confound-checklisted

# Confound Check ## The rule A confound is any explanation, other than the claimed cause, that produces the same observed pattern. Two consequences drive everything below: 1. **Naming a confound is worthless without the observation that would distinguish it.** Every rival explanation gets a discriminating test, or it stays live and the claim gets softened. "This could be confounded by seasonality" is not a finding. "If it were seasonality, last year's same-week data would show the same bump — check it" is. 2. **If you cannot name the units that did *not* get the treatment, and say why they were comparable, there is no causal estimate.** There is a description with a causal verb attached to it. ## Mode Route on what already exists. Do not run the full pass when a shorter one is called for. | Situation | Mode | Run | |---|---|---| | Nothing measured yet | **Design** | §1, §2, §6, then write the §4 ledger as the pre-registration: arms, controls, the confounds being accepted, and the checks that will decide each one | | A result exists and is being reviewed | **Review** | §1–§8, full | | Causal language appeared mid-analysis | **Guardrail** | §1 and §2 only, no ledger. If §2 produces an untested list, stop and surface it *before* the conclusion is written down | Guardrail triggers — fire on these in any draft, comment, commit message, or chart title over non-randomized data: *drove, caused, led to, resulted in, because of, thanks to, the impact of, X% lift