The ATL vs. GB game inspired this AI-aided discovery flow. I haven't solved anything yet. I just wanted to share my thinking on this issue: a breakable assumption(s). 

The Line Is a Prior. Week 3 Is Where It Starts to Crack.

Thursday night was not a soft open for the model. Green Bay closed a 4.5-point favorite at Lambeau, and the Premortem Value readout was more sure of the Packers than the market was. Vegas had them at 67 percent to win. The model had them at 89. Atlanta won, 35–14.

That is the right place to start, because Premortem Value isn't a pick sheet, and it isn't a secret ranking of “who's good.” It is a readout of how much of a closing line is still a named, breakable assumption. The math that turns tape, line movement, and injury news into that readout stays in the engine. Call it the secret sauce. What follows is the sauce's Week 3 printout, and what it's willing to say out loud.

Premortem Value (PMV) scores a game before kickoff by how much of its line is still a named, historically breakable assumption. It does not estimate who covers. It estimates how much failure information is still in the number.

Does a higher score mean less failure information in the number?

No. A higher Premortem Value means more failure information is still in the number.

The score measures how much of the line is still a named, breakable assumption. A high number means several of those assumptions are stacked. A low number means the market, the tape, the public, and the injury report are telling the same story, so there is less left to be specifically wrong about.

Score Meaning

Low: Little failure information. Do not invent a story.

High A lot of failure information. Tuesday can confirm it or kill it.

=======================================================

The two failure modes

A, steam overshoot. The market already updated. Points rise with the size of the move toward the closing favorite and with a crossing of 3 or 7. If the efficiency gap still wants the favorite by 4 or more points, that move is catch-up, not overshoot, and A is halved. The attack location is the dog.

The weight rests on TeamRankings since 2008: an opening favorite that steamed 3 points covered 45.6% (n=83); a 4-point steam covered 41.5% (n=67). Unconditionally, the opening favorite covered 49.0%.

B, identity lag. The market did not update. Points rise with the absolute efficiency gap, plus adders when a 2-0 team is still a dog. If the line has already moved 3 or more toward that new identity, cut B by 30%. The attack location is the team the line is still refusing to respect.

Net = B − A.

• Net above +3 and the other term is not also large: B-LAG.

• Net below −3: A-STEAM.

• Both A and B at least 9 or 10: MIXED. Both stories are live. Do not force a side.

• Otherwise: BALANCED. Little directional information.

A quarterback-out on the closing favorite that has not moved 3.5 points is forced toward the dog. That injury is underpriced information, not proof the favorite’s old identity is correct.

Public and injury, briefly

Tickets at 80% or more score the full 15. The bands step down through 70, 65, and 60. A ticket-versus-handle gap of 15 points or more adds 2, capped at 15. The 2025 reference class: Covers sides at 60% or more of picks went 42% against the spread; 65% or more went 31%. That rate isn't stable year to year, which is why the public is only 15 of 95.

Injury is 10 only when the starter is out, and the close has not already moved about 4 points. If it has, the points are cut, because the absence is already in the price.

=========================================================

What the columns mean

Every game on the board carries two labels.

Mode is the kind of argument still sitting in the number.

A B-LAG game is one where the market looks late. A team has already shown something, a 2–0 start or a statement win, and the line has not fully accepted it. The old identity is still the price.

An A-STEAM game is the opposite. The number has already jumped toward a favorite, sometimes through the numbers bettors feel in their stomach, and the jump itself is the thing that might be wrong.

MIXED means both arguments are alive in the same game. The index will not choose a side for you.

BALANCED means the line, the recent tape, and the public are roughly telling the same story. There is less to interrogate.

Attack surface is not a bet. It is the side that gets embarrassed if the named argument is true. 

A team name means one side is exposed. 

BOTH LIVE means the game is a knife fight between two failure stories. 

NONE means the board has no side to expose.

The win probabilities next to those labels are the model's, set beside Vegas on purpose. Where they match, the week is quiet. Where they split, the week is the article.

