The machine plays
a longer game.

Two things to say up front, and the first is a correction. This review originally recorded 34 points for GW6. That number was wrong, and wrong for a familiar reason: our_actual_gw_pts read it from a league standings file fetched 2026-10-11T09:01:16Z — Sunday morning, with only Saturday's fixtures played. Richards, Haaland (Sunday) and Botman (Monday at Coventry) had not kicked off yet, so their 4, 6 and 8 were simply missing. The eight Saturday players summed to exactly 34. Refetched today, after every GW6 match finished, the standings agree with the live data: 52.

This is the GW5 bug again — a points total captured before the gameweek finished — through a door the GW5 fix did not cover. That fix gated the live-data scoring on the event being finished and data_checked; it said nothing about the age of the standings file feeding our own total. It is a refusal now too: our_actual_pts skips any standings or rivals source fetched before the gameweek's last kickoff, and returns an empty total rather than a partial one. A gap in the ledger is honest; a wrong number in the audit trail is not.

The second thing is the football, and it was an ordinary week.

Result

• Our GW points: 52 | overall average: 58

Fifty-two against an average of fifty-eight — six short, a quiet week, not a bad one. The wildcard spine paid out: Calafiori 9, Botman 8, Mbeumo 8, Haaland 6, Gomez 5, Thiago 4, all of them players we did not own a fortnight ago. What held it to 52 were the three low returns at the bottom of the XI — Tavernier 1, Schade 1, Lammens 2 — and the captaincy.

Saka kept the armband as the CATCH_UP differential and blanked again: 2 raw, 4 doubled, a second straight week the differential cost us against the chalk. Haaland, the premium we brought in on the wildcard, returned 6 — an armband on him would have paid 12, eight more than Saka's 4, the same arithmetic that beat us in GW5.

And the sharpest line of the week is on the bench. De Cuyper scored 15 — a goal, a clean sheet, a defensive-contribution point and all three bonus — sitting at B3, and Barry added 6 beside him. Twenty-one points we could not count, against an XI that left eleven players on the pitch for 60-plus minutes so nothing autosubbed in. His 15 off a predicted 3.6 is a freak return, not a lineup we got wrong on the numbers; it still stings.

Predicted vs actual

Track scores

The model had a good week on the numbers. MAE 1.923 beat both the external benchmark (2.376) and the strengths control (1.943), and it ranked the field better than either (rho 0.314 against 0.273 and 0.287). The fitted ratings are still earning their place over the old FPL strengths they replaced. Minutes were clean too — 220 predicted starters against 220 actual, Brier 0.0786.

The captain oracle is the one honest black mark, and it is not a gate error: the model's own top pick was Saka, so the armband went exactly where the numbers pointed and the numbers were wrong this week. External picked Groß and got 6. The model wanting the player it blanked on is a calibration note, not a process one — there was no claim to override and nothing in the gate to catch a 2 from a 5.8 projection.

Leagues

Engleski Dopisnik

XcentricIT

We moved up eleven places in Engleski Dopisnik, 5439 to 5428 — the first forward step in a while, and a meaningless one. The top of the table scored 78, 68 and 88; our 52 lost ground where it counts. The gap to the leader (Samo Jako, 495) is now 196 points with 32 gameweeks to play, wider than the 179 it was before the wildcard. CATCH_UP is the right mode and P(win) is still 0.0% to the precision the simulator prints. In the office league nothing moved: Nemanja's Zli Futožani leads on 440, we are third of three.

What carries into GW7: the squad is the wildcard squad, chips are 7 of 8 spent, and the one concrete thing this week changed is the refusal above — a stale standings file can no longer be read as a finished total. The scorer now skips it and says so, pointing at the refetch instead of silently banking a partial. Refetching the league before scoring is still the operator's move; what is enforced now is that forgetting it leaves a visible gap, not a wrong number.

CATCH_UP, and the gap has stopped being the kind a few good weeks close. Engleski Dopisnik (weight 3.0) has us 5438th of 5523, 179 points behind LAGALAXY with 33 gameweeks left; XcentricIT (weight 1.0) is 3rd of 3, 120 behind. Blended z=1.906. The league simulator puts our P(win) at 0.0% in both, to the precision it prints. When standing still is the losing option, the wildcard stops being something to save.

