The model's record

Picks are recorded before the deadline, or not at all. A scheduled job writes the round's file inside the day and a half before the lockout, and stores both times — when the picks were written and when the round locked — so the gap is on the page. The job refuses to write after a lockout has passed and refuses to overwrite a round that already exists. There is no flag that turns either refusal off. A missing round therefore means the job did not run or the feed was down; it can never be filled in later.

The picks are the ones the site showed. The job imports the same model the dashboard runs — the same file, not a copy — so the record grades what visitors were actually told.

Points come from a subtraction. The official game publishes a player's season total, not a per-round figure. A round's points are the difference between the total stored at its lockout and the total after it finished. That difference is only this round if nothing else was played in between, so the window closes the moment a later round kicks off. A round graded outside it is marked ambiguous and left out of the season figures — a number that might be two rounds added together is worse than a gap, because a gap is visibly a gap.

The captain is graded without a multiplier. This app has never verified what the official captain multiplier is, so nothing here assumes one. What is measured instead: where the armband landed inside our own seven, and how many points a perfect armband would have added per extra multiple. Whatever the multiplier turns out to be, that gap is what the decision cost.

Clubs are graded on official points when the feed publishes them, and on the match result — win, draw, loss, goals, clean sheet — when it does not. Which basis was used is recorded per round and shown, because averaging the two would be meaningless. The official club scoring tariff is not published to this app and this page will not invent one.

The baselines are decisions somebody could have made. "Our seven scored 41" says nothing on its own. The random seven is the floor; the naive seven — best points per appearance at the time, chosen before the round, never in hindsight — is the bar that actually matters; the best possible seven is the ceiling nobody reaches. The headline figure is the share of random sevens ours beat, because it survives a week in which nobody scored.

The rank correlation is the model's real exam. Seven picks is seven data points a week. The correlation between the rating this site gave every player before the round and what they actually scored is about 1,800 data points a week, and it grades the model rather than the squad rules wrapped around it. It is a correlation, not an accuracy: the rating is a 0-100 ranking, not a points forecast, so only the order it puts players in can be right or wrong.

None of this makes a rating a prediction. A model that beats these baselines over a season is a model worth listening to alongside what you know; it is not one that knows what is going to happen on Saturday.