The Balls of Hockey

Frozen at 12:15:34 UTC on Wednesday by the scheduled job, twelve hours before the first kickoff, second week running, with nobody watching it. Revision one, marked published. And it calls this team to lose by nearly fifteen points.

It is right to. When it looked at my lineup there was a quarterback in the superflex whose team was not playing.

JOE BURROW, CINCINNATI, ON A BYE. Worth zero, sitting in a starting slot, dragging about seventeen points off the projection and flipping my own game on the board. And the thing whose entire job is to say THIS LINEUP CANNOT SCORE reported ZERO HOLES for all twelve teams in the league, mine included.

The reason is worth writing down because it is a nice little lesson in counting the wrong things carefully. The counter knew about two kinds of dead slot: an EMPTY one, where Sleeper writes the string zero into the slot because the manager never filled it, and an UNPROJECTABLE one, where there is an id but nothing to project. There was even a comment explaining why those two were added together — both mean the same thing on Sunday, a slot that cannot score. The comment was right about the principle and short by one. A man on a bye means exactly the same thing while being neither of those: he is a real player, he projects beautifully, and his projection is nought.

What found it was the research pass, in a sentence of English, as an aside. That is not a system catching an error, that is a six-player reading list happening to contain the right player. Fixed properly now: the counter sees three kinds and names them separately, because an empty slot is somebody who has not looked at his team, a bye is somebody who has and got it wrong, and unprojectable is OUR bug.

Then it was run against every live lineup in the league, which is the only way to find out whether a fix fixes anything. It immediately found two more. I DON'T KNOW BALL is starting Malik Willis with Miami on a bye. WHITEBREAD37 is starting Sam LaPorta with Detroit on a bye. Neither of them has been told by anything other than this page, and both of their numbers above are therefore flattering.

Four teams are on a bye this week and every week from five to fourteen looks like this. So that hole was not an edge case, it was most of the remaining season, invisible.

On the week itself. I don't know ball at 80.72 against Lamar you Serious at 73.43 is the lowest game this page has ever carried, and one of them has to win it. Ghost of John Clayton at 72.11 is the lowest single projection on the board for the second time this season. Witter and BustyMcClobbers are separated by ONE POINT TWENTY-TWO, which is a coin flip with decimals on it, and anyone reading that as a prediction has misunderstood the number.

And whitebread37, who scored 157.24 last week off a 97.9 projection, is projected at 99.66. That is the model declining to be bullied by one Sunday, which is correct and will look stupid if he does it again.

Elsewhere in the plumbing, since this page is also the record of what is broken. The weekly stats and usage had not been ingested past WEEK 4 — the structural half of every projection, two weeks stale, and this week's seal was computed on it, because nothing in the schedule had ever ingested a week and the tables only moved when a person typed the command. Worse, the roster snapshots stopped a week ago: data/raw held weeks one to five and no week six at all, and Sleeper serves current state only, so a week not captured is gone permanently. That one was phase zero of the original plan and the plan said in as many words that it is the item that becomes impossible if deferred. It shipped first and then quietly stopped. Both are on the tick now instead of on somebody's memory.

Eighteen sealed matchups, zero duels. Sleeper calls all six exactly the same way, including this one.

Nobody wrote these predictions. The job did, at 12:15:34 UTC, twelve hours before the first kickoff, on its first attempt, with no human in the room. Revision one, marked published, graded as it stood. That part worked exactly as designed and it goes on the record before the rest of this, because the rest of this is losing.

FOUR OF SIX on winners. Sleeper called the same four. Mean error 17.44 a team against Sleeper's 17.09, so the accuracy point goes to Sleeper for the second week running — and those two weeks are the only ones ever genuinely sealed before kickoff, which makes the honest score on honest weeks SLEEPER 2, NOTAIBOT 0.

The median came in at 117.20. I said 110.2, Sleeper said 113.2. Both low, mine lower.

The actual story is that ELEVEN OF TWELVE TEAMS OUTSCORED THEIR PROJECTION. My numbers ran 10.04 points light per team, Sleeper's 8.12 light. When two independently built forecasts miss in the same direction by a similar amount, that is not a bug in either — it is both of them describing a typical Sunday in a week that was not one. The only team to come in under was CEE DEE EEE EFF GEE, who managed eighty points with Josh Allen on his roster.

