Notes · 21 September 2026
The week we lost the record
There is no Gameweek 5 scorecard on this site, and there is not going to be one. Building it today would take about ten minutes. That is exactly the problem.
The point of this site is a simple one: write the projections down before the deadline, leave them alone, and score them afterwards whether they were any good or not. That only means something if the writing-down genuinely happened first. For Gameweek 5, it did not.
What is missing, and what is not
Gameweek 5 was played. Its results exist, they are in Fantasy Premier League like every other week, and they feed this site's model exactly as they should. The player pages, the projections, the price data, the fixture grid — none of it is affected, and nothing about that week's football has been lost.
What does not exist is the list of what this model predicted before the deadline. That is a different object from the results, and it is the only one that cannot be produced later, because anyone producing it now would be writing with the answers already in front of them.
So Gameweek 5 can still be analysed. It cannot be used to score the model, because there is nothing on record to score it against.
What was supposed to happen
The Gameweek 5 deadline was 18 September, 18:30 UK time. A job was scheduled to capture the model's top twenty two hours before that, at 16:30. It was scheduled on 16 September, for the correct minute, and then it was left to run.
What happened instead
It never ran. The capture depended on a particular computer being awake and reachable at that minute, and it was not. The job waited, the window it was allowed to wait in expired, and it eventually fired at 08:52 on 21 September — three days after the deadline, with the gameweek played and every result known.
By then the site was serving projections for Gameweek 6 and the model had already been refitted on Gameweek 5's results. There was nothing left to capture that could honestly be called a prediction.
A correction
The Gameweek 4 review ends with the sentence "Gameweek 5's projections are already locked." That was written on 16 September, when the job had just been scheduled and there was no particular reason to doubt it. It was wrong. It has been corrected in place rather than quietly deleted, because a site that edits its own claims without saying so is worse than one that gets things wrong.
Why it is not being backfilled
This is the part worth spelling out, because the temptation is real and the excuse is easy. Nobody would know. The model's Gameweek 5 numbers could be reconstructed to a reasonable approximation, formatted into the usual table, dated 18 September and published. It would look exactly like the Gameweek 4 file.
And that is the reason not to do it. A reconstructed week does not damage itself — it damages every other week beside it, because from the outside there is no way to tell which is which. The Gameweek 4 list only means something if you believe it was written before the matches. One fabricated entry in the folder and you have no reason to believe any of them.
A missing week costs one data point. A fabricated week costs the entire record. That trade is not close, and anyone keeping a public prediction record who tells you otherwise is keeping a different kind of record.
What has been rebuilt
Two things were wrong, and both have been fixed by the same change.
The first was that the capture ran on a machine that could be asleep. It now runs on the server that serves this site, triggered two hours before the deadline, with a second independent trigger on the site's own infrastructure as a backstop. Neither depends on anyone being awake, at a desk, or online.
The second was subtler and mattered more. The old scorecard was produced by something that read the projections and formatted a table — and anything that can format a table can format a different table. The file is now generated by the server itself, from the same model that produces every number on this site. The trigger decides only when to capture. It has no way to influence what gets captured.
Three rules are now enforced rather than intended. A capture after the deadline is refused outright. A gameweek can be written once, and the second attempt fails in storage rather than relying on anyone to remember. And every file records how long before the deadline it was taken, because a list locked twelve hours out is a harder test than one locked at two hours and you deserve to know which one you are reading.
The files themselves are now public at the scorecard page, served from storage that anyone can fetch without going through this site at all.
What it costs
The running record skips from Gameweek 4 to Gameweek 6. That is stated plainly wherever the numbering appears, so a gap is never mistaken for a lost file.
The open question from Gameweek 4 — whether the model's captain pick can beat the most-captained player in the game — is still sitting on a sample of one, and a bad one, since Isak returned 2 while the crowd's Haaland returned 9. It should have been two samples by now. Gameweek 6 becomes the second.
Next
The Gameweek 6 deadline is 11:00 on 10 October. The capture will happen at 09:00, without anyone touching anything, and it will appear on the scorecard page the moment it does.
If it does not, that will be written up too.