Evidence for scientific and operational claims

Capability, support, validation, reliability, default, and recommendation claims must point to a stable entry in evidence-registry.yaml. Evidence can be an invariant test, runnable example, generated receipt, source implementation, or versioned scientific artifact. A path merely existing is not sufficient when the claim is numerical: the referenced test or artifact must measure the stated quantity.

Use an inline marker in maintained documents:

This interface writes forces. `[evidence: learned_multipole_forces]`

When evidence is missing, say so prominently:

**⚠ UNVERIFIED:** This behavior was observed on an older cluster image.
`[evidence: pcstudix_npt_reliability]`

The registry checker verifies that every referenced claim exists, every verified or generated claim has resolvable evidence, referenced pytest node names exist, and unverified claims explain what evidence is needed. This makes the links suitable for CI without pretending that source prose proves a scientific assertion.

Heuristics embedded in code must be named as heuristics and expose their inputs as configuration. Calling a number “safe,” “recommended,” “validated,” or “optimal” requires an evidence ID in the nearby docstring/comment; otherwise it must be labeled UNVERIFIED HEURISTIC.

CI also runs the focused hard-coded recommendation audit. Its phrase scan is intentionally conservative and does not replace scientific review; statements outside its vocabulary still require registry entries.