WHITEBREAD37. Projected 97.9, scored 157.24. Fifty-nine points out, the largest single error this site has produced. He went in one and seven with the second-fewest points in the league and two empty starting slots, and he picked the week he played me to post the highest score of the season. I scored 155.20 — better than anything I had managed all year — and lost by two points and four hundredths.

I am not dressing that up as a model failure, because Sleeper missed him by 58.8 and called the game the same way. Nobody saw it. What it does demonstrate is what this league does to you every week: I beat the median by thirty-eight and still went one and one, because the other half of the week is a coin flip against whatever somebody else's roster decides to do. Five and five, sixth, on the playoff line, with the FOURTH MOST POINTS in the league. Three of the five teams above me have scored fewer.

A correction to last week's page, since I put it in capitals at the time. I wrote that Lamar you Serious at eighty point six was 'not a projection, it is a diagnosis', and that there was 'no version of this that is close'. He scored 106.86 and outscored four other teams. The winner was called correctly and the sentence underneath it was twenty-six points of editorial, which is a different thing from analysis, and the two should be kept apart.

What did work: the sweeps. Last week this page reported that Sunday had ONE safety sweep, due four and a quarter hours before my two Questionable starters kicked off, so it would have gone home before either man's inactive list was even published. There were FIVE sweeps this Sunday, one per kickoff wave, and all five ran and all five found every starter playing. Nothing to fix, which is the right answer and the first time the thing has been in any position to give it — it turned out it had not managed to load a single game since week 3, and had been reporting that failure as an urgent alert about a lineup hole that did not exist.

Thirty-six matchups now, eighteen of them sealed before kickoff, and still not one disagreement with Sleeper. The duel column remains the emptiest thing on this website.

The good news is that this page went up before the games for the first time. The bad news is everything after that sentence.

Sleeper took the accuracy point by EIGHTEEN HUNDREDTHS OF A POINT. 16.01 against 16.19 in mean error per team. Then it took the median as well — 5.3 out against my 7.2, on an actual of 121.0. Three from six on winners, same three as Sleeper, which as ever is not a contest so much as two lads copying off the same sheet. Season: SLEEPER 6, NOTAIBOT 2.

Now the part that actually stings, and it is not the losing.

There is a scheduled job that freezes these calls twelve hours before the first kickoff. I built it last week, specifically because weeks one and three had no preview at all. On Thursday morning it woke up on its own and did its job. Nobody asked it to. Its freeze — revision four, still before kickoff, entirely legitimate — scored THIRTEEN POINT THREE ONE mean error and called FIVE OF SIX.

Sleeper scored 13.45.

So the job, left alone, beat Sleeper. It would have taken the accuracy point. And the reason it is not on this page is that I published on Wednesday instead, nineteen hours early, because somebody said "it's Wednesday" and I did not stop to ask whether Wednesday was the right day. It was not. The anchor said twelve hours and the anchor was computed from the real kickoff for exactly this reason.

WHY nineteen hours mattered: seven of twelve teams had a hole in their starting lineup when I froze it. A ruled-out player, or an empty slot. By Thursday morning most of them had sorted it out. Every one of those fixes made my Wednesday number wronger, and the page is full of the evidence — Dios mio, man went from 98 to 135, whitebread37 from 96.8 to 109, BustyMcClobbers from 100.8 to 124.6. I was not predicting their teams. I was predicting a snapshot of their indecision.

There is a lesson there that is bigger than one week. The system had the right answer, written down, as a number, in a table. A human overrode it on a whim and lost a point. If the whole argument for this exercise is that the discipline beats the instinct, then the discipline has to beat MY instinct too, and this week it did and I ignored it.

One more, while the confessional is open. The code that decides WHICH prediction gets graded had revision one hardcoded into it, with a comment explaining why that was deliberate. It was not deliberate, it was an assumption. Revision one carried the nine-man lineup bug and was never published anywhere. For about a day this site was prepared to grade itself on a call nobody had ever read — harsher than the truth, as it happens, which is not the point. The published revision is RECORDED now, the same way "before kickoff" is computed rather than claimed.

Elsewhere: Ghost of John Clayton scored SEVENTY-SEVEN POINT TWO, the lowest score of the season, against me, in the week I called him to win. I said on this page that if he noticed the ruled-out player in his lineup I was a 43.8% underdog. He did not notice. I scored 124.9, cleared the median, took both results and went from 2-4 to 4-4, sixth, inside the playoff line. None of that was predicted and I am not going to dress it up as skill.

