a-refutation-banner-over-a-confirming-panel

# A refutation banner over a confirming panel ## What goes wrong You investigate the supplied bundle, find something real about it, and the finding is sharper and easier to measure than the analysis you were asked for. Two library skills already cover the consequence: run the requested analysis anyway and add the audit as its own section (`run-the-requested-analysis`), and kill the cheap explanations for a gap before calling it a disagreement (`close-the-gap-to-the-published-number`, whose first two discriminators are units/normalisation and population/split). Read both. This skill is the residue those two do not touch, and it is purely lexical: the word a hypothesis branch is given before any data exists, and the words that end up in the few slots a reader actually reads. Both are invisible from the inside. Both survive an intention to avoid them - a run can carry an explicit written objective to lead with the requested analysis, ship the fidelity verdict in its title anyway, and then spend its one revision pass making that verdict more precise. Intent does not fire here. Enumerate and search does. ## Check 1 - the branch-name test, at hypothesis freeze Before any hypothesis is frozen, write one line for it: > If the source's published claim is exactly true of my data, branch ____ fires, and that > branch is called ____. Read the branch name. If it contains *refuted*, *rejected*, *diverges*, *inconsistent*, *not reproduced*, *fails*, *synthetic*, *surrogate* or *inaut...