The Bean Counter · Leeds United

Twenty-six sold players were sitting in “Retained” columns across fifteen clubs SITE-WIDE CORRECTION

Villa’s board was not the exception. It was the one somebody happened to be looking at. After Emiliano Martínez was found under “Retained” four days after his sale to Chelsea, the same pattern was checked across the other nineteen pages. Every club page carries these position boards — three columns each, who left, who stayed, who arrived — and at roughly forty-five chips a page that is about nine hundred hand-written claims, none of which any gate on this site could read, because they name players in free HTML rather than through a data array.

Twenty-six of them said a player had stayed when the club’s own deal table said he had been sold. The list is not marginal. Rodri sat under Manchester City’s “Retained · 6” tagged “REAL MADRID TALKS”; he had gone to Barcelona for £51m. Tijjani Reijnders (£52m, Al Qadsiah) and Nico González (£50m, Newcastle) were beside him. Carlos Baleba was retained at Brighton after a £65m move to Manchester United. Liam Delap was retained at Chelsea after £45m to Forest. Everton had five in one page — Ndiaye, McNeil, Beto, Iroegbunam and Patterson. Arsenal had Gabriel Jesus, Ipswich three, Hull two, Palace two, and so on across fifteen clubs.

Each one is now in the column it belongs in, with the destination and fee taken from the deal table rather than retyped. Column counts were recomputed from the chips actually present, so the headings match their own contents for the first time in a while. “Out · 0 — Nobody” placeholders were removed where a real departure now fills the column.

Worth recording how nearly this went wrong, because it is the second time today. The first script moved the chips and reported twenty-two successes. It had in fact deleted all twenty-two players: the insertion point never matched, the removal did, and the counts were recomputed afterwards so every column looked internally consistent. Nothing in the output said anything was missing. It was caught by reading one board back rather than by trusting the tally, and reverted. The rewrite asserts the set of names in a board before and after the edit and refuses to write when they differ — which then immediately caught a second, smaller error, and later a third: the pattern used to find a column was swallowing the last chip’s closing tag, so the final name in every column was invisible, which is why four clubs looked clean on the first pass and were not.

And a gate, the twenty-fifth. tools/check-flowboards.js asserts two things about every chip on every board: a name in a retained column must not be a completed permanent sale in that page’s OUT26, and a name in a retained column must exist in that page’s data at all. It says nothing about incoming or outgoing columns, where naming a target from another club is correct, and it exempts chips whose own tag places them outside the squad table — a third-choice keeper in contract talks, a U21 needing no registration. Its first run reported seventy-one problems, of which twenty-four were its own fault: typographic apostrophes it could not match to straight ones, and editorial labels like “Third keeper slot” that are not players. Tested against the pre-fix pages, where it finds exactly twenty-six.

The reason this sat unseen is worth stating plainly. The deal arrays are gated. The ratio inputs are gated. The squad table is gated. The answer block, the league tables and the share cards are all derived. These boards were the one substantial thing on every club page that was neither derived nor checked — prose in the shape of a table, in a heavily-read tab — and they had been drifting for as long as they have existed. It took a reader with the live page open to find the first one.

The Saudi deadline was five weeks out, and it is four days CORRECTION

This morning’s work put the Saudi Pro League window at 12 October and built a countdown, a gate and seven card tags on it. A reader said that sounded wrong — they had read it was days away. They were right.

The sources disagree. Wikipedia and Goal put the registration period at 22 July to 6 September; one tracker says 11 September; Sports Mole and AgentFlow say 12 October. The tell is arithmetic: 22 July to 12 October is 11.7 weeks, essentially FIFA’s twelve-week maximum for a registration period. That reads like somebody published the longest period permissible rather than the date actually announced. 6 September is 6.6 weeks, an ordinary window.

It is now carried at 6 September — the earliest credible date, chosen deliberately. The conservative error here is the early one: saying a shut window is open invents deals, which is the exact failure the window gate was written this morning to prevent. The countdown now reads four days rather than forty.

What is still unresolved: no source reachable from here separates Saudi domestic registration from international, so no such split is modelled. If one exists it is the international deadline that governs a sale out of England, and 12 October may well be a domestic-only extension rather than simply wrong. The date is flagged as contested on the deadline page rather than asserted.

Worth naming the pattern: the 12 October figure came from a single search result, and the source family it came from has now been wrong four times in this project — Hincapié’s fee, Trafford’s club, Fofana’s club, and this. A date that anchors a gate deserved a second source before it was written into twenty-one pages.

The window is shut, and the site now says so — except where it isn’t CORRECTION

The Premier League window closed at 11pm on 1 September, so from the 2nd no rumour card here can describe a signing as something still pending: a registration cannot be lodged outside the window, however well sourced the link was. 139 cards are now marked JANUARY and priced as January deals rather than deals that can be done today.

But “the window” is not one thing, and flattening it would be the opposite error. A club that cannot buy can still sell: the Saudi Pro League runs to 6 September and the Süper Lig to the 4th; MLS closed on the 2nd. Bundesliga clubs finished earliest, on 31 August; Serie A, LaLiga and the Premier League all shut on the 1st. Six cards survive because they point at a league that outlasted England’s, and each has had its destination narrowed to the club that can actually still act. The seventh has already gone. Leon Bailey, whose card read “Al Diriyah” after Sevilla and Hull ran out of time, signed for Olympiacos on 2 September for £3.4m — which is the point of this section in one transfer. The English window shutting does not stop an English club selling, and it did not stop this one. It also proves a window this table does not list: Greece was still open on 2 September, and no closing date for it could be sourced, so none is stored rather than guessed.

The first countdown clock had been reading DONE against a label saying “days to window close”. It now counts to the Saudi deadline on 6 September — the one transfer deadline that still constrains these clubs, because they can still sell into it.

tools/check-window-state.js holds it: a card with no jan flag is asserting the deal can happen now, so its destination must sit in a window open today. It is date-driven rather than a one-off sweep — on 5 September the Türkiye cards stop being live and the gate says so without anyone remembering. It found six on Fulham that the first pass missed, because those rows are written with double quotes where the rest of the site uses single.

The countdown gate had a matching blind spot: it skipped any page without a club nav and only knew one label shape, so deadline/ was invisible to it — and was still counting to 1am on 1 October, the exact GMT-versus-BST bug that gate was written for, fixed on twenty pages and left on the twenty-first. It now reads both shapes, both const and var, and every page that carries a clock.

Jack Harrison: an open question, now closed DATA

This page said no reachable source settled what happened to Jack Harrison after his loan return from Everton. One does: he joined New England Revolution permanently in July 2026 on a Designated Player contract to 2029/30.

The fee was not disclosed, so it is carried at zero and no profit is claimed. The deal moves the wage off the numerator and nothing else can honestly be modelled from it. His £11m from Manchester City in 2021 was fully written down before this window opened, so there was no book value left to sell against either way.

The note on his 2025/26 row, which had flagged the question as open, now points at the answer.

Gelhardt was sold in August, and this page still had him as a Leeds asset CORRECTION

Leeds sold him to Hull on 14 August for £4.5m guaranteed, up to £6.5m — and Hull loaned him straight back, so he has been at Elland Road all along and the sale went unrecorded. He now appears in both directions, correctly: sold in the Outs table, borrowed in the Ins.

His squad row carried a £0.9m book value that is no longer Leeds’. He is Hull’s asset; Leeds have the player, the wage and one of their 25 registrations, and none of the amortisation. Academy product, so the whole guaranteed fee is profit — sales banked £16.1m → £20.6m, ratio 68.9% → 68.4%, headroom £45.2m.

It needed a gate to learn a new shape. Two invariants say a sold player must be out of the squad, which is right almost always and wrong here. The exemption added is deliberately narrow: the squad row has to be flagged as a loan in and carry a zero fee. A sold player still sitting there with a fee on him — the error those checks exist for — still fails, which was verified by putting the £0.9m back and watching both fire.

Elvedi’s fee was about half the real number CORRECTION

This page carried Nico Elvedi at £4.3m, taken from an early ~€5m report. The settled agreement is €10m — about £8.5m, with Borussia Mönchengladbach also keeping a 15% sell-on, and his contract runs to 2029 rather than the three years this row stated.

Committed spend goes £78.4m → £82.6m and the window swing to +£24.0m/yr. The ratio holds at 68.9% against the 85% limit, so nothing turns on it here — but a fee that is out by half is out by half whether or not it crosses a line.

The sell-on is a liability rather than an asset, and an unusual one on this page: 15% of whatever Leeds eventually receive for him goes back to Gladbach. It is not modelled, because it is contingent on a sale that has not happened, but it is recorded so that any future Elvedi profit is read net rather than gross.