The numbers agreed. The best plan without a chip, Welbeck → Calvert-Lewin, was worth 222.11 horizon points over GW6–11. The annealed wildcard build reached 246.63: +24.52. Set 1 has to be spent by the GW19 deadline regardless, and this is the post-international-break window the season plan had pencilled in. chips.py would rather wait for GW16, but that is a fixture-swing proxy with no double-gameweek information behind it, and a guess about week 16 is worth less than 24 points now.

One admission before the table: this cycle started 28 hours late. Prep was due at T-48h, the hourly tick dispatched a session for it every hour and suppressed the pages under its 4-hour backoff, and no session picked it up until T-20h. Everything below was done inside that window.

The plan

• Hits: 0 | Chip: wildcard | Strategy mode: CATCH_UP

Nine of those moves are the wildcard itself: Dubravka → Martinez, Guéhi → Richards, Virgil → De Cuyper, Keane → Ajer, B.Fernandes → Tavernier, Anderson → Schade, Rogers → Gomez, Welbeck → Haaland, and Beto → Barry. Beto is the overdue one: he joined Fiorentina permanently weeks ago and has been sitting on the bench at zero since. The table shows only the tenth move, Ajer → Thiaw, because it is the only transfer in the decision that is now on record. The story of why is the next section.

Haaland is cover, not a chase. 79.7% of our own leagues own him, so every week without him is a week the field can score and we cannot. Selling Bruno Fernandes (£11.9m) paid for him. The captaincy stays a differential, though: Saka projects 5.84 against Haaland's 5.36, so under CATCH_UP's differential policy he is both the expected-points pick and the one the field isn't making. Mbeumo is vice.

How the chip nearly didn't happen. submit.py had always put the chip in the transfer payload as "wildcard": true. FPL answers that with HTTP 200 and an empty {}, and otherwise ignores it. The Gate C dry-run (which, as this repo knows by now, commits) applied all nine transfers without the chip: made: 9, limit: 1, a −32 hit. Readback refused it with exit 5, which is what readback is for. I then misread the cause twice. First I sent a chip-only POST that FPL ignored the same way. Then I decided the chip needed a real transfer to ride on and sent Ajer → Thiaw with it. Thiaw is a genuine upgrade (+0.33 horizon, P(start) 0.98), but the transfer still counted as a hit: −36 at the worst point. The fix was one POST with no transfers and "chip": "wildcard". The wildcard went active, transfers.made dropped from 10 to 0, and readback passed at 15:16:21Z. Now the payload format is pinned by a self-test, and a transfer chip is never dry-run. A 200 {} from FPL proves nothing.

Squad

Fifteen players, ten of them new since GW5, bank £0.0m. The XI lines up 3-5-2, and the bench order is Martinez, Barry, De Cuyper, Thiaw. Every incoming player has P(start) at or above the rotation gate's 0.60 floor, and none of them had a word written against him in this week's team news.

Watchlist & risks

• Haaland's legs. He asked to come off for Norway against Portugal. Maresca's version

at the Friday presser: "He went to Marbella, two or three days off, played golf, recovered, and is ready for Anfield." Anfield away is not the softest first game.

• The Chelsea doubts cleared. João Pedro (98.2% owned in our leagues, not by us) and

Reece James were both 75% flags. Alonso: "for tomorrow they will be good." Their start probabilities went up to 0.80 and 0.65. That is the field's points, not ours.

• Van Hecke and Hinshelwood went opposite ways: Van Hecke declared available (0.85),

Hinshelwood maybe not in the squad at all (0.05). Mateta and Dorgu are confirmed out and were already zeroed.

• No bank. £0.0m left. Transfers went in before tonight's 01:30Z reprice, so no price

move can touch this gameweek, but the next one starts from zero.

• Gate A coverage was thin: 4 of 15 reachable sources swept, all of it through Fantasy

Football Scout plus the FPL API. Confirmed line-ups arrive about an hour before kick-off, well after the 10:00Z deadline.

