Kernels

Executable kernels with declared verification boundaries and linked theorem cards. Each kernel has an input/output schema, optional unit_constraints (strings), and (for test_status: "tested") a structured contract_v1 describing units, domains, tolerances, and formal obligations. The portal renders kernel pages from corpus/kernels.json and linked theorem cards. See Contributor playbook – Verification boundary.

Shared test helpers (kernel-conformance)

Cross-family numeric helpers (for example assert_close, monotonicity grids) live in the workspace package kernels/conformance/, and new kernel families should depend on that package in tests instead of duplicating helper code.

Numeric witness contract (SPEC v0.2, numerically_witnessed)

Kernels that implement empirical or computational isotherms use boundary numerically_witnessed: correctness is attested by:

1. Pytest suite in kernels/adsorption/tests/ — bounds plus property-style checks (monotonicity, invariants, scaling), and I/O schema validation. CI runs uv run --project kernels/adsorption pytest (see JUSTFILE just test).
2. Reproducible examples in kernels/adsorption/examples/ — scripts with fixed inputs and small sweeps to validate behavior across ranges.

Formal proofs for related claims live under formal/ScientificMemory/; the kernel is a separate numeric witness, not a substitute for machine-checked Lean.

Corpus enforcement

Pipeline validation requires every kernel with non-empty linked_theorem_cards to have test_status: "tested" (fail-fast). Kernels marked tested must also provide a complete contract_v1 (units/domains/tolerances/witness/obligations) and keep passing the adsorption package pytest suite in CI.