The confidently wrong test, and why it looks right
Lesson 1 of 6 · Statistics you can defend
1 · Learn the move · Red-team audit
Ask a model which test to run and it will answer fluently, with example code, in seconds. That is the problem. It accepted your framing without asking about your design: whether your replicates are technical or biological, whether the groups are paired, whether the wells share a plate. Fluency is not evidence, and a recommendation built on unasked questions is the confidently wrong test. The move: after any recommended test, make the model attack its own recommendation. List the design facts it assumed without asking, the assumptions the test requires, and the conditions under which it is the wrong choice here. This area's standing discipline is the same idea: re-derive any statistic you plan to report by a second method, and match it before you trust it.
You recommended [test] for my experiment. Red-team that recommendation, do not defend it. List: (1) every design fact you assumed without asking me, marked [NEEDS: ...] where I have not supplied it, (2) the assumptions [test] requires and which ones you cannot check from what I gave you, (3) realistic conditions under which [test] is the wrong choice for this design, (4) the questions I must answer before any test is defensible. Do not pick a test for me.
2 · Your turn. You write the prompt
You asked for help comparing viability across three drug concentrations in a plate-based assay, and the model suggested t-tests between each pair of groups, with tidy example code. Something nags: three groups, several tests, and all the wells came off the same two plates. Write the prompt that makes the model audit its own suggestion before you run anything.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.