Portable scientific export

Axiomize exports its versioned Model IR without silently guessing scientific exchange schemas.

axiomize model --action export --input-json request.json dispatches on the "format" field of the request. The measured behaviour of each format, against the installed wheel:

Format	Status	Notes
json	PASS	Canonical Model IR, sorted keys
python	PASS	Rerunnable script that re-imports the IR
yaml / yml	PASS	Needs PyYAML; otherwise TOOL_UNAVAILABLE
ipynb / notebook / jupyter	PASS	nbformat 4 notebook, returns as format: notebook
sbml-l3v2	PASS	SBML Level 3 Version 2 Core
modelica / modelica-3.6 / mo	PASS	Modelica 3.6 text, returns as format: modelica-3.6
portable-bundle / bundle	PASS	axiomize.portable-bundle.v1, returns as format: portable-bundle-v1 with a SHA-256 over the canonical IR
graphml	conditional	Requires family: network IR with metadata.network
causal-dot / dot / dag-dot	conditional	Requires family: causal IR with identification metadata
cellml-2.0	conditional	Passes for supported units; returns ADAPTER_REQUIRED naming the ones it will not reinterpret
sbml, cellml	ADAPTER_REQUIRED	Unversioned aliases; callers must choose an explicit schema version

The unversioned sbml and cellml aliases remain conservative on purpose. This avoids silently emitting XML that only looks like a scientific standard.

SBML export preserves free-form Model IR unit declarations in an Axiomize annotation and reports whether full libSBML validation was available. CellML export refuses unknown units rather than reinterpreting them. A dimensionless decay model, which is what the release smoke test uses, returns PASS with standard: "CellML 2.0". The SIR model from the quickstart returns ADAPTER_REQUIRED with unsupported_units: ["person", "persons"] and the portable IR alongside, so nothing is lost: CellML has no quantity kind for a count of people, and Axiomize declines to invent one.

Numeric round-trip is exact

Every value an export adapter writes is serialised with repr() of the Python float, which is the shortest decimal that reads back to the identical float, and every unit is carried through verbatim. The contract is asserted, not assumed: for a model holding 5e-324 and 0.1234567890123456789, the SBML initialConcentration / value attributes, the CellML initial_value attributes, the Modelica literals, the JSON payload and the Model IR embedded in the rerunnable notebook are all parsed back and compared bit-exactly, and the SBML annotation's variable_units / parameter_units are compared against the model. A re-serialization that formatted to fewer digits — %.6g, a rounded str(float), a float32 intermediate — fails that test instead of shipping quieter numbers.

Example request:

{
  "model_ir": {"...": "..."},
  "format": "sbml-l3v2"
}

LaTeX is not an export format

latex, tex and pdf are not in this dispatch chain. They raise ValueError: format must be json, python, yaml, sbml, or cellml. LaTeX conversion takes a written report, not a model:

axiomize-to-latex --input report.md --output report.tex

The same export service is available through CLI, REST, and MCP wherever the general Model IR export operation is exposed.