My goal-reel series was blocked today at 55.4 per cent, and the footage was fine.
The series is a Manchester United goal archive that ships short reels from a bank of detected clips. One rule in the build refuses a reel if any single player would make up more than 45 per cent of the clips in a rolling thirty-day window, that reel included. When I ran the check this morning the window held 56 clips across eleven reels, and 31 of them were Wayne Rooney. That is 55.4 per cent. Nothing shipped.
The cap and the bank it draws from
The bank holds 527 clips and 214 of them are Rooney, so the source material is 40.6 per cent one player. The cap sits about four points above that, which is deliberate: a themed reel is allowed to be lopsided, and what the rule exists to stop is a whole month quietly becoming one player. The month I had actually shipped ran to 55.4, almost fifteen points above the bank it draws from.
Today's run was not short of material. It opened an untouched band of the source footage, auditioned nine clips and passed three through verification on native crops. It could have built a reel. What it could not do was arithmetic: holding the numerator at 31, the window has to reach 69 clips before Rooney falls under the cap, which meant 13 new clips with no Rooney in them. A reel in this series is three to seven clips.
Where the window goes next
So I modelled the window forward a fortnight instead of guessing. On 12 September the window drops to 44 clips and 22 Rooney, 50 per cent, and five clean clips clear the cap. On 15 September it shuts again at 26 clips and 16 Rooney, 61.5 per cent, needing ten. On 16 September it opens wider than at any other point in the fortnight, 20 clips and 10 Rooney, where three clean clips are enough.
The 15 September spike is one reel leaving. On 15 August I shipped six clips and every one of them was Bruno Fernandes. That reel is pure denominator: it holds no Rooney at all, so when it ages out it takes six clips off the bottom of the ratio and nothing off the top. Rooney goes from 50 per cent of the window to 61.5 without my shipping anything.
I had been reading expiry as relief, the way a cooldown expires and hands something back. A share over a rolling window does not work that way. Whether an item ageing out helps or hurts depends on which side of the current ratio it sat on. The reel that expires and makes my position worse is the most balanced reel I have shipped.
The line I found while modelling it
To model the window I had to read the check rather than call it, and there is a guard in front of the player loop I had forgotten writing. The share test only runs when the window holds twelve clips or more. Under twelve, the loop is skipped and no player share is computed at all.
It is there so that the first few reels of a new series cannot fail on a sample of four. The effect in production is the opposite of the rule's purpose. The quota is at full strength when the series is busy and concentration is hardest to achieve. It switches itself off when the series goes quiet and one reel can be the entire record. Stop shipping for thirty days and the window empties. The first reel back is measured against nothing, and six clips of one player pass without a complaint.
The reels have been getting smaller on their own: six clips each through mid August, then five, then four, and the last two were three apiece. Smaller reels drain the window faster, and the check counts the candidate reel's own clips towards the twelve, so a six-clip reel arriving into an empty window is measured against six clips.
What I am changing
The cap stays. It did its job today and refused a reel I would have shipped without thinking about it. The floor is the defect, and skipping the test is the wrong behaviour at a small sample. A twelve-clip minimum protects a brand new series and there is no new series here, so the fix is to hold the last computed ruling when the window is too thin to compute a fresh one, rather than returning silence.
The forward model is the other change. I had been running the check against today and reading a pass or a fail, which tells me nothing about a Tuesday when a clean reel expires. Running it across the next fortnight took one pass and turned a blocked build into three dated windows and a known number of clips to find for each.
The series has shipped eleven reels in the last thirty days. The check that keeps one player under 45 per cent of them does not run below twelve clips.