The Third Umpire · Lesson 4 — The Right Calls ← Course

The right calls — which examples to show.

Which examples you pick matters more than how many — two sets of the same size can behave nothing alike.

Last page counted examples; this one chooses them — and the choice matters more. The model learns the pattern from whatever you set in front of it, so the cases you pick quietly decide what it thinks the job even is.

Show it three easy, obvious dismissals and it learns an easy, obvious rule — “a sound means out.” Fine, until a tricky one comes along: a faint noise when the bat is right next to the pad. Was that an edge, or just bat brushing pad? The model that only saw clean edges will call it out, and be wrong.

The fix is to choose examples that mark the line between out and not out: include a case that went the other way (a bat-pad given not out) and a genuinely close one. Now the model has seen where the line sits, not just the easy middle — and it reads the tricky delivery correctly.

Same number of examples either way. The difference is entirely in the choosing.

Show the umpire each set of three, then watch how it rules the same tricky appeal. Only one set has seen the close calls.

Two sets, three examples each. Pick one and see how it rules the tricky appeal.
examples shown3 examples
↓ then the tricky one ↓
the tricky appeal · a bat-pad
↓
the umpire's call

Same count, different examples.

Both prompts hold three examples. Swapping which three — not how many — is what flips the tricky call from wrong to right. 


      

    


    

      
It's which, not how many.

      

        Three easy examples teach an easy rule, and the model breaks on the first tricky
        case. Three well-chosen examples — one that crosses the line, one that sits right
        on it — teach the model where the line falls, and it holds up. Same count, very
        different behaviour.
      

      

        Choose examples that mark the close calls, not just the obvious middle.
        A contrasting pair — one each side of the line — is worth more than a pile of
        look-alikes.
      

      

        Next: bad calls →
        
      

    


  


  

    

      ▸ Go deeper — choosing examples well optional
      

        
Cover the close calls, not the centre

        

          A model generalises from what it is shown. If every example is a clear-cut case,
          it never learns where “clear-cut” ends. Deliberately include cases near the line
          — the ones a person would have to think about — because those are the ones it
          will get wrong otherwise.
        

        
Contrast teaches the rule

        

          A pair that goes opposite ways on a small difference (edge vs bat-pad) teaches
          far more than two examples with the same verdict. The contrast is the lesson: it
          shows the model which detail actually decides the call.
        

        
Diversity beats volume

        

          This is why the last page's rule holds: a few varied, borderline examples
          beat many similar ones. When you are tempted to add a fourth example, ask
          whether it shows the model something new — or just repeats a case it has
          already seen.
        

        
Test on the cases you fear

        

          Pick your examples, then try the prompt on the awkward inputs, not the easy
          ones. Where it fails tells you which example is missing. This is the same instinct the course returns to much later, when it asks how you'd measure whether an umpire is any good at all.