Perri is official, the buy option was £2.8m light, and the loan was not free OFFICIAL + CORRECTION

Torino and Leeds have both announced it, so the row moves from advanced to done. Two figures on it were wrong while it sat there.

The option was carried at ~£9.4m. Torino’s deal is reported at €14m, about £12.2m, with Perri having pre-agreed a three-year contract and a further year if it is taken up.

And the loan is not free. Torino pay a €1m (~£0.9m) loan fee, which this page had at zero. A loan fee is income, not cost relief, so it sits on the other side of the ratio from the wage rather than being netted into the relief figure. It is too small to move the headline, and it is recorded anyway because this site’s standing claim has been that none of its outgoing loans carries a disclosed fee. That is now one exception, and an unrecorded exception is how a claim becomes false.

The wage moves for the season; the £13.9m fee keeps amortising at Elland Road unless the option is exercised, because a loan never frees the amortisation.

Three pages had the window shutting four hours early CORRECTION

Hull, Coventry and Newcastle counted down to 7pm on 1 September. The window shuts at 11pm, which is what the other seventeen pages said and what their clocks ran to. On the eve of deadline day, three club pages were telling readers they had four hours less than they did.

Both halves were wrong together, which is why nothing looked odd on those pages: the label said 7PM and the clock counted to 19:00, so they agreed with each other and disagreed with the rest of the site. The first fix here only corrected the labels, and the new gate immediately failed for the right reason, catching a symptom fixed without its cause.

Also tidied: a squad-list deadline written as “after the window closes (~2 September)”, where the date belongs to the list and the parenthetical made it read as hedging on the window itself. And the Premier League checkpoint countdown was set to midnight UTC on 1 October, which is one o’clock in the morning here, because it was stored as though Britain were on GMT in October rather than BST.

tools/check-countdowns.js now renders each stored target through Europe/London and compares it to the label written beside it. Timezone is the whole difficulty: 22:00 UTC is 11pm in September and would not be in December, so the two can only be compared after conversion. Broken on purpose in both directions before being trusted, once with a wrong label against a right clock and once the other way about.

This was found because a reader corrected something said in conversation, not by anything that runs. Sixteen gates check this site and none of them read a clock.

The bolding was doing too much work — 2,316 signposts dropped EDITORIAL

A reader said the site reads as though a machine wrote it. That is worth measuring rather than arguing with, so it was measured.

The design came out fine. Claret and blue are the club’s own rather than the cream-and-terracotta that generated pages default to, the display face is Barlow Condensed rather than Inter, corners are three or four pixels rather than fully rounded, and there are five emoji on a 530KB page. Nothing there needed changing.

The prose was the problem, and one habit dominated: a bolded editorial sentence opening almost every paragraph. There were 2,612 of them, roughly one bold span every 62 words across half a million words of text. Bold is for things a reader scans for — a fee, a name, a date, a verdict. Used on the thesis of every paragraph it stops marking anything at all, and it is one of the most recognisable tells there is.

2,316 of those have gone. Figures, names, labels and short verdicts keep theirs. No wording changed anywhere, only tags, so nothing about the analysis moved and the whole pass reverts with one command. Density goes from one bold every 62 words to one every 85.

Two things this does not fix, stated plainly because the difference matters. The em-dash rate is 17.3 per thousand words against one to three in edited prose, and that needs a slower pass: a blanket substitution produces comma splices. And underneath both is a rhetorical habit no script reaches — signpost, then aphorism, then antithesis, three times a paragraph. That is the thing the reader actually noticed, and it has to be rewritten by hand.

One note on method: the first attempt stripped the bold from the two sentences that state the medical rule, which tools/check-deal-stage.js holds byte-identical across all twenty pages. The gate failed and the sentences are excluded now. It was written to guard a transfer convention and it caught a typography change, which is the argument for gating wording as well as arithmetic.

Agent fees to 10%, the wage gross-up to 1.55x — both were too low, and the FA publishes the proof MODEL CHANGE

Two constants underneath every figure on this site have moved, at a reader’s prompting. The agent fee charged on a signing goes from 5% of the transfer fee to 10%, and the wage gross-up from 1.45x to 1.55x. Both were estimates that had never been checked against a published source. Both now have one.

The agent rate was checkable all along. The FA publishes what every club pays intermediaries. Between 4 February 2025 and 2 February 2026, Premier League clubs paid £460.3m — against the £3,137m of incoming fees this site records for that window, an implied rate of 14.7%, nearly three times the 5% being charged. Chelsea alone paid £65.1m, Aston Villa £38.4m. 10% is the figure taken, deliberately below the implied ceiling: the FA total also covers sales, loans and contract renewals, and only the buy-side portion is capitalised into a transfer’s cost. It is a floor with evidence under it rather than a rule of thumb.

The gross-up was missing a whole category of pay. It covered employer NI, bonuses and image rights — but not signing-on fees, and UEFA is explicit that they count. Annex G.3 of the Club Licensing and Financial Sustainability Regulations: all forms of consideration to and/or benefit of players, such as sign-on fees, must be treated as employee benefit expenses and not costs of a player’s registration. So they belong in the numerator, they sit with the wage rather than the amortisation, and they spread across the contract as a prepayment for services. Employer NI alone is 15%, so of the old 45% only about 30 points were doing the work of bonuses, image rights and signing-on fees together. 55% gives that last item somewhere to live.

What actually moved. League-wide, the annual cost this window adds rises from £1,257m to £1,330m. Twelve clubs’ wage inputs were re-derived by the bridge and reset to it; no club crossed a limit, and headline ratios moved by at most 0.9 of a point — mostly downwards, which is worth understanding rather than glossing: the gross-up applies to the net wage movement, and most clubs sold more wage than they bought, so a bigger multiplier makes a negative movement more negative. Every projection built on an old baseline was recomputed too, not just the baselines.

One piece of tidying that mattered more than it looks. The gross-up existed as a named constant and then as five hard-coded copies of 1.45 in the same file, on all twenty pages. They all now read the constant. A number duplicated a hundred times across a site is not a convention, it is a hundred chances to change ninety-nine of them.

Nothing below a medical belongs in the Ins & Outs tables RULE CHANGE

Julian Brandt leaves the Ins table — and moving him turned up something worse. He signed for Ajax on a free at the end of July, choosing them over Leeds and Fulham, and this page had carried him as a live target for a month: an “In talks” row in the Ins table and an “OFFER MADE” card in the Rumour Mill, both offering a player who had been someone else’s for weeks. The site already knew. Villa’s page says in passing that Leeds “missed out on Julian Brandt” — the correct fact, one page over, contradicting this one and checked by nothing. His card is now marked dead and off the modeller, and the squad-list board says Ajax. It was found only because the rule change put a human back in front of a row nobody had reason to look at, which is the argument for the rule in one example: a table of settled business gets read, and a row that counts towards nothing does not.

The status ladder ran done → advanced → in talks → offer, and all four rungs were rendered as rows in the deal tables. The bottom two contributed nothing to any figure — every total on this site filters to done and advanced — which made them arithmetically harmless and editorially loud. A reader scanning the Ins table saw rows the club had never agreed to, in the same table, in the same shape, as business that is done. Being excluded from the totals is not visible; being in the table is.

The threshold is now a medical. A deal enters at advanced — terms agreed and the medical taken or booked, awaiting only an announcement — and moves to done when a club announces it. A loan row is an agreed loan. Everything short of that is interest, however well sourced, and interest lives in the Rumour Mill, where it can still be ticked and modelled against the ratio. Nothing has been deleted. A row that counted towards nothing still counts towards nothing; it is just no longer sitting where settled business sits.

Four figures did move, and they moved because they had been wrong. The headline ratios, the swing and the window balance all filter to done and advanced and were untouched. But the amortisation column added to the deal tables sums every row in the table with no filter at all, and writes its total into a footer labelled “TOTALS IN (confirmed/advanced)” — a total that named a filter it did not apply. The same total feeds the “Added by this window’s deals” and “Net change” lines under the run-off chart. So unagreed deals were being amortised into figures that said they excluded them:

• Crystal Palace — added by this window’s deals +£12.4m/yr → +£6.6m/yr, nearly halved, because Le Guen (£12m) and Njie (£16m) were being amortised though neither fee exists; freed −£18.3m/yr → −£12.3m/yr; net −£5.9m → −£5.7m/yr.
• Sunderland — added +£11.2m/yr → +£6.7m/yr and net change +£9.2m → +£4.8m/yr, on an asking price for Poku that no one had agreed to pay.
• Brentford — added +£16.9m/yr → +£16.1m/yr, net +£16.5m → +£15.6m/yr.
• Liverpool — totals out, wage ~£735k/wk → ~£725k/wk and cost freed £61m → £60.2m/yr, for a loan with no agreed destination.

