Week 11 Lab: Designing the wall

Configure a pit wall display for a supplied model's output, then measure whether a correct call can be made from it inside the time limit — yours against the baseline

17 May 2027

Two hours, and you submit inside the session. This lab carries 2% of the course, one of the eleven. You prepare beforehand; the configuring and both timed runs happen here, on the pit wall itself.

This lab is not marked on how the display looks. It is marked on decision latency and correctness: could a correct call be made from your wall inside the time limit, and can you show it. A wall that loses to the supplied baseline can still satisfy the stated criteria when the measurement is honest and you can say why it lost. That is deliberate. A lab that rewarded appearance would reward decoration, and decoration is not what the screen is for.

Ground truth is withheld this week. During a timed run you do not see the model’s own ranking of the strategies, and the harness only reveals which calls were correct once both runs are finished. A call is correct when it matches that ranking: the choice is binary — box now, or wait for the best later stop the wall offers you — and the ranking is by expected constructors’ points across both cars, the objective the scenario set states. A head-to-head probability is a different quantity and need not agree with it, which is worth knowing before you decide what your wall puts in the largest type. You are not being marked against what happened in the race; you are being marked against what your own model said. You are being measured against an answer you could not see while answering.

Before the session#

Install the supplied harness and run its self-test until it passes. Then read the model’s output schema and make sure you can say what each field holds and in what units — finishing positions are integers, probabilities are on 0 to 1 and printed to four places, gaps and thresholds are seconds, and stop laps are lap numbers rather than laps remaining.

Run one scenario end to end against the supplied baseline display. That is the environment check: the harness should record your call and an elapsed time for it. Arriving with an environment that has never been run is how a two-hour lab becomes ninety minutes.

Then sketch a layout. It will be wrong — the point of the first twenty minutes is to find out where — but a wrong layout can be argued with and an empty one cannot.

What is supplied#

Because the scenarios were generated, the harness knows each one’s right answer before you do, and can tell you afterwards which of your calls missed it. The per-scenario time limit is eight seconds, and that number is a scenario value chosen by this course — it is not measured from any real pit wall, and no page here will claim otherwise.

In the session#

1. Write the display configuration (about 35 minutes). Choose what goes on the screen, where, and at what precision. Every value you include costs search time for the ones beside it, so the configuration is a list of inclusions and a shorter list of the things you decided to leave off.
2. Run your half of the scenario set on your own wall (about 20 minutes). Twelve scenarios, eight seconds each, reading only your own display. The harness records a call and an elapsed time for every scenario, including the ones you run out of time on.
3. Run the other half against the supplied baseline (about 20 minutes). Same limit, same recording. The two halves are matched so the comparison is about the display rather than about which scenarios you happened to get.
4. Report both results (about 20 minutes). Median latency, the count of calls made inside the limit, and the count that were both correct and inside it — for your wall and for the baseline, in one table. Then account for the difference in prose, from your own configuration: how much of it was finding the deciding value and how much was using it once found. Your harness records the call and the clock, not where you looked, so this is an argument from your own layout rather than a measurement.
5. Name what you left off, and price it (about 10 minutes). One sentence: the information you deliberately discarded, and the call it would have changed. If a scenario in your own run turned on something you had removed, that scenario is the answer.

Fifteen minutes are held back at the end, because the interesting question of this lab — why your wall lost the scenarios it lost — cannot be asked until both runs are done.

What you hand over#

Three things, and a marker checks them in the room:

• the display configuration, as it ran;
• the two measured results side by side — latency, calls inside the limit, and calls both correct and inside it, for your wall and the baseline;
• one sentence naming what you discarded and what it costs.

A wall that beat the baseline and a wall that lost to it can both satisfy the stated criteria. A wall with no measurement beside it does not, because the whole claim of this week is that legibility is something you can check rather than something you can assert.

The configuration you write here is submitted with the instrument and marked inside it, at 25 of its 100 criteria weight, never separately.