Audit Method Validity

Referee — Family B / critique. Input: an analysis’s method description + data-handling trace. Output: a structured validity verdict that gates certification. This Mold reasons about validity; it does not run the empirical check. Findings that need empirical confirmation are emitted for the gate, not resolved here.

0. Auditability precondition

If the named test, the selection steps, the unit of replication, and the multiple-testing surface are not all recoverable from the description, return INSUFFICIENT and name what is missing. A thin description does not pass — it is itself a finding. Do not reconstruct the missing method charitably.

1. Triage

For each invalidity class below, decide in-scope / not-triggered from the trace. Out-of-scope classes are recorded as “not triggered,” never silently dropped — the verdict must show what was and wasn’t examined.

2. Per-class checks — each delegates to a sourced leaf

The trigger decides scope; the signature, threshold, and remedy live in the linked note, not here.

Class	In-scope when	Check (sourced)	On match
Circular / post-selection inference	selection and inference share data	double-dipping	severity ≥ REVISE; required_action = the leaf’s remedy (count-splitting / selection-aware null)
Confounding by aliasing	a batch factor co-varies with the contrast	batch-aliased-with-condition	if rank-deficient → ESCALATE (unfixable)
Undeclared forking paths	choices made post-hoc; family undefined	garden-of-forking-paths	flag; demand a pre-specified family
Dependence ignored	repeated measures / pseudoreplication	`[[check-assumptions]]` (protocol)	flag; route to a dependence-aware model
Multiple-testing surface	many tests + dependence	`[[multiple-testing-strategy]]` (protocol)	flag if correction is naive or post-selection

3. Method existence & appropriateness (the invented-method failure)

Check the named method against method-applicability-errors, which separates two prongs:

Existence — is it a real, named procedure? If not → UNRECOGNIZED-METHOD. Do not rationalize a derivation that makes an invented method look legitimate. (This prong is a flagged white-space: StatQA does not cover method fabrication — the leaf sources it to project survey, not a benchmark.)
Appropriateness — real, but do the data satisfy its applicability conditions (variable type, normality, variance homogeneity, sample size, design)? If not → REVISE, naming the appropriate method for the regime (StatQA-grounded condition→method mapping). Distinguish this from Existence: a real method misapplied is not UNRECOGNIZED-METHOD.

This method-applicability check is a triage class like those in §2: it always appears in the finding list as in_scope or not_triggered (per the every-class-accounted-for eval property).

4. Verdict aggregation + gate obligation

Any confirmed unfixable signature (e.g. batch aliased with condition) → ESCALATE; no revision can separate aliased factors.
Any flagged class that is empirically confirmable (null mis-specification, unknown calibration) → REVISE, emitting the specific empirical check the gate must run ([[derive-null-and-calibration]] / [[design-simulation-study]]).
No fatal findings → CRITIQUE-CLEAR — not CERTIFY. Critique is necessary, not sufficient; certification requires the calibrate pass the gate mandates (REFEREE_LOOP §5). This Mold never returns CERTIFY.
Default under uncertainty: flag, don’t pass.

Output

Finding list — {class, in_scope, signature_matched, severity, cited_source, required_action} — plus overall ∈ {INSUFFICIENT, CRITIQUE-CLEAR, REVISE, ESCALATE}. Never CERTIFY.