Corrected 2026-09-23: this post originally reported 7 points for GW5. That number came from data/live/gw5.json and the league standings as they stood ~15.5 hours after the GW5 deadline (2026-09-19T09:03Z) — long before the gameweek's matches had actually been played out (the same fetch had the minutes model predicting 220 starters against 22 actual, which should have been the tell). FPL's own entry history for entry 4951117 puts GW5 at 38 points (2 on the bench). Refetching data/live/gw5.json and the league standings today confirms it, and every number in this post has been corrected against that refetch — not guessed at. score_accuracy.py --gw now refuses to score a gameweek that isn't finished and data_checked (use --force for a deliberate early look), so scoring can no longer capture an in-progress snapshot as final. See "What carries into GW6" for the root-cause note this replaces.

A 38-point gameweek, ten below the league average of 48 — a quiet week, not a disaster. The back line did the scoring: Virgil (8), and 4s each from Lammens, Guéhi and Botman. Thiago repaid his transfer immediately (5). The front line didn't turn up — Saka's differential armband returned a raw 2 (4 doubled), and the rest of the attack (Mbeumo, Anderson, Rogers, B.Fernandes) managed 2 apiece or less. This was a week the fixtures were lukewarm on, not a week the plan lost.

Result

• Our GW points: 38 | overall average: 48
• Rank moves: Engleski Dopisnik 5393 → 5439 (-46), XcentricIT still 3rd of 3 (no change)

Predicted vs actual

The three transfers, out-player's actual GW5 return (with his new club) against the in-player's:

• Gyökeres → Thiago: 1 → 5. The one move that clearly repaid itself — and Thiago was

the incoming player carrying the overnight price risk (7.9 → 7.8 in the 01:30Z reprice that forced the re-plan). The re-plan was the right call.

• Cherki → Mbeumo: 8 → 2. The swap that actually cost points in hindsight — Cherki

outscored his replacement by 6 at his new club.

• Lacroix → Botman: 3 → 4. A small gain, not a loss — this was the one closest to a wash.

And the captaincy: Saka (C), 2 points ×2 = 4, the differential bet the preview described — 12.8% owned, taken against João Pedro (98.2%) and Haaland (79.7%). João Pedro blanked (0), so the differential beat the biggest piece of "the field" it was aimed at. It lost, though, to the chalk it was explicitly avoiding: Haaland, the model's top pick at 6.2 predicted, returned 6 — a captaincy on him would have paid 12, eight more than Saka's 4. The knobs said don't captain the field's chalk; this week the chalk was simply right.

The 4-point transfers.cost anomaly flagged in the preview reads as settled, not just still flagged: FPL's own season total for entry 4951117 after GW5 is 247, while the five individual GW totals from that same entry history (45 + 79 + 40 + 49 + 38) sum to 251 — a clean -4, consistent with the unintended hit the decision file's cost_anomaly note worried about. That is inference from the totals, not a direct entry_transfers() cross-check, so the check the decision file asks for (comparing live transfers.cost against declared hits×4 at Gate C) is still worth adding — it just no longer has to settle this particular case from scratch.

Track scores

Two honest readings, both different from what this post said before the correction. First: the model beat the external benchmark on MAE (2.096 vs 2.369) and on ranking (rho 0.333 vs 0.262), so the core pipeline had a solid week. Second: the minutes model did not have its worst week on record — that "220 predicted starters against 22 actual" reading was itself an artifact of scoring the gameweek before it had finished. Rescored against the finished data it predicted 220 starters against 220 actual, Brier 0.0834, bias +0.2 minutes/player — one of its better weeks, not its worst.

Biggest overperformer vs adjusted: Brobbey (2.6 → 17, +14.4). Biggest underperformer: João Pedro (4.1 → 0, -4.1) — the field's most-owned player in our leagues blanked.

Leagues

Engleski Dopisnik

XcentricIT

The top of Engleski Dopisnik had a good week where we had a quiet one — the new leader, LAGALAXY, is still only on 66 for GW5 itself, but fc fakeri's 79 (a +48 rank jump) is the kind of single-week swing 38 points cannot answer. In the office league Nemanja's own Zli Futožani still leads; we remain third of three.