The amortisation code has deliberately not been changed. Its filter is now the whole table, because the whole table is now business at a medical or beyond — and the fix belongs at the input, in one gate, rather than as a twenty-first copy of the same rule pasted into twenty pages of inline script. That is also the answer to why this survived: nothing compared a table’s footer against the label printed on it, and the rows doing the damage were invisible in every figure a reader would have checked.

Twelve rows moved across eight clubs, and the wording moved on all twenty. The ledes described the threshold nine different ways — “medical/verbal”, “option/verbal”, “talks/verbal”, “medical or negotiating” — and three of those promised a table that now excludes what they promised. One sentence states the rule on the Ins & Outs tab and its counterpart on the Rumours tab, identical on every page, and tools/check-deal-stage.js holds both to the letter as well as failing any deal row below the threshold. The gate was broken on purpose twice before being trusted: once by dropping a row’s stage entirely, which is the obvious way round a stage check, and once by adding a season array no hard-coded list would have known about.

The Ins & Outs tab scrolled sideways on a phone, and so did every other tab CORRECTION

One table on this site sat outside a scroll container. The amortisation runoff — Season, Already here, This window, Combined — is 560px wide and had nothing but a margin on its wrapper, while every other table on the page sits inside .twrap with overflow-x:auto. At phone width it stretched the whole document about 200px past the viewport, and because it was the document scrolling rather than a box inside it, every tab slid sideways, not just the one containing the table. All twenty club pages, identically.

It is not new and it was not the Read more button. The version of this page from before that change overflows by exactly the same 597 pixels against a 390-pixel viewport, which is worth stating plainly rather than letting the newest change take the blame for the oldest bug.

Why it went unseen is the more useful part. The browser check that runs against every page loaded it at desktop width and never clicked a tab, so it could not have seen either half of this: not the narrow viewport, and not the tab the table lives on. It now runs at 390px and walks all twelve tabs on all twenty pages, failing on any page whose document scrolls horizontally. That costs real time on each run, which is the honest price of catching a class of fault that only exists on the device most readers use.

The deal notes now fold away behind a button LAYOUT

The note under a player’s name is where the reasoning for a deal lives, and on the busiest rows it runs to several hundred words. Read end to end that is the point of the table; scrolled past on a phone it buries the row underneath and everything after it. Long notes now clamp to two lines with a Read more button, and short ones are left exactly as they were — a note that fits gets no furniture around it. The full text stays in the page whether or not it is open, so it is still searchable, copyable and printable.

It did not work the first time, and the reason is worth writing down. The clamp was styled on a single class, and the table already carried a td .glide-cap rule setting the notes to display as blocks — a more specific selector, so it won. The result was a button that toggled its own label and changed nothing else: the line-clamp property applied, the display property it depends on did not, and the note stayed full height. Checked in a browser at phone width rather than by reading the stylesheet, which is the only way that particular failure shows up.

A sell-on reduces the profit, not just the cash — and the gate that should have said so was reading a different page CORRECTION

Asked whether everything a transfer actually contains was applied to the deals on this site, and one thing was not. A sell-on owed to a former club is a cost of the disposal: it comes off the profit booked, not only off the cash banked. This site deducted it from the window balance and the List A test and then left it out of profit entirely — and profit is what feeds the squad-cost denominator through the three-year average. It now comes off booked profit, so the only quantified clause on the site, Villa’s £15m to Middlesbrough on the £117m Rogers fee, is handled by the model rather than by hand.

The gate was passing green on a £15m disagreement, and the reason is worth recording. The trading bridge’s check pulls the profit function out of the page and runs it — but that function reaches outside its own block for the sell-on data, which the extract did not include. Undefined dependency, guard returns zero, and the checker quietly computed a different profit from the one the page computes. A checker that does not mirror the page is not checking the page. It now pulls its dependencies across too, and it failed immediately when it did.

No figure moves today. The £15m had already been taken off this club’s denominator by hand when the sell-on was first noticed; the anchor gives that third back so the arithmetic lands in the same place, and from here it flows automatically. The three other clauses on the site — Baleba, Reijnders and Chérif — are recorded as unquantified and still contribute nothing, which is honest rather than convenient: the exposure is real and the percentage has never been reported.

Exactly full: 25 of 25, and nobody under 21 NEW

The Premier League’s own squad rule arrives on this page. Until today it carried UEFA’s List A and nothing about the domestic quota — a different rule with different arithmetic: 25 senior places, no more than 17 non-homegrown, under-21s outside the list altogether, and players out on loan not registered at all.

Where this squad stands: 25 of the 25 places used, 13 homegrown and 12 non-homegrown, with 0 under-21s playing free of the list. They can name a full 25, which on this tracker is the minority position.

Homegrown is a registration test, not a passport test, which is why every player now carries a stated basis rather than a bare flag: three entire seasons, or 36 months, at any club affiliated to the FA or the FAW before the end of the season in which he turns 21. Nationality is not the question and repeatedly gives the wrong answer.

Why it was missing until now. The section was built for one club eleven months ago and the commit that shipped it said the other nineteen would follow. They did not, because an absent feature breaks no gate and nothing was watching for absence. Every count here derives from the squad table, and the same headcount gate that holds every other page holds this one.

The sale-profit input now derives itself, like the wage input above it METHOD

in-psp is the three-year average player-trading profit in the squad-cost denominator, and it was hand-set on all twenty pages. The failure mode is the one the wage bridge was built to stop, one input along: a sale completes, its profit is booked in the table, and the figure meant to carry a third of it stays where somebody last typed it. It only ever moved when whoever booked the deal remembered — which is the definition of a number that will eventually be wrong.

The page now shows its working beside the input: the earlier years’ contribution, plus this window’s booked profit divided by three, because UEFA smooths player-trading profit over three years so one window enters at roughly a third. Booked means done and not a loan — an advanced deal has not completed and its profit is not in the accounts, which is also the convention this page was already kept on. The profit is net of the FIFA solidarity taken out of a foreign sale.

One thing about the anchor is worth stating rather than glossing. The earlier years’ figure is carried forward from the number this page already used, not sourced from published accounts, so the bridge agrees with today’s input by construction. It earns its keep from here on, by moving when a deal lands. Tested by planting a completed sale into the tables and leaving the input alone: the gate said the denominator was £20m short, which is exactly the silence it exists to break.

A sale can lose money once FIFA has taken its 5%, and the formatter could not write it down CORRECTION

Modelling solidarity produced the site’s first negative profit on a completed sale. Villa banked ~£8.08m of Lucas Digne’s £8.5m fee against a book value of £8.5m, so that deal is a small loss rather than a break-even — which is the correct answer, and one the gross-fee basis could never produce.

The money formatter could not render it. It put the pound sign first and the minus wherever the number happened to start, so the cell read “£-0.4m” against a site that writes −£0.4m everywhere else. Nineteen pages had that version. Bournemouth had the correct one, negative-aware, written when its own trading profit went below zero — so the fix was to adopt the version already in the codebase rather than invent a twentieth. All twenty now share one formatter.

Checked in a browser rather than by inspection, because the failure this could introduce is a doubled sign where a caller already prepends its own minus: no malformed negative on any of the twenty, and no JavaScript error.

FIFA takes 5% of every fee sold abroad, and this site had been booking 100% of it CORRECTION

Asked what else is inside a completed deal that these numbers miss, and the answer was the solidarity contribution. 5% of any international transfer fee is deducted from the fee and shared among the clubs that trained the player between 12 and 23. The mechanics decide who pays: the buying club hands over the headline number and distributes the 5% out of it, so a purchase costs exactly the fee already capitalised here — nothing was wrong on that side. The selling club banks 95%. Every profit booked on a sale abroad was therefore 5% too high, across all twenty pages, from launch. Around £28m league-wide.

On this page it is £0.06m, from the foreign sales of Isaac Schmidt. That comes off booked profit, and a third of it off the denominator, because sale profit feeds the squad-cost ratio through a three-year average.

The glossary had been claiming this was modelled the whole time. It said the 4% Premier League levy and FIFA’s 5% solidarity contribution both capitalise into registration cost — true of the levy, true of solidarity for a buyer, and quietly untrue of the thing that actually matters, which is what a seller keeps. Documentation that describes a calculation the code does not perform is worse than no documentation, because it stops anyone looking.

It is derived, not typed. The deduction comes from where the player went, which puts all the weight on classifying a destination — and getting that wrong is silent: an English club missing from the list costs its seller 5% of a fee it actually keeps, and no total looks odd. So tools/check-solidarity.js refuses to classify by default. Every fee-carrying destination must be a club one of the two lists already knows; a name in neither fails until a human decides which it is. The gate also holds the twenty inline copies of the classifier to being identical, for the same reason the wage gross-up is a single constant.

