Check, run, inspect, compare

One project remains one readable case.py, with operational metadata in agentfem.toml. These commands do not require an AI assistant.

agentfem check --runtime
agentfem run --name baseline
agentfem show latest
agentfem show latest --files

check checks project structure and Python syntax without executing the case. --runtime also checks numerical runtime readiness. Neither promises that a particular mesh, material, boundary condition or solver combination is supported; the model's own preflight remains authoritative.

show --files lists recorded artifacts and reports missing files with exit code 2. Keep an XDMF file together with its referenced HDF5 files. This command checks recorded paths, not the internal contents of visualization files. Normal show keeps the file list short; JSON and --files retain the complete list.

Compare selected results

agentfem compare outputs/001-baseline outputs/002-candidate \
  --quantity displacement_max --atol 1e-8 --rtol 1e-5

Choose quantities explicitly, with an absolute tolerance in their recorded units. Repeat --quantity or use --history reaction. One invocation applies the same tolerances to all selected observables; use separate invocations when different quantities need different tolerances.

The comparison requires completed result manifests, identical declared units and shapes, and, for histories, stored samples with identical increasing coordinates. It never silently converts units, interpolates histories or maps fields between meshes. Missing units are not assumed dimensionless: record 1 explicitly. Missing or incompatible data fails closed. Relative error against a zero reference norm is unavailable; absolute error is still reported.

Acceptance is componentwise abs(candidate-reference) <= atol + rtol*abs(reference). Exit code 0 means the selected observables meet these tolerances; code 2 means failure or incompatibility. Add --json for machine-readable output. This is not independent scientific validation and does not promote either result's trust.

A small real finite-element campaign

agentfem init wall-study --template heat-campaign
cd wall-study
agentfem check --runtime
agentfem run --name first -- --samples 3
agentfem run --name resumed -- --samples 3
agentfem show latest --files

The installed template solves steady heat conduction, varies the hot boundary temperature deterministically, and exports a scalar-response dataset and audit. The project-local campaign-data ledger lets the unchanged plan reuse completed cases. Edit the physical problem in case.py; consult the template README before changing an existing campaign. This is a bounded workflow example, not a trained surrogate or a claim that arbitrary thermal models are validated.