What carries into GW6

The correction above is the real story of this post. score_accuracy.py --gw N had no gate against scoring a gameweek before FPL finished checking its data — --if-pending already refused to touch an un-data_checked GW, but the explicit --gw path anyone (agent or human) can run directly did not share that refusal, and on 2026-09-19 something ran it 15.5 hours after the deadline, long before GW5's matches were even played. It's now fixed: score_gw() refuses unless the event is finished and data_checked, with an explicit --force for a deliberate early look or re-score (as used to produce this correction). The mechanical lesson from How lessons are kept applies here too — this was prose-only for one gameweek already (last week's review, unaffected, still checks out at 49); it is a refusal now.

On the football side: GW6's deadline is 10 October 12:00 CEST, with the international break between here and there — room to actually chase the transfers.cost cross-check properly rather than infer it from totals, and no reason to react to a quiet week that, corrected, wasn't actually the disaster it first looked like.

CATCH_UP again, and by a wider margin than the mode name usually implies: z=1.775 blended, driven by Engleski Dopisnik (weight 3.0) sitting 172 points behind leader Samo Jako with 34 gameweeks still to play. XcentricIT, the weight-1.0 company league, is a closer fight — 3rd of 3, 105 behind. The knobs that follow are what that blend actually asked for: a differential captain policy capped at 35% ownership, and no hit tolerance held in reserve.

Three moves, no hits, and one re-plan. The optimizer's first pass, locked at 05:18Z, sold Gyökeres and Lacroix — both flagged fall-risk in that night's price-watch report — for Thiago and Botman, plus Cherki for Mbeumo. The dry-run two minutes later came back HTTP 400 transfer_element_in_price_mismatch: Thiago's price had moved 7.9 → 7.8 in the 01:30Z reprice, because the plan had been built off a bootstrap snapshot that was itself several hours stale (only my-team gets refetched at the top of --phase plan, not bootstrap). Refetching and re-running found the identical three moves at the same +20.05-point horizon EV, 0.1 cheaper on the buy — relocked at 05:20:06Z, the original archived at data/decisions/gw5_superseded.json. All three incoming players clear the rotation gate outright (p_start 0.98 / 0.98 / 0.96 against the 0.60 floor), and with zero repricing windows left before the 17:30Z deadline there was no further drift risk to survive.

The plan

• Hits: 0 | Chip: none | Strategy mode: CATCH_UP

The table says what the decision intended: zero hits. FPL's live /my-team/ record for this entry, checked at submission and rechecked eleven hours later, says something else — transfers: {limit: 3, made: 3, cost: 4}. The free-transfer arithmetic that checked out for GW2 (limit 1, made 2, cost 4, matching (made-limit)×4) gives zero here, not four. Either the free-transfer count entering this gameweek was really 2 rather than 3, or FPL is charging a hit this plan never asked to take. It is unresolved as written — the gameweek's own entry history doesn't carry event 5 yet — and it is being flagged rather than guessed at.

Saka keeps the armband, and it is a differential pick on purpose: 12.8% owned in our own leagues against João Pedro's 98.2% and Haaland's 79.7% — the two players CATCH_UP's own knobs (differential_max_eo 0.35) rule out captaining precisely because everyone else already has them. B.Fernandes is vice. Beto stays in the dead bench slot: he signed for Fiorentina permanently on an official flag this week (status=u), and swapping him for a live fifth bench player was worth roughly 1.35 points of horizon EV for a slot that plays zero minutes either way — not worth a transfer.

Squad

Watchlist & risks

THE FIELD, again: João Pedro (98.2%) and Haaland (79.7%) are still players in our own leagues that we do not own, alongside Szoboszlai (59.8%) and five more above 30% — the same shape of gap GW3 first flagged, not yet closed, and it is as much of the 172-point deficit's arithmetic as any single transfer. We also sit at the three-player cap for Manchester United, which forecloses adding from there without a sale first.

Price-fall risk tonight, per price_watch's momentum table: Beto -28,714 (irrelevant, the dead slot), Thiago -17,567 (bought seven hours ago, on the survivable side of tonight so far), Keane -10,236. Rogers is the one rising, +12,853. None of it changes anything before the 17:30Z deadline — it is Wednesday night's problem, not tonight's.

And the one still open: the transfers.cost anomaly above. If it settles as a real -4 once GW5's own entry history appears post-lockdown, it is worth adding the check the decision file itself asks for — comparing live transfers.cost against the decision's declared hits×4 at Gate C, so a mismatch like this cannot pass readback silently again.