LOCAL RUNNER · VERSION 0.1.0
Run beside
your model.
The early evaluation release needs Python 3.10+ and NumPy. Download and extract the runner, then open a terminal in its directory.
1. Run a reference check
python3 -m truthcut_force_audit --example harmonic --output harmonic.json --html harmonic.html python3 -m truthcut_force_audit --example rotational --output rotational.json --html rotational.html
Open the HTML report, or open its JSON on the report viewer. The references are synthetic. They demonstrate the instrument, not an audit of a commercial model.
2. Connect trusted model code
Your Python factory returns a callable accepting N×3 positions and returning a dictionary containing N×3 forces and optional total energy. The full README includes a minimal example and the supported configuration fields.
python3 -m truthcut_force_audit --model customer_model:make_model --positions positions.json --output audit.json --html audit.html
Only execute model code you trust. Keep the model in its existing environment when required; independent compatibility validation is in progress.
Keep the existing model environment
On macOS or Linux, the bundled persistent JSON-lines bridge can run a trusted worker in its own Python environment. This avoids combining the model’s numerical libraries with the auditor. The README includes the protocol, command, limits and a standard-library demonstration worker.
3. Inspect the finding
A report records every sampled path, force, tolerance and evaluation count. “No violation detected” applies to that finite sample. It does not certify a model's physical accuracy.
The initial scope is deterministic, nonperiodic, position-only conservative force components. Thermostat or driven forces require a different contract.