How accurate is TencolineAI?
On one whole match checked by hand, TencolineAI found 357 of the 359 shots (99.4 %), wrote 3 marks where there was no shot to mark, and named the shot type right on 356 of the 357 it found. It is one match, checked by the maker of Tencoline and not tested blind.
TencolineAI is part of Premium, which is not on sale yet. On the free plan you mark the shots by hand.
The results, measured on 25/10/2026
| What was measured | Count | Share |
|---|---|---|
| Shots found | 357 of 359 | 99.4 % |
| Marks that are a shot | 357 of 360 | 99.2 % |
| Shot type named right, of the shots found | 356 of 357 | 99.7 % |
| Player right, of the shots found | 357 of 357 | 100.0 % |
| End-of-point marks kept | 104 of 108 | 96.3 % |
| How a point ended (winner, error) | Not named yet | - |
For each shot type
| Shot | In the match | Found | Named right | Extra marks |
|---|---|---|---|---|
| First serve | 58 | 58 | 57 | 0 |
| Second serve | 51 | 51 | 51 | 0 |
| Return | 81 | 81 | 81 | 0 |
| Forehand | 124 | 124 | 124 | 3 |
| Backhand | 37 | 36 | 36 | 0 |
| Smash | 7 | 7 | 7 | 0 |
| Backhand volley | 1 | 0 | 0 | 0 |
| All shots | 359 | 357 | 356 | 3 |
Two found shots had their time moved by the person who checked, the larger by 0.57 s. Of the 3 extra marks, 2 were at a moment with no shot and 1 marked one shot a second time.
How was the accuracy measured?
TencolineAI marked a whole recorded match with the model that ran in production on 25/10/2026. A person then checked every mark against the video: removed each mark with no shot, added each missed shot, and corrected names and times. A shot in both lists counts as found, and a mark only the engine wrote counts as extra.
Where do the true values come from?
From the checked list, not from a list marked blind. The person who checked it is the maker of Tencoline. Checking a list is easier than marking from nothing: a time that looks right is kept, and a missed shot is found only by someone who looks for it. A blind test gives a stricter number.
How large is the test set?
One match: 359 shots and 360 engine marks in 66 minutes of play. That shows where the engine fails, but it is too small for an exact rate: one more missed shot would move the share of shots found by about 0.3 points. Each new match goes on this page as a new row with its own date.
What does TencolineAI do badly?
It missed 2 shots, one of them the only volley in the match, and it wrote 3 extra marks. It named one first serve as a second serve. It does not yet say how a point ended, as a winner or an error, and it never calls a line: it reads the video, and it has no radar.
It also wrote 108 end-of-point marks, and 4 of them were removed in the check. Telling how a point ended is the next part of the engine, and this page will measure it the same way when it ships.
Can I compare these numbers with another app?
Only when the method is the same. A share of shots found depends on the test set, on who made the true values, and on how close a mark must be to count. Ask how many shots were in the test, where the true values came from, and if it was marked blind. This page answers all three.