One assumption is recorded rather than hidden: Cardiff, Newport, Swansea and Wrexham are Welsh clubs inside the English pyramid, and are treated here as domestic because their registrations run through the English system.

Also documented, deliberately not changed: add-ons are modelled all-in on a purchase and guaranteed-only on a sale. That asymmetry makes the buy column a best case and the sell column a conservative one, and it now says so on the page instead of being visible only to whoever recomputes the fees.

Wages in the deal tables were understated by 45%, and the sentence describing the fix was wrong on nine other pages CORRECTION

The site grosses wages by 1.45 to cover employer NI, bonuses and image rights — everywhere except the two columns readers actually look at. The cost of an arrival and the relief from a departure in the Incomings and Outgoings tables were built from raw weekly wages, so every one of them understated the real charge. Both columns are now grossed through a single GROSS constant, so they cannot come apart again. The headline ratio does not move: it is computed from the wage, amortisation and agent inputs, which were already on the grossed basis. What moves is the Net SCR Swing card and the two columns feeding it.

Which is where the second problem was. The swing is derived from the deal arrays and was therefore right the moment the gross-up landed — but the same figure is also typed into the overview callout, and typed figures do not move. The cost swing in the overview read +£20m/yr against a computed +£21.9m/yr. And the ratio quoted in prose was ~63% against a computed ~69%. Palace’s was the clearest case: it read “still negative” where the page computed a positive swing, so the sentence was not stale but inverted.

Nothing showed this in a browser. A small script copies the derived card into the overview span at load, so the screen self-healed while the file stayed wrong — the same failure as the share card that displayed correctly and shipped a stale number to anyone who read the source. tools/check-overview-figures.js now runs the page’s own formulas against the page’s own arrays and fails on any stated swing or profit that disagrees, in the callout or in prose. The changelog is excluded, because a dated entry is a record.

Two more holes closed while the gate was being written. The baseline gate knew five ways of phrasing a ratio and not the one in the most-read sentence on every page, which is how City carried “~65% ratio vs the 70% wall” against a computed 57.1%; it now reads that shape too. And tools/check-parity.js checks that CI runs every gate the local suite runs — twice a gate had been wired into one and not the other, and both times it was caught by eye rather than by anything.

Ten answer blocks disagreed with their own pages — and the rule that lets that happen is now written down CORRECTION

The answer block is the first paragraph on every club page and the one a search engine quotes. It states four things that are all derivable — the ratio, the headroom or overshoot, what the club has banked from sales, what it has committed on signings — and only the ratio had a gate. Ten of the twenty disagreed with their own data. Brighton said £91.9m banked against £156.9m of completed sales; Manchester United £89m committed against £154m; Sunderland £0m against £32.6m; Forest was wrong on all three. Every answer block is now written from the deal tables and the ratio inputs rather than typed, and a gate holds them there.

One of the ten was two days old and mine. Manchester City’s said “£338m banked from sales”, which came from reading a total that includes advanced deals into a sentence that says banked. It is £288m. Advanced business is committed, not banked, and the gate now applies that distinction rather than trusting whoever writes the sentence to remember it.

Chelsea, Fulham and Ipswich stated no headroom at all, because they are over their line and the sentence had been written for clubs that are under it. They now say by how much: −£23.3m, −£19.2m and −£17.7m per year.

And the underlying rule is now recorded in the repo rather than relied on. A club page is twelve sections plus the head, the same quantity is stated in several of them, and every significant error here has had one shape: a figure corrected where somebody noticed it and left standing everywhere else. tools/echoes.js reports which sections state a given figure before it is edited, and the gates that exist because of this rule are listed alongside it. Anything derivable gets derived; anything typed gets a gate.

No gate had ever read a headcount CORRECTION

Ten gates run on every build and not one of them counted players. They read percentages, deal statuses, transfer rows, bar widths and card images — so a sentence saying “nineteen senior players, six places unused” could sit beside a card correctly showing 22 and 3, indefinitely, with everything green.

There is now a gate that recounts from the squad table, using the page’s own registration rule rather than a second copy of it: a player is on the list unless he is loaned out or under 21 on the season’s cut-off, and the cut-off is read from the page rather than assumed. It checks the registered total, the homegrown and non-homegrown split, the free places and the under-21 count, in both digits and words.

It missed one of its own test cases on the first run. “Six more players are under 21” begins a sentence, so the word is capitalised, and a lowercase-only pattern walked straight past it — found only by planting each stale figure back and checking the build failed. Four claim types, four confirmed failures, then green. A check you have not watched fail is a guess about what it catches.

The hero stat cards were three-across on a phone, and four of them left an orphan LAYOUT

Reported on Villa: three columns on a 430px screen squeezed the longest card into a ribbon, and because that page carries four cards rather than three, the fourth dropped alone onto a second row beside two empty cells. Below 760px the grid is now two columns, which gives Villa a clean 2×2.

The other nineteen pages carry three cards, so two columns would have left an orphan there instead. A lone last card now spans the full row rather than sitting beside a hole — so three reads as two-then-one and four reads as two-and-two, without a special case for either.

Desktop had the same orphan and nobody had mentioned it. Four cards in a fixed three-column grid is unbalanced at every width, not just on a phone. The grid fits its own content now, so Villa gets four across on a wide screen and the three-card pages are unchanged. Measured at 390, 430 and 1400px rather than eyeballed.

And the card that caused it got shorter. Villa’s List A churn description had just been lengthened to fix an arrival count, which is what made it tower over its neighbour. The names moved out — they are on the Ins & Outs tab — and every figure in it stayed.

A stale ratio phrased one way slipped a gate built for another CORRECTION

The prose-baseline gate catches “roughly N% of football revenue”. It did not catch “sit at roughly N%”, which is how the glossary phrases the same claim — so a stale ratio could sit in a definition indefinitely while the identical statement in the FAQ was checked on every build. Found on Villa, where the glossary was two revisions behind the page.

Extending the rule found the same fault on three more pages nobody had asked about. Arsenal’s glossary said 64% against 66.1% computed; Fulham’s said 90% against 94.0%; Ipswich’s said the ratio “sits at roughly 85% — on the line” when it is 94.6%, about ten points over. Ipswich’s sentence also quoted a “~£106m gross window” that the deal table puts at ~£147m. That one was wrong twice and read as reassurance both times.

A gate that matches one phrasing of a claim is a gate against that phrasing, not against the claim. Both forms are checked now, and the extension was tested by planting the old figure back and confirming the build fails.

Both loan bars read as full when only one quota was used CORRECTION

Reported by a reader looking at Villa on a phone, and right. The card showed two full amber bars, so both ceilings looked hit — when the line under the second one said 6 of 6 international slots still free. Only the domestic quota was used. The bar was hard-coded to 100% and had nothing to do with the sentence beneath it.

Every one of these bars was a literal width typed in beside the text it was meant to summarise. That is the same fault as the headline reading “1 + 6 Left” above a body saying none left, which already had a gate; the bar had simply never been looked at. They are derived now — “N of M remaining” means M−N used, and the width and the colour come from that, so the picture and the sentence cannot disagree. Amber only when nothing is left, which is the point at which a quota starts deciding what a club can do.

Writing the check found four more pages the report had not mentioned. Fulham, Ipswich, Tottenham and Coventry each had a quota line with no bar above it at all, so the international ceiling was invisible on those cards. And Chelsea’s international bar sat at 33% beside a sentence that stated no count whatsoever — nothing could have derived it and nobody could have checked it. That line now says 5 of 6 senior slots remaining and shows its working: Badiashile at Napoli is the one senior non-club-trained loan abroad, with Antwi and Murray-Campbell treated as academy and therefore exempt. Disagree with that classification if you like — the point is that it is now possible to.

The gate checks the source, not just the screen. Villa’s rendered card was fixed by the derivation the moment it was added, while the hard-coded 100% was still sitting in the file — correct in the browser, wrong in the thing a checker reads and the thing that renders before any script runs. Both are right now, and the gate compares the typed width against the stated numbers so a hand edit cannot quietly reintroduce it.

Deep audit of the rules and the inputs: the wage figure had drifted from its own derivation on five clubs CORRECTION

The rule constants are unanimous across all twenty pages and all correct — the £105m PSR cap, the €60m earnings limit, the five-year amortisation cap and its 1 July 2023 start, the 4% Premier League levy, the 1 January 2005 under-21 cut-off, 17 non-homegrown of 25, two concurrent domestic loans and six international. Thirty-nine separate statements of the five-year cap; not one disagreed with another. That is the half of this audit with nothing to report, and it is worth reporting that it has nothing to report.

