| “Users who received X improved, so X works.” | Observational rung; association supported; intervention effect unsupported unless identification/design results close the gap. |
| “We changed X once, so the policy works everywhere.” | Interventional result limited to its population/environment/window; transport requires exact endpoints and assumptions. |
| “The simulator shows what would have happened.” | With no causal reliance, exit to model reporting. With causal reliance, cite the simulation result, assumptions, validation, supported model use, and unsupported realized/interventional use. |
| “The trial was randomized, therefore the estimate is valid.” | Run the common threats: interference, attrition, measurement, adherence, and analysis can still lower the result. |
| “The observational estimand is identified.” | Cite the identifying expression/derivation for identified; a bound supports bounded wording, and a nonidentification witness supports nonidentified wording. The label alone is incomplete. |
| “The fairness metric improved, therefore the intervention is fair.” | Report metric change. A counterfactual-fairness claim additionally needs its causal estimand, counterfactual-identifiability assumptions, estimate-consistency basis when used, and bounded C.28 support before D.5 audits it. |
| “Logged replay says this policy is optimal.” | Cite behaviour/evaluation policies, overlap, confounding, transport, uncertainty, and bounded support; unqualified optimality is unsupported. |
| “Method A beats Method B causally.” | Use G.9; different rungs, estimands, support components, endpoints, or windows require a bridge with stated loss, degraded parity, or abstention. |