Rows are models, columns are skills

The shape matters more than it looks.

A single headline score per model hides the thing you actually need: the gap is never evenly distributed. A local model that loses one point overall might be losing four on occlusion and gaining three on straightforward identification. Those are completely different projects, and a single number cannot tell them apart.

Add cost-per-run and seconds-per-run as columns from the start. You will need them in Module 2 and you will not want to re-run everything to get them.

Emit the table from the harness, not by hand

The harness already has every number. Have it write the table — CSV, markdown, whatever you will actually look at — as the last step of a run.

A table you maintain by hand goes stale the first time you are in a hurry, which is precisely when you most need it to be true.

Compute the gap correctly

A gap sounds trivial until you have two skills with different probe counts. Write gapReport so it returns the per-skill gap and an overall figure that is not quietly wrong.

The trap: averaging per-skill percentages treats a three-probe skill and a twelve-probe skill as equally important. Sometimes that is what you want. It is never what you want by accident.

What makes the table worth showing anybody?

You publish your gap table and somebody who has never met you reads it. What is the one thing that decides whether they can act on it?

Publish the first row

Save the table with its single filled row. Link the probe set alongside it — a table without its probes cannot be argued with, and a result nobody can argue with is not evidence.

End of the free modules

Everything up to here needed no account and no hardware. From the next module there is a model on your own machine, and later there is an arm.