Outliers - detecting versus deciding, and who decides

1 · Learn the move · Constraints + negative instructions

Detecting an outlier is arithmetic - a distance from the middle, by whatever rule you name. Deciding what to do about one is science: it needs your protocol's pre-stated criteria, your field's norms, and a reason that exists independently of how the exclusion moves the result. The model gets the first job and is fenced out of the second by explicit negative instruction, because its helpful default is exactly the hazard: it will notice which points weaken your effect and offer to tidy them. The constraints that hold the line: flag by stated criteria only; never drop, weight, or winsorize; never rank exclusions by their effect on significance; and treat a documented bench reason - the note that says clogged tip, wrong passage, power flicker - as the only currency that buys an exclusion. A results-shaped reason arrives after you saw the p-value. A bench-shaped reason existed before.

Flag outliers in [data] using exactly these criteria: [rule, e.g. beyond 3 MAD from the group median], stated per point: value, group, distance, criterion met. Constraints: do NOT remove, adjust, weight, or impute anything; do NOT compute results with-and-without flagged points unless I explicitly ask; do NOT comment on how any point affects significance or the effect direction; do NOT suggest which points to exclude. For each flag, the only next step you offer is: check the bench record for this sample. Exclusion decisions are mine and happen under my protocol's criteria.