Two hours, in two halves. First half, laptops closed. Each team projects its best figure and the room asks three questions, every time: what is the control, what is the confound, what would falsify this. Everything visible in a figure asserts something. The assertions you made on purpose are usually fine. The ones you made by default are where figures go wrong.
Second half, you rebuild the figure the room attacked. The house spec is broadly what a reviewer expects: an interval on every estimate, the control on the plot rather than in the text, direct labels at the line ends, a title that states the finding and carries the sample size. Then you make it harder to attack by triangulating: the same measurement across seeds, then across a second model.
What you will do
- Present your figure and take notes on every question you could not answer.
- Rebuild it against the checklist, literally.
- Wrap your measurement in a function of the seed and make sure the seed reaches everything random in it.
- Re-run on a second model. Everything you wrote in weeks one to five is already model-agnostic.
What you leave with
- The figure that becomes Figure 1 of your write-up.
- An interval on your main result, and a second model that either agrees or does not.
Materials
Notebooks and slides are released to each cohort at the start of the session.