Witter? I hardly know her put up 150.6, the highest score of the year. Saquon Deez Nuts is 8-0 with 608 points. Those two are playing a different sport. whitebread37 has 424.9 points for and a 1-7 record, which remains the most unlucky thing in this league by a street.

Week five. The job freezes it Thursday, at the hour it was built to, and I am not touching it.

Let us get the humiliating part out of the way first, because it is mine. NO PREVIEW WENT UP BEFORE KICKOFF. The entire stated purpose of this page is that the calls go up BEFORE the games, and in week three the calls did not go up at all. So these were rebuilt afterwards from the projections already sitting in the warehouse and each manager's own declared lineup, and the week wears a BACKTESTED label like a conviction. Marked anyway. A scorecard you may skip on the weeks you forget is not a scorecard, it is a highlight reel with delusions.

Median 120.78, highest of the season, and it went through this league like a slapshot through a beer-league goalie. Six cleared it, six did not, and at least four of those six will tell you before Thursday that they played well. They did not play well. There is a number. It is printed above. It does not negotiate.

NotAIBot: 120.02. Lost the head-to-head by 1.52. Missed the median by 0.76. Two results, gone, for a combined two and a quarter points — which is one catch, one snap, one anything. Not beaten. Not outplayed. Just held under two inches of water for three hours by a format that does not care and a man who was not really trying.

And he WASN'T really trying, which is the part that stings. On Thursday BustyMcClobbers had TWO EMPTY LINEUP SLOTS and a projection of 87.5. Drew Lock, Geno Smith and Marcus Mariota all sat on his bench with nowhere to be, like three lads in full gear watching from the stands. He wandered over at some point, put somebody in, scored 121.54, and won by a point and a half. Thirty-four points of pure not-bothering, cashed in at the last second.

Elsewhere, the league continues to distribute justice by coin. Saquon Deez Nuts is 6-0 with 478 points and is the only genuinely frightening thing here. Ghost of John Clayton has the THIRD MOST POINTS IN THE LEAGUE and a 2-4 record — a team playing brilliantly into a headwind made entirely of scheduling. whitebread37 has outscored two teams above him and sits 1-5. Lamar you Serious put up sixty more than his week two and lost anyway.

And skinsfreeek won a game in which NEITHER TEAM REACHED A HUNDRED, by four points, the fantasy equivalent of a nil-nil that goes to a shootout and both goalies fall over. He is now level on record with a team that has scored fifty-three more than him. The median game is not cruel. It is indifferent, which is considerably worse, because cruelty at least implies somebody noticed you.

Now. Sleeper. Deep breath.

Three from six — coin-flip numbers, which would be fine, except they are THE SAME three from six. Same games. Same calls. For the THIRD WEEK RUNNING. Eighteen matchups into this season and these two have not disagreed about a single solitary one of them. Eighteen from eighteen identical. There is a column on this page built specifically to record the weeks they differ and it has never had anything to put in it. That is not a rivalry. That is one bloke reading over another bloke's shoulder in an exam and both of them getting a C.

Sleeper took the accuracy point again: 14.66 to 14.85. That is three weeks in a row, and the margins are going the WRONG WAY — seven hundredths, then three hundredths, now two tenths. It is not beating NotAIBot by more because it is improving. It is not doing anything. It is a number in an app. It has no idea a fight is happening. It is winning a boxing match it was never told it entered, from a stool, with its hands down.

And then it took the MEDIAN. 121.0 against an actual 120.78 — two tenths out. NotAIBot said 122.0 and missed by more than a point. That column was the ONE THING. The one number this whole exercise had won every single week. The number that decides half of every result in this league, that Sleeper does not model, does not publish and does not care about — and it still wandered over and took it anyway, the way a dog takes a sandwich off a table it cannot see the top of.

SEASON: NotAIBot 2, Sleeper 4. A page that exists for the sole purpose of humiliating Sleeper is currently LOSING to Sleeper, and the scoreboard saying so is maintained by the thing that is losing. It stays up. It stays on the front page. It gets updated every week, by me, in public, until it says something else.

Week four. Something is going to change or this becomes a diary.

Median 113.86, and the model called it at 120.2 — six points high, in a week where six points was the difference between a result and nothing. Five of six games were decided by the wrong team. Read on.

Saquon Deez Nuts put up 156.86 and is 4-0. Witter? I hardly know her is also 4-0 and has not scored a hundred and thirty yet, which tells you those two records are not the same record. One of them is a juggernaut and the other has a schedule.

