompt detail levels (Short, Middle, Detailed). 1,687 evaluations across 1,728 generated files.

Erstelle eine einzelne HTML-Datei die das archimedische Prinzip visualisiert. Keine externen Abhängigkeiten.

Each level also has ASCII and SVG variants that add a renderer constraint line. The core prompt structure is shown above for the HTML variant.

Contact

Built by kugelmensch. Powered by onprem.ai.