ineedvalidation

Declare which physical case, kind of evidence and response quantity a test exercises; collect the declarations across repositories into an evidence ledger; lint and render a validation hierarchy from node notes plus that ledger.

The vocabulary follows Oberkampf and Roy, Verification and Validation in Scientific Computing (2010): a validation hierarchy of unit, benchmark, subsystem and system cases, and a strict separation between verification (the code solves the equations correctly) and validation (the equations describe the measured world).

Marking a test

import ineedvalidation as vv

@vv.case("free-space-propagation", "code-verification", srq="on-axis intensity")
def test_matches_closed_form():
    ...

@vv.case("free-space-propagation", "solution-verification", refined_parameter="grid-spacing")
def test_second_order_in_the_step():
    ...

@vv.case("free-space-propagation", "cross-code-benchmark", reference_code="other-code")
def test_matches_independent_code():
    ...

@vv.case("detector-frame", "validation", referent="lab-flat-field-2026")
def test_matches_measured_frames():
    ...

@vv.seam("optics", "detector", case="detector-frame")
def test_photon_scale_anchor():
    ...

@vv.regression
def test_frozen_output():
    ...

Evidence kinds: code-verification, solution-verification, cross-code-benchmark, validation (measured data only) and uncertainty-quantification. They are spelled out rather than lettered so a marked test reads without a key. reference_code= belongs to a cross-code benchmark, refined_parameter= to solution verification and referent= to validation; each raises when used with another kind, so a mislabelled test fails at import. Regression tests claim no kind of evidence.

The markers register through a pytest plugin entry point, so installing the package is the only setup a repository needs.

Tagging by path

A table in pyproject.toml assigns a case and evidence kind by test path, so the tagging cost is per file or directory rather than per test. The longest matching prefix wins, and an explicit marker overrides the table.

[tool.pytest.ini_options]
testpaths = ["tests"]

[tool.ineedvalidation]
library = "mylib"

[tool.ineedvalidation.defaults]
"tests/test_kernel.py" = { case = "free-space-propagation", evidence = "code-verification" }
"tests/benchmark/" = { case = "free-space-propagation", evidence = "cross-code-benchmark" }

Paths are matched relative to the pytest rootdir. A [tool.pytest.ini_options] table in the same pyproject.toml pins the rootdir to the repository; without one, pytest can settle on a directory above it and no default matches.

Collecting the evidence

$ pytest --vv-evidence evidence/mylib.json

One JSON file per repository lists every marked or defaulted test with its case, kind of evidence, response quantity, seam and the outcome it reached, stamped with the commit. The library name comes from --vv-library NAME, else [tool.ineedvalidation] library, else the rootdir name.

Linting, scaffolding and rendering a hierarchy

$ ineedvalidation lint hierarchy/ --library-map libraries.md
$ ineedvalidation scaffold hierarchy/
$ ineedvalidation render hierarchy/

The hierarchy is a directory holding nodes/ (one Markdown note with YAML frontmatter per node, listing the cases it owns under cases:), evidence/ (the ledgers) and optionally views/; --nodes and --evidence point elsewhere. The lint checks the notes for internal consistency and checks them against the collected tests in both directions: a case no node owns, a node with no tests, a response quantity a node does not list, or a validation level claimed without a passing test that supports it. scaffold writes a stub note for every case a test names that no node owns. The renderer draws the levelled figure through d2, with named views (--view, --all-views), a single branch (--root) and optional hand-edited overrides for figures that have to look right in a talk.

Status

Pre-alpha. The markers, the evidence collector, the hierarchy lint, the scaffolder and the renderer work; the API is not yet stable.