Next to score modification suggest
I have one suggestion after testing the Who Scores Next signal.
Using Fenerbahçe vs Sturm Graz (match ID 223286) as an example, the first signal was generated after only 2 minutes of play. At that stage there isn't enough match data yet, so the prediction is based on very limited information.
I think it would be better to introduce a minimum amount of match data before generating the first signal (for example waiting 5-10 minutes, or until a minimum number of events has been recorded).
I would also recommend adding a cooldown after every goal. In this match Fenerbahçe scored in the 8th minute, and the system immediately generated a new "Who Scores Next" prediction. In my opinion, a few minutes should pass after a goal before calculating the next signal, allowing the match to stabilize and the statistics to reflect the new game state.
This could reduce noisy signals and make the predictions more reliable.
Comments
1
Bzzoiro · Admin
Aug 06, 2026 21:48
Good catch, and better than you knew - the first half of this was a bug, not a missing feature.
The minimum you asked for already existed. There is a floor of ten minutes before the model is allowed to speak, with a comment in the code saying exactly that. It was measured wrong: the check counted how many momentum readings we held rather than what match minute those readings reached. The feed can deliver several readings for a single minute, so the count crossed ten while the match was two minutes old. Your example is the clearest evidence of it - the call is timestamped 18:04:08 against an 18:00 kick-off. The guard believed it was holding ten minutes of evidence and was holding four.
Fixed and live. It now tests the match minute. Of the 22 calls published so far, 3 were below minute 10 and would now be held back, including the one you saw.
On the post-goal cooldown: you are right that there isn't one, and I am not adding it yet. I want to explain why rather than just say no.
Half of your reasoning does not apply, and half of it does. The model reads the current score, so after a goal it is not working from a stale picture of the game - that part is fine. But it also uses rolling averages of momentum, and those genuinely do take a few minutes to reflect that the match has changed character. So the instinct is sound.
What stops me is that I have 22 resolved calls in total. That is not enough to tell whether the signal is too eager or not eager enough, and adding a second gate on top of a floor that was broken until today would mean tuning against noise I have just removed the source of. The sensible order is: fix the floor, let it run, then look at whether calls immediately after a goal actually underperform. If they do, the cooldown goes in with a number behind it instead of a guess.
I would rather tell you that than quietly implement it and imply it was measured.
Leaving this open until I have enough calls to judge the second half. Thanks for testing it properly - noting the exact match and the exact minutes is what made this findable.
- Bzzoiro
Log in to comment.