Where the model is louder than Vegas

Four favorites come out of the sauce hotter than the market.

Kansas City is a 10.5-point favorite at Miami, 83 percent in the logistic the market implies, 96 percent in the model. Buffalo is laying 7 to the Chargers, 75 percent versus 97. San Francisco is laying 8.5 to Arizona, 79 versus 94. Seattle is laying 7.5 in Washington, 76 percent versus 90 percent. Carolina is the quieter version of the same shape: a short road favorite, 60 percent in the market, 80 percent in the model, tagged B-LAG with the Panthers as the exposed side.

These are not four copies of one opinion. Kansas City, San Francisco, and Carolina are B-LAG. In the model's view, the market still hasn't caught the team it is already favoring. Buffalo and Seattle are MIXED, both live. The number moved hard, the public is piled on the favorite, and the tape still says the favorite might be the better team by more than the line. Fading them because the line moved, and backing them because the line did not move enough, are both still legal arguments. The index refuses to pick.

Where the model walks away from the favorite

Five favorites are under 50 percent in the model while still being the Vegas favorite. That is the actual news in the file.

Detroit is laying 6.5 to the Jets. The market has the Lions at 73 percent to win. The model has them at 40, and the Jets at 60. Mode is B-LAG. Attack surface is the Jets. The line opened nearer a touchdown and came backward. The readout treats that retreat as unfinished.

New Orleans is laying 3 to the Raiders. Vegas has the Saints at 62 percent. The model has the Saints at 28 and the Raiders at 72. That is the cleanest identity-lag game on the sheet: a 2–0 team still priced like last year's version of itself, and the attack surface is Las Vegas.

The other three are narrower, and they should be read as narrow. Houston is a 1.5-point favorite at Indianapolis, 57 percent in the market, 48 percent in the model, and the mode is BALANCED with no attack surface. The model is shrugging. Los Angeles is laying 2.5 at Denver, 60 versus 48, B-LAG, attack surface the Broncos. Philadelphia is laying 4.5 at Chicago on Monday night, 67 versus 45, and the tag is MIXED, both live. A hamstring changed the number. Whether the number changed too much is exactly the argument the index will not settle before kickoff.

Green Bay belonged on this list in spirit and did not make it in the file. The model did not walk away from the Packers. It leaned in, harder than Vegas, and Thursday cashed the other side. Anyone using the Sunday sheet as scripture should have to say that sentence first.

The short list of single-sided games

Strip out MIXED and BALANCED, and the board still has a one-sided exposure on ten games. Exposure is not an instruction. It is the name of the assumption.

Game

Mode

If the assumption breaks, it breaks on

Packers–Falcons

B-LAG

Packers. It already broke.

Chiefs–Dolphins

B-LAG

Chiefs

Bengals–Steelers

A-STEAM

Steelers

Lions–Jets

B-LAG

Jets

Panthers–Browns

B-LAG

Panthers

Jaguars–Patriots

A-STEAM

Patriots

49ers–Cardinals

B-LAG

49ers

Vikings–Buccaneers

A-STEAM

Buccaneers

Ravens–Cowboys

B-LAG

Cowboys

Saints–Raiders

B-LAG

Raiders

Rams–Broncos

B-LAG

Broncos

=======================================================

The A-STEAM trio is the public-looking side of the sheet. Cincinnati moved from a short number to 3.5 at Pittsburgh, and the exposed side is the Steelers. Jacksonville is a field-goal favorite against New England; exposed side: the Patriots. Minnesota flipped from dog to a short favorite at Tampa Bay, exposed side the Buccaneers. In each case, the claim is modest: the move is the thing that might be early, not a verdict that the favorite cannot win.

Baltimore in Brazil is the smallest B-LAG on the board. The market has the Ravens at 64 percent. The model has them at 53, with Dallas as the exposed side. International games are where home-field habits go to get audited. This one is an audit, not a shout.

Texans–Colts is the game the index cares least about. Two winless teams, a 1.5-point line, mode balanced, attack surface: none. If you need a story there, it will not come from this readout.

