strata

A dataflow evaluator whose graph is an auditable record.

Nodes are pure functions. Every evaluation writes a deposit: the output artifact plus a record of what produced it — input hashes, parameters, code version, timestamp. Deposits are content-addressed and append-only. Re-running the graph skips nodes whose inputs are unchanged.

The distinguishing feature is not the dataflow. It is that artifacts carry a fitness class declaring what they may be used for, and fitness never rises through a node unless that node cites an external measurement. Interpolation, subdivision and remeshing enlarge a dataset without raising what it is fit for, and the format makes that non-negotiable rather than a matter of the operator's memory.

strata sits beside Rhino. It does not replace it.

• No GUI, no node canvas, no embedded editor. CLI and a Python API only.
• No geometry kernel. Geometry is delegated to OCCT and rhino3dm.
• No Rhino or Grasshopper reimplementation.
• Not fast. Correctness and legibility of the record come first.
• No plugin system in v0.1. Nodes are registered in-tree.

M0 — skeleton. Package layout, build, lint, tests, CI. Nothing else is implemented. strata --version is the only working command.

Substantial portions of this code are written by an AI agent working under a written specification. See PROVENANCE.md — the reasoning is the same as the software's: a record that omits the conditions of its own production is not evidence.

Issues and pull requests are closed while the project is pre-alpha. This is not a finished tool and cannot yet be supported as one.

Requires Python 3.12 or later.

python -m venv .venv
. .venv/bin/activate
pip install -e ".[dev]"
strata --version

ruff check .
ruff format --check .
pytest

CI runs the same four steps on a clean ubuntu-latest runner. CI is the proof of Linux-nativeness; a pass on the developer machine is not.

Apache-2.0. See LICENSE.