Week 11: The Screen#

The question#

Your instrument has an answer. The question this week is whether you can get it off the screen and into a call before the lap it applies to is gone.

A display is a lossy compression of a distribution. The design decision is what you choose to lose — and the only way to find out whether you chose well is to measure a person using it.

One situation, all week#

Scenario wall-2027-05, lap 31 of 58, car 14 running third. The stop window closes on lap 38, so four stop laps are still open: 31, 34, 36 and 38.

The call is the decision you make on the radio: box now — bring the car into the pit lane — or wait for a later lap. The wall is the display you read it off. The window is how long you have to read it: eight seconds, on this course.

Eight seconds is a scenario value. It was chosen for the course, not measured from a real pit wall, and so is every other number in this deck. The archive is generated, so the arithmetic closes exactly.

What the model hands the screen#

Week 9 turned each strategy into a distribution over finishing positions rather than a number. For each of the four stop laps the model publishes eight quantities off it: the mean, the 10th and 90th percentiles, expected constructors’ points (the team’s score, which week 10 settled as what we rank under), the position you rejoin in, tyre age at the flag, the pit loss applied (the seconds a stop costs), and which stop laps remain.

Then the pairwise matrix — the probability each strategy beats each other — twelve cells once the diagonal goes. Then three situation values: the gap ahead, the undercut threshold (the gap below which pitting first puts you out ahead), and the hazard of an interruption before the window closes.

4 × 8 + 12 + 3 = 47, and that is only what the baseline prints.

The over-dense wall (1 of 2)#

Everything, at the precision the model produced it.

Two of the four strategy columns, four of the eight rows.

The over-dense wall (2 of 2)#

Below these two blocks the real wall carries the other two strategy columns, the twelve-cell pairwise matrix, and the three situation values — gap ahead 1.9042 s, undercut threshold 2.2000 s, interruption hazard 0.1173.

Forty-seven numbers, on one screen, at once.

Eight seconds against forty-seven numbers#

Nothing on that wall is wrong. Every figure is one the model genuinely produced, and a strategist who had a minute could use most of them.

The failure is not accuracy and it is not taste:

• the value that decides this call is one cell of a twelve-cell matrix, and the cell beside it is its complement, so reading the wrong one inverts the answer
• every value carries four decimal places. 20.4000 is a pit loss of twenty and a bit seconds; those three zeros appear in every scenario and have never changed a call
• nothing on the screen says which of the forty-seven the call depends on, so the reader searches instead of deciding

Four honest ways to compress#

Each of these is defensible. None of them is neutral.

1. A point estimate with an interval — “P3, likely P1 to P5.”
2. A probability one option beats another — “boxing now scores more points than waiting in 58% of races.”
3. A two-option comparison — box now against wait, side by side, and nothing else on the screen.
4. A single recommendation with a confidence — “box now, 58%.”

Communicating a probabilistic recommendation, and calibrating one, are live research problems rather than settled ones. There is no answer key at the back of this week.

What each of them hides#

The fourth hides the most and reads the fastest. That is the trade the week is about, and you are allowed to come down on either side of it as long as you say which.

What this call actually uses (1 of 2)#

Start from forty-seven and take things away, saying why.

• The call is binary — box now, or wait for the best later stop under the objective, the one with the highest expected constructors’ points. Two strategies matter, not four. Sixteen of the thirty-two go.
• Rejoin position, pit loss and tyre age at the flag are inputs the model already consumed to build these distributions. Reprinting them asks the reader to redo the model’s work. Six more go.
• Expected constructors’ points is the objective and ranks the options. It is not mean finishing position, and not what a reader checks a call against — so this wall prints the position and assumes the head-to-head probability stands in for the ranking. That assumption is the display’s own. Two more go.

What this call actually uses (2 of 2)#

• Of the twelve pairwise cells, exactly one is between the two live options. Eleven go.
• All three situation values go. The gap and the undercut threshold are already priced into the distributions, and the interruption hazard is inside the probability you are about to read.
• Which stop laps are still available stays, reduced to the one fact this call turns on: the lap the window closes. It is what makes this a decision now rather than later.

Left: a position and a range for each live option, the one probability between them, and the lap the window closes. Five lines, plus one saying what they were all computed under.

The legible wall#

Same 0.5831, printed as 58%: the probability box now outscores lap 36 on points. It agrees with the ranking here — and the wall bets it always will.

The objective is on the screen because it has to be#

Week 10 left you with the same distribution ranking options differently under a team objective and a driver objective. So a wall showing P3 · 58% without naming which one is showing a number whose meaning is unstated.

It is one line of small type, and it is the line that makes the recommendation checkable at all: without it nobody can say which ranking the number came from. Everything above it is conditional on it.

Legibility is a property you can lose#

Three more ways, all available to a careful designer who is not thinking about the reader:

• Precision the decision cannot use. The fourth decimal of 0.5831 has never changed a call and it costs a fixation every time.
• Colour carrying information on its own. Green-for-box is a fine second channel and nothing on its own — the recommendation above is carried by size and position, which survive a colour-blind reader, a sun-struck screen and a photocopy.
• A value in the wrong place. The 58% is in the largest type at the centre because it is the number the call turns on. Put it in the corner and you have not removed it; you have removed it from the eight seconds.

The measurement#

Twenty-four scenarios from the archive, split into two matched halves of twelve. Eight seconds per scenario. Reading only the display in front of you.

The harness shows the situation, records the call and the elapsed time, and withholds the model’s own ranking until both halves are done. You cannot tune a display against an answer you are looking at.

Three numbers come out, per wall: median latency, how many calls were made inside the limit, and how many were both inside the limit and correct. The third is the one that matters, because a fast wrong call is worse than a slow one.

What the two walls produced#

Both walls held the same forty-seven numbers behind them. Neither model was changed. Every figure here is a scenario value.

A result is not an explanation, so the next slide takes the 6.5 s apart.

Where the six and a half seconds went#

A course-instrumented run timestamped first fixation on the deciding value, so latency splits into finding and using.

Finding: 4.7 s of the 6.5. Forty-seven values against five lines, and on the baseline the deciding one is an unlabelled matrix cell.

Comparing: 1.8 s. 0.5831 beside its complement 0.4169, against 58% on the row of the option it favours — the second has the comparison already done.

Where the legible wall lost (1 of 2)#

Two of its twelve went wrong: one over the limit, and scenario 7.

A different situation: lap 26 of 58, box now against lap 33, the best later stop under the objective. The headline said box now takes the head-to-head 62% of the time. The reader boxed.

Synthetic, like every figure in this deck. The teammate’s points are held equal across both strategies, so this one car’s points settle the team’s ranking too.

Where the legible wall lost (2 of 2)#

Box now wins more often and scores less: its nineteen bad races are further down than its thirty-one good ones are up.

This is the assumption breaking. The headline is a probability of winning; the objective ranks by expected points. Different questions — and this wall put the one that does not settle the objective in the largest type.

That is a design choice, and a fixable one — print the points, or the stake. It is not proof that a compressed display must mislead. The cost of compression is real; which summary you promote is yours, and you answer for it.