Every club’s cap now agrees with its European status. In Europe the UEFA line is 70%, outside it the Premier League’s is 85% — fifteen points of headroom turning on a single yes/no, asserted in three places at once and checked in none. All twenty agree: nine in Europe at 70%, eleven outside at 85%. A gate now enforces it, because a club’s European status is exactly what changes between seasons.

The real finding is the wage input. This site’s stated convention is that the wage figure IS the wage bridge’s own derivation on every accounts-basis club — that is what makes the league table comparable, and what makes the input track the window instead of sitting at a pre-window figure nobody updates. It had decayed on five clubs, because deals get booked by hand and the input gets nudged by hand with them, and a nudge drifts from the derivation the moment the arithmetic differs by a rounding. Tottenham were £22m adrift. All five are now set to the derivation and a gate holds them there. Promoted clubs are correctly exempt and the bridge explains why in its own text: a wage bill roughly doubles on going up, so last season’s accounts are the wrong anchor in both directions at once.

One club changes compliance status as a result. Bournemouth go from 69.8% to 71.2% — from just inside the 70% cap to about £3.4m/yr over it, in their first European season, on the tighter cap that European qualification brings. Tottenham move 64.6% to 68.3%, Forest 81.9% to 82.7%, Manchester City 57.7% to 57.1%, Villa 64.0% to 63.8%.

And two more pages were flooring their headroom at zero. Ipswich read “£0m of room” while sitting £17.8m/yr over the 85% line; Bournemouth would have done the same on crossing it. The same defect was fixed on Palace four days ago and not looked for elsewhere — which is the lesson, not the bug. Both now state the overshoot.

What the audit could not do. Every football site and reference database is blocked from this environment, so the transfer tables were reconciled internally — squad rows against deal rows, amortisation floors against inputs, both sides of every intra-league move — and checked externally only as far as search summaries allow. Internally they reconcile on all twenty clubs. That is a real guarantee about consistency and not a guarantee about completeness, and the £50m signing found missing yesterday is what the difference looks like.

The share card is now checked like everything else — and two gates had never run in CI at all CORRECTION

Rebuilding the cards fixed today’s figures and fixed nothing about tomorrow’s. A generator someone has to remember to run is exactly what produced a card reading 60.2% four days out of date, so the question is not whether the cards are right now but whether they can go wrong again quietly. A gate now recomputes the card’s figures from the page and compares them against the token the builder stamped into the meta tag. Move any input — a wage, a fee, a sale — and the build fails until the card is rebuilt. It was tested by nudging Villa’s wage input and confirming the failure, and by stripping the token to confirm that fails too.

The builder and the gate read the figures through one shared module. Two copies of the same calculation drifting apart is the failure this site has hit more than any other, and it would be absurd to reintroduce it in the check designed to prevent it.

And checking the CI list against the local one turned up a real gap. Two gates — the one that catches stale ratio figures in prose, and the loan-quota headline check — existed, worked, and had never run in CI. They only ran when someone ran the full suite by hand, which means a direct push could have deployed stale numbers with a green tick. Both now run on every push, and the two lists are asserted to match. The prose-baseline gate is the one that has caught the most this week, and it was the one running least often.

The share cards were showing a ratio four days and six deals out of date CORRECTION

The squad cost ratio is printed into the card image itself, which is fine until the ratio moves. Villa’s card said 60.2% against a page saying 64.0% — the figure from before the Goretzka wage correction, and before Jackson, Martínez, Suzuki, Ruggeri and the rest. Anyone sharing the link was posting a number this site had stopped believing on 27 August. Chelsea’s said 83% against 88.7%; Manchester City’s said ~65% against 57.7%.

There was no generator. The cards were made once, by hand, on 26 August and left, so nothing could have kept them current. There is one now: tools/build-og-cards.js reads every figure from the page it belongs to and rebuilds all twenty, and it runs like any other build step.

Cache-busting is the part that actually matters for sharing. X and Facebook cache a card image against its URL for days, so replacing the file in place does nothing for a link anyone has already posted. Each card’s meta tag now carries a version token derived from the figures on it, so the URL changes exactly when the picture changes and not otherwise.

The number is now coloured the way the page colours it — green three points clear of the limit, amber up to it, red past it. A card showing Ipswich at 94.6% against an 85% line in the same neutral cream as one showing Manchester United at 56.9% throws away the only thing a reader takes from a shared image.

And a gate now checks the alt text against the computed ratio, because that was stale too and in a way nobody would ever notice: six pages carried figures in the card’s alt attribute that no longer matched, including one that had never matched any published base. It was tested by planting the old 60.2% and confirming the build fails.

Full pre-deadline audit of all 20 clubs — one £50m transfer was missing entirely SWEEP

Nico González, Manchester City to Newcastle, ~£50m, completed and announced. Medical taken, presented to supporters before the West Brom cup tie, No.6 shirt issued — and absent from both pages. It is the largest single omission this site has had to correct, and it moved two ratios materially: City 60.4% → 57.7% with headroom out to £100.2m, Newcastle 70.8% → 74.7% with headroom down to £46.8m against their 85% line. The fee is reported as reaching £50m after add-ons, so the guaranteed portion is lower and unpublished; £50m is carried on both sides because a counterpart has to agree, and it is flagged. The same number cannot be conservative for both clubs at once — it flatters City’s profit while overstating Newcastle’s charge — and that is stated rather than smoothed over.

Everything else checked out, which is the more useful result. Forty-odd names drawn from external reporting were tested against the pages one by one and every single one was already there. Where the reporting and this site disagreed, this site was right: one summary placed Detourbet, Monga and Rulli at Nottingham Forest when all three are Manchester City signings, and another had Bouaddi at Forest when City announced him.

Two numeric errors found by audit rather than by reading. Manchester City’s Kalvin Phillips loan credited £11.3m of relief when a loan frees the wage and none of the amortisation — the registration stays, and so does the charge. Corrected to £7.8m. Nottingham Forest’s Gustavo Sá loan freed nothing at all when Olympiacos pay his wages; corrected to £1.8m. Two more outgoing loans, at Leeds and Liverpool, carried no loan flag, so the amortisation book and the List A gate were treating them as permanent sales.

Eight new checks, and they were run against real data before being trusted. A completed permanent signing must have a squad row; a sold player must not still be in one; a destination reading “loan” must carry a loan flag; no duplicate squad rows; no fee without contract dates; no end date before the signing date; no signing dated in the future; no expired contract still in the squad. All eight are clean across all twenty clubs, and two were deliberately broken to confirm they fail. Three earlier checks this week passed on their first run because they were wrong, not because the data was right, so nothing goes in now without being made to fail first.

What this sweep cannot claim. Every football site, news aggregator and reference database is blocked from this environment, so external checking is search summaries only — and those summaries mixed up clubs twice in this pass alone. The internal consistency of these pages is now verified exhaustively and mechanically. Their completeness against the real world is verified as far as search allows, which is not the same thing, and the González miss is what that gap looks like.

The stale-rumour gate could only see half the market — two more boards were selling players who had left CORRECTION

Yesterday’s gate checked live rumours against completed incomings at other clubs. That catches a player who moved between Premier League clubs and nothing else. Two whole classes of exit were invisible to it, and both had live cards sitting on boards.

A player sold abroad appears in no incoming table anywhere. Spurs sold Cristian Romero to Atlético Madrid and announced it on 15 August; Arsenal’s board went on offering him to the scenario modeller for a fortnight. The gate now reads completed outgoing rows as well, taking the destination from the row itself — which is derivable, and should have been there first time.

And a move between two non-Premier-League clubs is recorded on no page at all. Bayern sold João Palhinha to Benfica on 26 August — €14m plus up to €5m, contract to 2030 — and Newcastle’s card still read “door ajar again”. Nothing structural could ever have caught that, so there is now a small list at tools/gone.json for exactly this case, and the gate fails against it. Two names so far: Palhinha, and Leão to Galatasaray.

One false positive, worth recording because it nearly shipped. The first version read Jadon Sancho’s outgoing row — destination “Free agent” — as him having joined a club called Free agent, and would have killed two perfectly live cards. A released player has left a club and joined nobody: he is more signable, not less. Release destinations are now excluded, while a row naming an onward club after an arrow still counts. All three arms were then tested by un-flagging a closed card and checking the gate fired, because a check nobody has seen fail is not yet a check.

Cards closed: Romero (Arsenal), Palhinha (Newcastle), Cho (Hull, where Hull themselves ended the talks on 10 August). Each keeps its note as the record and leaves the modeller.