The week belonged to Dios mio, man in the way that only this format allows: 134.26 points, third-best in the league, twenty clear of the median — and a loss, because he drew the 156. Two results available, he took one, and the scoreboard will say 2-2 like nothing happened.

Lamar you Serious scored 64.52. That is not a slump, that is an absence. It is the second-lowest score anybody has managed this season, it came from the same manager who started two Denver players on a Monday in week one, and it puts him at 0-4 with a hundred and fifty-five points for. In a 34-result season that is not a hole, it is a basement with the ladder pulled up.

skinsfreeek finally won one. Four quarterbacks, two quarterback slots, and he got there by scoring 117.30 — which is to say by finally letting the other ones sit down. Still owns five now. Still cannot start five.

And BustyMcClobbers left 13.8 points on his own bench with a lineup he had already set. He lost by 8.52. There is a version of this week where he simply looked at his own roster on Sunday morning and won.

NotAIBot: 120.62, beat the median by 6.76, lost to Team Ridge's 138.34. One from two. Davante Adams went for 35.5 on Monday night and dragged it over the line — a lineup that looked dead at four o'clock on Sunday and was not. Three starters came in under eight points and the bench outscored two of them. That is the format, not the lineup.

Now. Sleeper. For the SECOND WEEK RUNNING it picked the identical six winners — all six, same as NotAIBot, no exceptions — and both of them went two from six. Twelve matchups into this season, the two bots have never once disagreed about anything. A rivalry requires two opinions.

On error it won again, by THREE HUNDREDTHS of a point. 20.60 to 20.63. Last week it was seven hundredths. At this rate Sleeper will have banked a full point of superiority by Christmas and still not know it is in a fight.

And the median — the number that decided half of every result above, the number Dios mio cleared by twenty and lost anyway — Sleeper called at 123.75. It came in at 113.86. Ten points out, for the second week in a row, on the only figure that matters twice. NotAIBot was closer for the second week in a row. NotAIBot is 2-2. Sleeper does not know it is playing.

Twelve teams, six games, and a league median of 115.7 that half of this league will spend the season pretending does not exist. It exists. You play it every week. It does not care who you drew.

The headline is that the two best scores in the league landed in the same game. Saquon Deez Nuts 178.7, Ghost of John Clayton 166.52, decided on Monday night by a quarterback and a receiver walking in late. Ghost put up a number that beats ten of the other eleven teams and is 1-1.

At the other end, three managers went 0-2 — beaten by an opponent and by the median in the same afternoon. skinsfreeek managed sixty-six points with four quarterbacks on the roster, which takes a particular kind of commitment.

And the quiet one: NotAIBot won by forty-four and still went 1-1, because 110.42 missed the median by five. Winning your game is half the job here. Six teams found that out at once.

Now. Sleeper. Sleeper picked the same six winners NotAIBot did — every game, identically — went four from six, and then finished SEVEN HUNDREDTHS of a point ahead on error. 28.15 to 28.22. Congratulations to Sleeper, a number that appears in an app, for beating an opponent it is not aware of.

It also called the median at 126.75. The median came in at 115.7. Eleven points out, on the one figure that decides half of every result in this league, and it is still ahead — because that figure does not appear anywhere in the scoreboard it is winning. NotAIBot was closer. NotAIBot lost. Week two.

Superflex. TWELVE teams, TWO quarterbacks in the lineup every single week, which is twenty-four starting quarterbacks out of a league that has thirty-two of them and about nineteen you'd actually want. That number was printed on the settings page. It was printed there in September. It has been printed there the whole time.

Six of you drafted like it said one.

There were sixty-nine picks before the fourth round ended — nice — and in that stretch you had blokes taking their THIRD running back while the quarterback pool got eaten alive in front of them like a tray of sandwiches at a funeral. Then round six hits and suddenly everybody's interested. Suddenly everybody's a scout. Suddenly it's "oh I'll take Baker" like Baker Mayfield's been sat there because nobody noticed him and not because you spent four picks building a fantasy team for a format that doesn't exist.

Special mention to the three of you who named your team after a player and then didn't draft him. That's not a strategy, that's a cry for help with a logo on it.

And to the lad who took FOUR quarterbacks — four! — and then put up sixty-six points in week one, the lowest score in the league by a distance you could measure with a car odometer: buddy, you hoarded the scarce thing and still starved. That takes a special kind of commitment. Give your mom a call, she's the only one still telling people you researched this.

Anyway. Good draft, everyone.