Writing down your prediction

One job is left before day 14: write down what you expect to happen, before Improve starts. It's a short document, a few numbers and a few sentences, and it's what turns the first loop from an initiative into a test. The pilot kit calls it the pre-registration. This lesson builds yours.

Why it has to be written first

Decide what success looks like after the results arrive, and you will always find some. Output went up, or quality did, or morale, or people learned a lot. It's like painting the target around wherever the arrow landed. Nobody is lying; there are just too many ways to read a result after the fact, and the kind ones are easier to believe.

Medicine learned this the hard way. Around 2000, large US clinical trials began registering their main outcome on a public register before they started. A study of the biggest heart and lung trials funded by the National Heart, Lung, and Blood Institute looked at what happened next.

The authors suggest that declaring outcomes in advance may have contributed. Once the outcome was fixed before the results came in, there was nowhere for a disappointing result to hide. That's uncomfortable, and it's the point. A prediction that can't count against you can't count for you either. Write it down first and the result counts, whichever way it goes.

What goes in it

The fields come straight from the pilot kit, so a team that finishes this course is ready to run a pilot with the same form.

• Today's level, and where the number comes from. Pick the system of record now and don't switch it mid-loop.
• The predicted level, at the bottleneck, in the aim's unit.
• The measurement window: the days you'll measure over, such as days 60 to 90. Not one good day.
• The expected early dip: how far things may get worse first, and when.
• Kill criteria: the point where you stop this approach and change it, or take it to Build.
• Success and failure thresholds: the result that counts as the framework working here, and the result that counts against it.

Expect the dip, and write it down

Changing how work is done nearly always makes it worse before it gets better. People learn new steps, templates change, and the old way and the new way run side by side for a while. Economists Erik Brynjolfsson, Daniel Rock and Chad Syverson call the pattern the productivity J-curve (American Economic Journal, 2021). With a general-purpose technology like AI, organisations first invest in things that don't show up as output, such as redesigned processes and new skills, so measured productivity dips before it rises.

A 2025 US Census Bureau study of industrial AI in American manufacturing, by Kristina McElheran, Brynjolfsson and colleagues, found the same shape inside firms: losses in productivity and profit in the short run, before longer-term gains. It also found industrial AI increased work-in-progress, meaning more work sitting between steps. That's the pile you learned to look for in lesson 1.

A dip nobody predicted looks like failure, and it gets the work stopped in week three. A dip written down in advance is just the plan. The pilot kit's rule is plain: keep the dip in the pre-registration, and don't read it as failure until the window closes.

Signed and locked on day 14

On day 14 the bottleneck lead and the person who runs the loop sign the prediction, and it's locked. You'll do that on your prediction sheet in the next lesson. From then on, if something has to change, the change is recorded with its reason, and the original is still the prediction that gets scored on day 90.

That sounds rigid. It's what makes the result worth having. If the bottleneck moves as predicted and the aim's outcome doesn't, that's a real finding: the wrong bottleneck was picked, and the readout will say so in its first line. You can only learn that if the prediction stood still while the results came in.