InternalAI pipeline eval

How right is the AI?

7 recordings, 824 clips with a human verdict. Last experiment logged Sep 11, 2026.

By recording

  1. Solo shooting, 15 June

    Solo

    Jun 15, 2026 · 12:21 of footage · 75 candidates

    Open review
    75 / 75 reviewed100%
    Real shots among proposals
    100%n=75
    Made/miss accuracy
    93%n=75

    made as positive:Precision 86%Recall 100%F1 93%

    Shooter agreement
    not yet
    Review speed
    7.4sper shot
  2. Solo shooting, 3 June

    Solo

    Jun 3, 2026 · 31:17 of footage · 238 candidates

    Open review
    238 / 238 reviewed100%
    Real shots among proposals
    84%n=237
    Made/miss accuracy
    89%n=198

    made as positive:Precision 79%Recall 96%F1 86%

    Shooter agreement
    not yet
    Review speed
    6.7sper shot
  3. 2v2, Hunters Hill, 18 Jan

    Small-sided

    Jan 18, 2026 · 18:44 of footage · 83 candidates · 4 filtered

    Open review
    83 / 83 reviewed100%
    Real shots among proposals
    77%n=82
    Made/miss accuracy
    87%n=63

    made as positive:Precision 76%Recall 96%F1 85%

    Shooter agreement
    96%n=49
    Review speed
    8.3sper shot
  4. Half-court pickup, 29 Mar

    Pickup

    Mar 29, 2026 · 55:03 of footage · 367 candidates

    Open review
    238 / 367 reviewed65%
    Real shots among proposals
    61%n=232
    Made/miss accuracy
    93%n=142

    made as positive:Precision 88%Recall 97%F1 92%

    Shooter agreement
    not yet
    Review speed
    8.4sper shot

Over time

226 logged runs from the experiment ledger. Tap a dot for the method and cost.

  • game-eltham-1
  • game_hd 2v2
  • Solo shooting, 15 June
  • other

Shot recall

30 runs
0%50%100%AprJunAug96.6%

Shot precision

30 runs
0%50%100%AprJunAug53.8%

Shot F1

29 runs
0%50%100%AprJunAug69.1%

Accuracy

86 runs
0%50%100%AprJunAug100.0%

Made/miss accuracy

1 run
0%50%100%AprJunAug76.9%

Candidate recall (5s)

18 runs
0%50%100%AprJunAug0.0%

CV-A accuracy

30 runs
0%50%100%AprJunAug82.8%

Made/miss AUC

2 runs
0%50%100%AprJunAug66.4%

How to read this

A clip's human verdict is the strict majority of its reviewers; ties count as split and drop out of every rate. Shot precision is the share of reviewed candidates that were real shots. Made/miss numbers compare the AI's call with that verdict, with "made" as the positive class, so precision is "when the AI says made, was it" and recall is "of the real makes, how many did it call". Recall of shots themselves is only measured where a recording has an audit game, a continuous interval cut into windows that a human labelled end to end; the other recordings show it as unmeasured.