Dead rumours were still tickable — three players could be “signed” on boards they had already left CORRECTION

The scenario modeller priced any card you ticked, including cards whose own tag said the player was gone. Newcastle’s board offered Nicolas Jackson after he had joined Villa; the tag read “DEAD — JOINED ASTON VILLA” and ticking it still moved the ratio. Liverpool’s offered Ezri Konsa, an Arsenal player since 20 August. Chelsea’s offered Ousmane Diomande, who is at Forest. This site has named that exact contradiction in its own changelog — “a completed deal and a live rumour for the same player” — and then kept shipping it, because nothing was looking for it.

Now something looks. A gate fails any live incoming rumour naming a player who has completed a move to another club, which is derivable from the transfer tables rather than from a tag somebody remembered to type. Only incoming cards are checked: an outgoing rumour naming a player another club has signed is the same move seen from the selling side, and that is correct. It found all three on its first run.

Dead cards now stay on the board and leave the model. They keep their note as the record of what a club looked at, greyed, with the tick removed and the impact line replaced by “no longer available”. Deleting them would lose the history; leaving them live was worse.

Deadline-week sweep: four announced deals the tables had missed, and a gate that was crying wolf SWEEP

Four deals were sitting at “advanced” after their clubs had announced them. Bouaddi to Manchester City — announced on the club’s own site on 26 August, and carried here as unannounced for four days, with the prose quoting “~£95m/€110m… agreed in principle” when the real deal is £81.3m fixed plus £4.3m of add-ons on a contract to 2031. Badiashile to Napoli, announced 22 August. Disasi to Crystal Palace, announced 26 August by both clubs. And Emi Martínez to Chelsea, announced by Villa on 30 August. All four are now booked, with the ratios on five pages moving as a result.

The Disasi loan is the one with a consequence nobody was watching. Palace were sitting a whisker inside the 70% cap that their European qualification imposes. Taking a £120k/wk centre-back on loan added ~£9.0m/yr of grossed wages and put them over it — a loan, with no fee, no amortisation and no transfer to report, moving a compliance position that a £30m signing would have been scrutinised for. This page has argued all summer that loans are treated as free when they are not; this is the arithmetic saying it.

And the prose-baseline gate was flagging a correct sentence. It reads any “roughly N% of football revenue” as a claim about the club’s own ratio — but “capped at roughly 70% of football revenue” describes the LIMIT, and Palace’s page says exactly that, twice. A gate that fails on correct text is worse than one that stays quiet, because the next person learns to skim past it. It now ignores the phrase when cap or limit wording precedes it, and it still fails on a genuinely stale figure — checked by planting one before shipping.

The league window table was wrong on 19 of its 20 rows — and an edit ate two deal rows without a single gate noticing CORRECTION

The “Who’s Selling to Survive” table is a derivation and had stopped being one. Its own caption defines it exactly — disclosed fees on officially completed 2026 deals, loans and frees excluded — so every figure in it can be recomputed from the transfer tables it summarises. Nineteen of the twenty rows disagreed with those tables. Six clubs were shown as having spent nothing at all, including Newcastle at £0 against £174.9m of completed incomings and Manchester City at £0 of sales against £238m. The rows were hand-maintained, the pages moved on, and nothing compared the two. All twenty are now derived from the pages, with the hand-written notes kept, and a gate recomputes them so the table cannot drift from the deals again.

The second fault is worse, because it was self-inflicted and invisible. Promoting Jackson spliced his row out of one array and into another by matching the next }, — and deal rows share source lines, so the edit swallowed whatever followed on the same line. Goretzka’s incoming row vanished from Villa and Kellyman’s outgoing row from Chelsea. All eight gates passed. A missing row breaks no parse, satisfies every invariant, and simply makes the totals smaller — and nothing on the page knew what they should have been.

Two things now catch it. A squad player signed inside the window must also appear in the incoming table, which is a real structural invariant: the squad row survives when the deal row dies, so it is the witness. That check found a third case immediately, one nobody had touched today — Forest’s Gustavo Sá has had a squad row since July, with a book value and an amortisation charge, and his incoming row was never written. A changelog entry from that sweep claims he was added “both halves of him”. One half was missing. And for outgoing rows, which leave no trace on the page once the player has gone, there is now tools/diff-rows.js: it reports every deal row added or removed against a git ref, to be read before committing rather than after.

Amortisation input raised to the squad floor: £51m → £62m INPUT CORRECTION

Why it moved. The amortisation input is the club-wide charge, which must be at least the current squad's, since it also covers players who have left mid-season and academy registrations the squad table does not carry. On this page it had fallen below that floor, which cannot be right — the summer's signings went into the squad table and the input stayed where it was. It is now raised to the floor.

Read it as a floor, not a figure. The true club-wide charge is higher by whatever departed players and fringe registrations add. This is a conservative correction, and it moves the ratio the wrong way for the club, which is the direction an honest correction usually goes.

A free transfer was costing the window totals nothing at all — £35.7m/yr across six pages CORRECTION

The cost-per-arrival calculation returned zero for any signing without a fee unless the row also carried a loan flag. So a free signing’s wages vanished from “cost added by this window” entirely. Goretzka was showing as costing Villa £0.0m. On the very page whose card argues that “the free one is the expensive one — a fee buys an asset that amortises and can be sold, a wage buys nothing and recurs”. The card was right and the arithmetic underneath it dropped the wage.

£35.7m/yr was missing across six pages: Villa £13.6m (Goretzka, Özcan, Guessand), Palace £8.1m (McNeil, Guessand), Liverpool £7.8m (Araújo), Everton £4.9m (Johnson), Coventry £1.1m, Fulham £0.2m. Villa’s window swing moves from +£12.4m/yr to +£26.0m/yr.

There were seven different versions of that one line across twenty pages, each special-casing frees and loans slightly differently, and the differences were exactly what caused this. They are now one formula — the one Hull already used: amortisation and the agent fee when there is a fee, and the wage on every arrival, fee or not.

And it exposes something this card cannot fix by itself. The List A gate carried a hand-typed margin of “Villa pass by ~£3.5m/yr” on the cost-based test. The page’s own sums now put the window at adding ~£26m/yr. That figure predates Jackson and Watkins, and the window totals count wages ungrossed while the ratio grosses them, so the two were never on one basis. The cost-based margin is now marked UNVERIFIED rather than passing — the fee-based balance above it is sound, that one is not, and saying so beats printing a number that cannot be reproduced.

A headline that contradicted its own paragraph — and nine more stale ratios behind it CORRECTION

Villa’s loan-quota card said “1 + 6 Left” in green, directly above a paragraph reading “0 of 2 concurrent remaining” and “ROOM: none left domestically”. Wan-Bissaka’s loan on 18 August took the second slot. The body was updated; the big number at the top was not. So the figure a reader sees first said a slot was free while the text underneath called the same thing a binding constraint. The bar sat at 50% with nothing left, and the card was still green. Crystal Palace had the same shape — “2 + 6” above “1 of 2”, left over from before the Guessand loan.

The failure runs one way every time: the prose gets edited and the summary above it does not. tools/check-quotas.js now fails when a loan-quota headline disagrees with its own body, including when the room note says “none left” and the headline disagrees. Verified by sabotage — put the wrong number back and it fails on exactly that.

Then the baseline gate was widened and found nine more. It was matching “ratio is X%” and “X% → Y%” but not the Quick Answers phrasing, “roughly 63% of football revenue” — which is the plain-English answer a reader or a search engine is most likely to lift. Brentford were claiming 75% while computing 84.9%. Manchester City 65% against 57.3%. Newcastle 74% against 70.8%. Ten pages rebased from their own inputs.

Two gates, two days, and both found more than the bug that prompted them. That is the argument for writing the check rather than the fix: the fix closes one hole, the check tells you how many others there were.

The baseline gate had a hole in it and was passing on what it could not see CORRECTION

The gate shipped this morning excluded the changelog by truncating each page at <section id="log">. The data arrays holding every rumour card sit in a script block after that point. So it was scanning 167KB of a 422KB page, never saw a single card’s projection, and reported green.

It now removes only that section and keeps the rest. On the first honest run it immediately found eleven stale figures across three pages — every Villa projection still quoting a baseline two changes old, and two on Chelsea. All recomputed.

A gate with a hole in it is worse than no gate, because no gate makes you check and a green one stops you looking. This one was written, tested against the case that prompted it, and shipped — and the case that prompted it happened to sit in the part of the file it could still see.

One rule came out of it. A live rumour card may state a projection, because the gate now keeps it honest against the current base. A completed deal may not: “60.2% → 63.4%” was true the day it was written and decays the moment anything else moves, and the effect is already in the headline. Those now state what the deal costs instead.

