Draft night, with the numbers already done.

On the clock

Four numbers, on one screen, updating as the room picks.

What the official game will not tell you

The game gives you last season's points. Points are a record of what happened, including the parts that will not happen again. These are the corrections, and each one is a mechanism rather than an opinion.

• Value over replacement, not raw totals A defender's 140 points and a striker's 140 points are not the same asset. Every player is scored against the man you would otherwise be forced to start at his position.
• Tiers, and an honest blank instead of a tier Players fall into tiers inside their own position. Anyone with under 270 minutes of Premier League evidence gets a question mark instead of a grade, because there is nothing there to grade.
• Bonus re-simulated under the new season's scoring Last season's bonus is re-run fixture by fixture under the confirmed change to defensive actions, which now score one point of BPS per three rather than per two. Changes that depend on data the public API never publishes are left alone rather than guessed, and if the re-simulation cannot reproduce the bonus that was actually awarded, the whole correction is thrown away rather than shipped.
• Reclassified players re-scored, not carried over When the game moves a player from defender to midfielder, his last season is re-scored at the new position's rates for goals, clean sheets and goals conceded. A keeper moving outfield is left blank, because that one cannot be done honestly.
• Finishing luck, separated from finishing Goals and assists set against the quality of the chances behind them. A striker well ahead of his underlying numbers is flagged HOT and a striker well behind is flagged COLD. It is a flag you read, never an input to the ranking, and the page says which.
• Injury flags that expire when the injury does A flag with a return date on or before the first deadline costs the player nothing at all. A return mid-season is discounted only for the part of the season he actually misses, so a stale February knock does not quietly bury a fit player in August.
• Departed players locked out, and still visible Anyone who has left the league is excluded from every pool the tool picks from. They stay on the board, struck through and marked, so you can see they were excluded rather than quietly dropped.
• Blank beats a guess When a number cannot be computed honestly, the field is empty rather than zero, and the player carries a short note saying why it deviates. A zero and a missing measurement are different facts and this tool never merges them.

News that is filtered, matched, and bounded

Every companion shows you headlines. The question worth asking is what a headline is allowed to do to a number, and here the answer is written into the schema rather than left to judgement.

• A headline cannot move a projection more than 20 percent up, ever Reports reach the projection as a single multiplier that the output schema pins to a fixed band. A report that is validated out of range is rejected outright, not quietly clipped. The ceiling is not the model choosing to be careful, it is arithmetic it cannot reach past.
• A headline can never override the official availability data If the first-party API says injured, suspended or unavailable, that player's projection is already zero, and zero multiplied by any report is still zero. No rumour resurrects a player the game itself has ruled out.
• Opinion is kept away from the numbers on purpose Community posts are read on the news page, where you apply your own judgement, which is exactly where an opinion belongs. They are barred by name from the one agent that moves your projections, because their headlines are opinions about the world rather than reports of it.
• Five feeds, and one broken feed cannot take down the rest A pair of national broadcasters, a fantasy specialist, a transfer aggregator and a community forum. Each is fetched inside its own guard and reports its own failure, so a dead source costs you that source and nothing else.
• Names are matched conservatively, and misses are the cheaper error Word boundaries, a capitalisation check, a minimum length so short surnames cannot substring-match, and a guard for a surname followed by somebody else's first name. Every rule errs toward missing a mention rather than pinning it on the wrong player.
• A relevance gate drops what is not Premier League football The cricket, the boxing, the Scottish and continental leagues that happen to share a surname with somebody in your squad. Filtered items stay readable behind a toggle rather than being silently destroyed.

Then twenty-nine more gameweeks

Draft night is one evening. The season is where leagues are actually lost, and it is lost slowly, one defensible-looking waiver at a time. This is the weekly brief. It is not in the Friends Edition yet.

• Your whole roster gets looked at, not just the obvious problems Every player in your squad is set against the best free agent at his position and given a verdict: upgrade, marginal, or hold. Candidates are taken two per drop slot before any global shortlist, so your bench cannot be crowded out by two exciting names competing for the same spot.
• Ranked claims, with a fallback where a real one exists A fallback has to clear a higher bar than the primary did to be worth naming. Where nothing clears it, the brief tells you to hold and not waiver if the claim is sniped, rather than inventing a second-best to fill the row.
• A submission queue with the odds each claim actually lands Ordered by your waiver priority, the round each claim processes in, and a prediction of what each rival's own roster will push them to go for. A fallback only counts for the times the claim above it fails, so the ordering is worth something rather than decorative.
• It grades its own past advice Once a gameweek is audited, every recommendation is re-scored on what the player you claimed actually scored against what the player you dropped actually scored. That produces a season record and a hit rate broken out by the confidence the tool stated at the time, and when a confidence tier has been over-promising lately the brief shows the realised rate next to the claimed one.
• Trades and opposition, scored both ways Straight swaps and packages that improve both teams' best legal elevens, not just yours, with a rough estimate of how likely each rival is to say yes, taken from how they have actually traded. Nobody is at 100 percent and nobody is at zero.

Why this and not one of the others

Anyone can put a projection on a page. There is no way to tell a good one from a confident one by looking, which is why every companion looks equally sure of itself. So here is the only axis that cannot be copied: this project publishes where its own model has been wrong, and kills its own work when the evidence says to.

• Both cleverer pick formulas written here were retired A smarter way to choose the pick, then a smarter one again, each made to play full drafts against the plain rule already shipping: twenty seeded drafts from all eight seats. Both lost from every seat, so both were retired and the simple rule kept the page.
• A projection candidate was killed by its own backtest The pass bar was written down before the run. Every point estimate pointed the right way and the confidence interval would not clear zero, so it was killed at the bar it had agreed to. Promoting it on the friendlier secondary reading would have been exactly the cherry-pick the rule existed to prevent. Exactly one candidate has ever been promoted, and it is on probation until it repeats on real gameweeks.
• An over-claim of ours is corrected in the ledger, in place A performance number this project had published was cut by roughly half when a stricter replay showed the earlier figure was inflated by availability the original test could not see. The old wording is kept, struck through rather than deleted, because the wording is the mistake worth remembering.

None of that is modesty. A tool that has never published a loss has either never checked or has decided not to tell you, and there is no third option.

What you get today

The Friends Edition is a real slice of this tool, not all of it. Here is the line, drawn plainly, so you never find it by clicking something that is not there.