TRUTHCUTFORCE AUDITStart a conversation ↗

INDEPENDENT MODEL CONSISTENCY · BY DYMENT.STUDIO

Find the inconsistency.
Before the long run.

An atomistic model can return plausible forces that do not belong to a consistent potential energy. We look for reproducible counterexamples—and give your team the exact path back to the problem.

Explore the $750 pilot ↗Inspect the reference tests ↓

A bounded investigation. Your model can stay on your machine.

THE QUESTION01 / CONSISTENCY
∮ F · dRDOES THE LOOP CLOSE ENERGETICALLY?

For a conservative force model, work around a closed configuration path should vanish within numerical error.

Conceptual illustration · not measured data
Full configuration spaceAll atoms move—not a 3D field-grid shortcut.
Numerical controlsRefined paths, reverse traversal, repeat evaluations.
Evidence you can replayCoordinates, forces, settings, and evaluation counts.

01 / INSPECT THE INSTRUMENT

A finding needs
a reproducible counterexample.

The examples below are executed synthetic reference tests. They demonstrate the instrument; they are not findings about a commercial AI model.

Loading the saved reference report…

Your selected report is read inside this browser. It is not uploaded. This viewer does not execute your model.

What we compare

Ordinary point checks, nearby energy-gradient checks, and closed-path evidence. We publish the benchmark with the runner, including cases where a simpler check is enough.

02 / THE DELIVERABLE

Less detective work
for your simulation team.

Use the audit before a larger simulation, while integrating a model, or when a regression makes results difficult to trust.

  1. 01

    A scoped finding

    What failed, where it failed, and whether refinement supports the conclusion.

  2. 02

    The exact counterexample

    Coordinates, force outputs, model settings, tolerances, and source identifiers.

  3. 03

    A repeatable regression

    A local runner and report your team can use after a model or integration change.

03 / FIRST PAID PILOTS

One model.
One clear scope.

For teams with a model intended to represent conservative forces and a specific simulation or release decision to make.

Existing verification tools, including OpenKIM, already test energy–force consistency. This service adds an independent, scoped investigation and a handoff your team can reproduce.

FIXED-SCOPE PILOTCOMPATIBILITY REVIEW FIRST

$750 USD

  • One accepted CPU-runnable model
  • Up to 10 nonperiodic configurations, 100 atoms each
  • Up to two CPU-hours of agreed tests
  • Findings report, recorded evidence, and local reproducer
  • Target: two business days after scope confirmation
Request a pilot ↗

We confirm compatibility and scope before invoicing. You pay for the investigation and deliverables; finding an error is not guaranteed. This is not a physics or regulatory certification.

Opens your email app. Nothing is submitted here. Please do not attach confidential models before the scope is agreed.

04 / WORK WHERE YOUR MODEL LIVES

Keep the model local.
Bring the evidence.

The Python runner evaluates trusted local model code. The website only reads the saved report. No account or model upload is needed to inspect the reference tests.

Download the local runner ↓Read the quick start ↗

Current boundaries

Initial pilot: position-dependent conservative force components, nonperiodic configurations, and accepted local adapters.

Thermostat forces, friction, external driving, periodic-cell workflows, expensive remote APIs, and unsupported model environments need a separately agreed scope.

Reference tests establish software behavior. We have not established accuracy or performance on your model.

GROUNDINGWhy nonconservative forces can cause problems ↗Existing OpenKIM verification checks ↗