What it does

Derivative, through its Forge pipeline, takes a plain-language software requirement and compiles it into a structured contract, builds a Python implementation, runs it in an isolated sandbox and validates it before packaging. The generated code does not get to declare itself correct. Each run ends as verified, validation_failed or infeasible_proven, the last being an evidence-backed certificate that the stated constraints contradict each other. Validation runs in an ephemeral Docker container with no network and a read-only filesystem, and failed checks trigger bounded repair attempts. Current scope is greenfield Python CLIs, REST services, data pipelines and libraries. The project says it does not cover editing existing repositories, other languages or frontends. It can run deterministically without an API key, or in a hybrid mode backed by an OpenAI model. The repository reports its benchmark results openly. On its latest frozen blind set, none of the six expected-verified cases reached the external oracle.

Capabilities

Requirements compiled into typed contracts
Validation in an isolated Docker sandbox
Independent checks and bounded repair
Packages only verified builds, with evidence files
Runs without an API key or in a model-backed hybrid mode

When to use it

You want to generate a small greenfield Python tool and keep a record of how it was checked
You are exploring verification-first approaches to AI code generation
You need a CLI, REST service or data pipeline scaffolded with tests