A new gate: does the prose still quote the ratio the page computes? CORRECTION

Twelve of the twenty pages were asserting a squad cost ratio their own inputs no longer produced. Manchester City claimed 63.0% while computing 57.3%. Manchester United said 53.9% against 56.9%, Sunderland 60.9% against 63.2%, Coventry 69.7% against 71.0%. Every one has been rebased from the page’s own figures.

Most of it sat in the answer paragraph and its structured-data twin — the text search engines and AI crawlers read before anything else, which makes it the worst place on the page to be stale. The headroom figures beside them were rebased too.

Nothing was catching this. The status gate reads deal status; the invariants gate reads the arrays; neither reads a percentage in a sentence. It surfaced only because a single free transfer moved one club’s baseline and the knock-on was visible. tools/check-baselines.js now recomputes every page’s ratio from its inputs and fails on prose asserting a different one, and it runs with the rest on every change.

Two things it deliberately does not flag. A sentence reading “68.6% → 62.0%” where 62.0% is the current ratio is a completed move being described, not a stale claim, so a done deal can still say what it did. And the changelog is excluded entirely: a dated entry recording August’s number is the record, and rewriting it to match today would be the opposite of what this tab is for.

The amortisation graph now splits the squad you had from the squad you bought FEATURE

One bar was hiding the more interesting number. The run-off now stacks two series — the squad that was already here in blue, and everyone signed this window in amber — with both totals and a combined figure above it, and a per-season split beside each bar.

It changes what the chart says about a club. A settled squad shows a thin amber band on a thick blue one; a rebuild inverts it. Coventry carry £14.0m already here against £28.9m signed this window, so two-thirds of their amortisation arrived in ten weeks, and their inherited book runs out entirely by 2028/29 while the new one still has three years to run. Liverpool are the other extreme at £143.6m against £18.4m. Same chart, opposite stories, and neither was visible when it was one bar.

Who counts as “this window” is taken from the signing date, not from the flag the squad rows carry. The two disagree: the date catches six players the flag had missed — Garnacho and Cissé here at Villa, Bentley at Coventry, Charles at Fulham, Sá and Diomande at Forest — and never disagrees the other way. A derived test cannot be forgotten the way a hand-set one can, which is the same reasoning that put the sell-ons and the outgoing amortisation on derivations.

Two rows of figures now sit above the chart and they will not always agree, so the panel says which is which rather than leaving a reader to spot it. The first row and the graph are the SQUAD, split by arrival date. The second row is the DEALS, including advanced ones whose players are not in the squad table yet. Where they differ the panel names the gap and what it is — agreed business that has not completed.

Every deal now shows what it costs per year — and the agent fee was being charged five times over FEATURE + FIX

A fee only means something against the charge it creates. £47.5m is a headline; £9.9m a year for five years is the thing that reaches the squad cost ratio. Until now this page printed the fee on every row and the amortisation on none of them. Both tables now carry an AMORT £m/yr column, and a panel underneath gives the number those rows add up into: the squad’s total annual amortisation, what this window has added and freed, and how the charge runs off over the next five seasons as existing deals finish.

None of it is newly typed data, which is the point. Incoming charges come from the player’s own squad row where one exists — real contract dates, five-year cap applied where it bites — and from fee over five years where a deal is advanced and the player is not in the squad yet, flagged est when that happens. Outgoing charges are derived rather than stored: a row’s cost-freed figure is the ungrossed wage plus the amortisation, so subtracting the wage leaves the charge that stops. They cannot drift from the totals above them because they are computed from them.

Loans show a zero, and that is the correct answer rather than a missing one. A loan moves the wage and leaves the registration here, so the amortisation stays too — the point this site has made about loans since the wage bridge, now visible as a column entry instead of a paragraph.

And a bug the new column sat straight on top of. The cost-per-year calculation behind these tables charged the agent fee in full every single year instead of capitalising it and spreading it over five — the same error corrected in the scenario modeller last week and missed here. On a £47.5m signing it overstated the annual cost by £1.9m. Every SCR-cost figure in the Ins and Outs tables, and the window’s cost-added total, move down accordingly.

One thing the new column makes visible and this page is not hiding: the SCR cost figures in these tables count wages at face value, while the rumour cards, the scenario panel and the wage bridge gross them up 1.45x for employer NI, bonuses and image rights. The same signing can therefore read £17.1m in the table and £20.2m on its card. The grossed figure is this site’s convention and the table is the odd one out. It is stated in the panel rather than quietly reconciled, because restating it moves every published window total on all twenty pages and deserves its own pass.

The “last updated” notice is gone — it had stopped telling the truth REMOVED

There were two of them. A line under the header reading “DEALS & FEES AS REPORTED [date] · BOOK VALUES & CALCS RECOMPUTE DAILY”, and a script that read that date and, once it was four days old, pushed an amber banner above the header warning that the page had last been checked N days ago.

The second one was firing on all twenty pages. The date it keyed off was hand-typed and ranged from 9 to 19 August across the site — while the deals underneath had in fact been swept on the 24th and 25th. So the site was leading with a warning about itself, generated from the one number on the page that nobody was maintaining. A staleness notice that is itself stale is worse than none: it tells readers to distrust figures that are current.

Both are removed, with their CSS. Nothing else changes — every ratio, book value and window total is unaffected, and the dated entries in this changelog remain the honest record of when each page was actually worked on.

You can now change the numbers — and the card line beside them was still wrong FEATURE + FIX

Every fee on the rumour board is somebody’s estimate. Brugge “value him around €45m”. Al-Hilal’s bid reached this site in the wrong currency and had to be chased down. Marmoush’s fee is arithmetic off a reported two-player budget rather than a figure anyone published. Readers were being asked to take those guesses as given.

So the amounts are editable now. Every card carries a fee box, and incoming cards a wage box too. Type your own number and the card’s cost line and the scenario panel both follow it immediately. Reset puts the reported figure back — overrides are held separately and the underlying data is never overwritten, so there is always something true to return to. Outgoing cards expose the fee but not the wage, because an outgoing player’s relief is derived from his squad row rather than typed, and the fee is the part that actually moves the profit on the sale.

And a fix that should have landed with the modeller rebuild. The card’s own “ADDS £Xm/yr” line was still running the OLD formula — ungrossed wages, and the agent fee charged in full every year. The calculation underneath was corrected; the display beside it was not, so every card showed one number while the panel below computed another. On this site’s own worked example the card said £10.1m and the panel used £10.7m — the exact discrepancy that rebuild was written to explain, sitting in the markup the whole time. Both now call the same helper, and that is checked on all twenty pages: tick a card, and its figure and the panel’s agree.

The 22 August sweep, and a modeller that was wrong three ways DATA + METHOD

Piroe has a bid on the table. West Ham have offered £15m and Leeds want £20m — so the £10m this page modelled sat below the figure that has already been rejected. Gnonto's suitor is Freiburg, not the unnamed Serie A clubs the card carried.

The scenario modeller was wrong in three ways at once, on every page on this site. An incoming signing's wage was charged with no gross-up, while the wage bridge, the ratio inputs and every note here gross wages by 1.45x for employer NI and image rights — so every modelled signing was 45% cheaper on the wage line than the same player would be in the ratio directly above it. The agent fee was charged in full every year instead of being capitalised with the fee and amortised, so a £60m signing carried £3m/yr of agent cost rather than £0.6m. And sale profit never reached the denominator: nineteen pages let a sale relieve the numerator and stop, when the profit also lands in the denominator through the three-year average, and the twentieth added the whole window's profit when one window belongs at roughly a third.

Why none of it showed up. The first two faults pull in opposite directions and roughly cancel at a typical fee-to-wage ratio — for a £25m signing on £70k/wk the old code gave £10.09m/yr against a true £10.73m, close enough to look right. They do not cancel at the edges: a free transfer on big wages was understated by 45%, a big fee on modest wages overstated by about 9%. Two errors that hide each other are worse than one that does not.

Outgoing relief is now derived, not typed. The hand-set relief on all eighty outgoing rumour cards carried the same ungrossed wage — checked against the squad, the stated figures reproduced the ungrossed derivation almost exactly. Relief and profit now come from the squad row, on the same gross-up and the same five-year cap as everything else, with the typed figure kept as a fallback for the handful of names with no squad row. Profit is allowed to be negative, because some of these sales are losses and the page should say so.

Elvedi signed — the rumour was right about the player and wrong about the price PROMOTED TO DONE

What happened. Leeds have announced Nico Elvedi from Borussia Mönchengladbach on a three-year deal, the club’s fourth signing of the summer. Held here since 14 August as a rumour with personal terms agreed and no club-to-club deal reported; the club side is now done.

What it does to the numbers. The fee came in at ~€5m (£4.3m) against the £6m modelled — he was deep into his final Gladbach year, which is the whole reason a 29-year-old with 364 club appearances and 70-plus Switzerland caps costs that little. On three years it is ~£1.4m/yr of amortisation; his wage, a tracker estimate at £65k/wk, is the larger line by a factor of three.

The shape promoted clubs want. Low fee, short contract, top-flight defending. It is the opposite of the Trafford deal on the same page — £40m over five years — and both are defensible for different reasons. Only one of them constrains the club in 2029.

Three new defensive threads: Elvedi terms agreed, Geertruida contact, Pope competition RUMOURS ADDED

Nico Elvedi. Personal terms fully agreed on a deal to 2029 and the player is reported very open to coming — the club-to-club piece (final-year Gladbach contract, so nominal money) is what remains.

Lutsharel Geertruida. Contact over a permanent at a ~€20m valuation — one fee, three positions of cover.

Nick Pope. Free to leave Newcastle; Leeds reported competing with Ipswich. Cheap fee, real wages, proven keeper.

Trafford goes official; Darlow joins Man Utd; Perri's Torino exit agreed; rumour board pruned

This entry covers 4–9 August in full — the page missed two scheduled refreshes, so everything below was re-verified from scratch rather than assumed.

James Trafford is officially a Leeds United player. The club announced the signing from Manchester City on 6 August on a five-year deal to June 2031 — the fee is officially undisclosed and stays at the reported £40m (+£5m add-ons per earlier reporting; one outlet puts the package at £45m, noted rather than adopted). Status moves Advanced → Done. The squad row now carries the announced contract dates (Aug 2026–Jun 2031), so the auto-derived amortisation (which comes straight from the squad contracts) resets to ~£8m/yr for Trafford (+~£1.4m/yr versus the longer contract previously assumed) — in the headline ratio that is almost exactly offset by Darlow's wages leaving the numerator, so the estimated ratio is essentially unchanged. No window-balance change either: the fee was already counted while the deal was advanced.

Karl Darlow → Manchester United, free. Leeds offered the 35-year-old a new deal; he rejected it, called Old Trafford back-up duty a "special opportunity", and left on an expired contract — no fee, ~£1m/yr of wages off the book. Removed from the squad table (the SCR wages estimate trims £110m → £109m accordingly) and from the homegrown pool, which still carries roughly eleven HG names against a need of eight.

Lucas Perri → Torino: agreed, not done. A loan with an option to buy reported at ~£9.4m is agreed between the clubs — Leeds withheld the green light until Trafford was official, and a minor injury pushed Perri's medical to Monday. Flagged ADVANCED, worded as reported, and classed out in the flows; because it is a loan, his £13.9m fee keeps amortising at Leeds unless the option is exercised, so no book value or window-balance figure moves today.

Rumour Mill pruned and refreshed. OUT: Carlos Espí — Real Madrid paid his £21.4m release clause, deal done, removed. Manor Solomon — West Ham agreed £5m + £2m with Spurs and Leeds passed on re-signing him, removed. UPDATED: Zian Flemming now leads the striker shortlist, but Burnley are braced for bids and reportedly want £20–30m (model fee raised £15m → £20m), with Brentford and Coventry rivals; Petar Sucic — Inter have shut the door, briefing "£90m or not at all". ADDED: Roony Bardghji — Leeds among several PL clubs interested in the Barcelona winger on loan; Barça prefer a sale with a buy-back and no formal proposal exists, so he is modelled fee-free. TEAMtalk's striker additions (Ermedin Demirović, Club Brugge's Nicolò Tresoldi) are noted in prose only — no fee or stage is public, so nothing is invented around them.

Untouched, deliberately: the 3-year smoothed sale-profit input (no completed sale this week), PSR estimates, book values other than Trafford's contract correction, and the Byram/Cairns contract-talks flags — reporting on their outcomes is ambiguous, so the amber stays until an official line exists.

Methodology: squad-cost amortisation now derives from the squad contracts

The headroom model's amortisation input used to be a hand-maintained estimate, while the Squad tab computed the same thing bottom-up from every player's fee and contract dates. The two had drifted apart — by up to £29m across the league, which is several points of squad cost ratio on pages whose whole purpose is showing how close a club sits to a compliance line. The hero model now takes its amortisation straight from the squad contracts, so the headline ratio and the squad table can no longer disagree, and the figure updates itself whenever the squad data changes. The field is still editable: type over it and the what-if model responds as before. Wages remain a stated estimate — clubs publish total staff costs, not the players-plus-head-coach slice the rule needs, so that split is still this tracker's assumption and is labelled as one.

Window rebuilt from a full source list — one loan return added, Meslier's destination corrected

Method changed. Previous passes treated the list already on this page as the baseline and only hunted for news since the last update — which is exactly how a completed deal goes missing. This pass rebuilt Leeds' summer-2026 ledger from scratch — every permanent signing, sale, loan in, loan out, loan return, free, release and academy exit — and only then compared it against the page. Caveat stated up front: Transfermarkt, the club site, Sky and the BBC all refused automated access on this pass, so the rebuild was assembled by cross-reading all nineteen counterpart club ledgers rather than a single canonical source. Anything those ledgers do not evidence has been left out and flagged below rather than guessed at.

Added: Joe Gelhardt, loan return from Hull City. His season-long Championship loan expired and he is back at Elland Road — Hull's ledger carries the matching exit and this page carried nothing. No fee moves in either direction. He was already inside the squad table and the 25-man list as retained depth, so his ~£1.0m/yr of wages were already in the wage base: the row is recorded as a Loan rather than a signing precisely so it is not double-counted into "Fees Committed" or "Annual Cost Added". His chip in the forwards flow now says back from Hull loan instead of implying he never left.

Corrected: Illan Meslier's destination. The Outs table said only "Released". He was released — his contract expired on 30 June and Leeds banked nothing — but he then signed for Arsenal as a free agent on an initial two-year deal plus a one-year option, and the row now says so. This is a labelling fix, not a financial one: the fee was and remains £0, there was never a sale to book against a £0 book value, and the £3.1m/yr of wage relief is unchanged. The 25-man list chip and the goalkeeper flow were updated to match.

Added to the 2025/26 table, deliberately not to this window: Jack Harrison. Everton's ledger records a Harrison exit to Leeds, and it is real — but it is a loan expiry in summer 2025, the previous window, so presenting it as current business would be wrong. It sits in the 2025/26 In table where it belongs and moves no 2026 figure. Honest gap: he is absent from this page's 2025/26 registered list and from the current squad, and no reachable source settles whether he was released, sold or simply left out — so nothing was invented around him.

Counterpart flags checked and closed without edits. Facundo Buonanotte — Brighton's page now carries the matching January-2026 loan out, and this page already has the arrival in the 2025/26 table plus the loan expiry in the 2026 Outs. The two agree; he is back at Brighton and no 2026 exit has completed. Harry Wilson — Fulham's page now carries the matching free transfer out, completed early July, status done on both sides. Both flags were stale rather than wrong.

Reconciled, no change made. Two fee differences with counterpart pages are base-versus-maximum, not errors: Brighton book Struijk at ~£18m, which is this page's £15m base plus the £3m of add-ons that are correctly not counted until triggered; Newcastle book Longstaff at £15m, which is this page's £12m base rising to £15m. The conservative base figures stay. Trafford stays flagged Advanced at £40m + £5m: there is still no official announcement, and Manchester City's page agrees the deal is unannounced while modelling a lower ~£25m fee from their own side — that gap is noted rather than silently split.

No financial figure moved, because no confirmed deal moved one. Fees out, fees committed, net balance, cost freed, cost added, the SCR estimate, PSR headroom, book values, amortisation and the 25-man homegrown arithmetic are all unchanged: the loan return is fee-free and already in the wage base, and the Meslier fix was always a £0 release.

Page launched

The Bean Counter's Leeds United tracker goes live at bean-counter.co.uk/leeds. Launch data set: the £40m + £5m Trafford agreement (announcement imminent — most expensive English goalkeeper ever), Muharemovic (£34.1m) and Wilson (free) in; Struijk (£15m + £3m to Brighton), Schmidt (£1.1m) and the Meslier release out. FY25 accounts digested (£137m record Championship revenue, £107.7m player wages, −£49.2m pre-tax); the closed 2023/24–2025/26 PSR cycle modelled against its reduced £61m limit; the new 85% squad cost ratio, SSR tests, 25-man list homegrown arithmetic and the Elland Road ~53,000 expansion all documented. Rumour Mill opens with Sucic, Brandt, Flemming, Espí and Solomon in, Piroe and Gnonto out.