=========================================================

How to read a score

PMV Meaning

0–24 Market, efficiency, public, and injury agree. Do not invent a story.

25–44 One team is live. Wait for Saturday.

45–64 A named failure is load-bearing. Read its falsifier before doing anything with it.

65–100 Several terms are stacked. Tuesday either supports the index or kills the weights.

Moneyline rule. 

PMV only becomes a win bet inside a field goal. Beyond 3.5 points, the story is about the cover, so the moneyline is a pass.

Action  When

Team ML: B-LAG, PMV ≥ 45, the lagged side is the favorite, spread is 3.5 or less

Dog ML: B-LAG, PMV ≥ 45, the lagged side is the dog, spread is 3.5 or less. Same action on A-STEAM if the dog is 3.5 or less and PMV ≥ 45, because missing the cover and losing are the same bet

Pass

MIXED, Balanced, PMV under 45, or any spread past 3.5

The Line Is a Prior. Week 3 Is Where It Starts to Crack.

Thursday night was not a soft open for the model. Green Bay closed a 4.5-point favorite at Lambeau, and the Premortem Value readout was more sure of the Packers than the market was. Vegas had them at 67 percent to win. The model had them at 89. Atlanta won, 35–14.

That is the right place to start, because Premortem Value is not a pick sheet and it is not a secret ranking of “who's good.” It is a readout of how much of a closing line is still a named, breakable assumption. The math that turns tape, line movement, and injury news into that readout stays in the engine. Call it the secret sauce. What follows is the sauce's Week 3 printout, and what it's willing to say out loud.

Premortem Value (PMV) scores a game before kickoff by how much of its line is still a named, historically breakable assumption. It does not estimate who covers. It estimates how much failure information is still in the number.

The two failure modes

A, steam overshoot. The market already updated. Points rise with the size of the move toward the closing favorite and with a crossing of 3 or 7. If the efficiency gap still wants the favorite by 4 or more points, that move is catch-up, not overshoot, and A is cut in half. The attack location is the dog.

The weight rests on TeamRankings since 2008: an opening favorite that steamed 3 points covered 45.6% (n=83); a 4-point steam covered 41.5% (n=67). Unconditionally the opening favorite covered 49.0%.

B, identity lag. The market did not update. Points rise with the absolute efficiency gap, plus adders when a 2-0 team is still a dog. If the line has already moved 3 or more toward that new identity, B is cut 30%. The attack location is the team the line is still refusing to respect.

Net = B − A.

• Net above +3 and the other term is not also large: B-LAG.

• Net below −3: A-STEAM.

• Both A and B at least 9 or 10: MIXED. Both stories are live. Do not force a side.

• Otherwise: BALANCED. Little directional information.

A quarterback-out on the closing favorite that has not moved 3.5 points is forced toward the dog. That injury is underpriced information, not proof the favorite’s old identity is correct.

Public and injury, briefly

Tickets at 80% or more score the full 15. The bands step down through 70, 65, and 60. A ticket-versus-handle gap of 15 points or more adds 2, capped at 15. The 2025 reference class: Covers sides at 60% or more of picks went 42% against the spread; 65% or more went 31%. That rate is not stable year to year, which is why public is only 15 of 95.

Injury is 10 only when the starter is out and the close has not already moved about 4 points. If it has, the points are cut, because the absence is already in the price.

What the columns mean

Every game on the board carries two labels.

Mode is the kind of argument still sitting in the number.

A B-LAG game is one where the market looks late. A team has already shown something, a 2–0 start or a statement win, and the line has not fully accepted it. The old identity is still the price.

An A-STEAM game is the opposite. The number has already jumped toward a favorite, sometimes through the numbers bettors feel in their stomach, and the jump itself is the thing that might be wrong.

MIXED means both arguments are alive in the same game. The index will not choose a side for you.

BALANCED means the line, the recent tape, and the public are roughly telling the same story. There is less to interrogate.

Attack surface is not a bet. It is the side that gets embarrassed if the named argument is true. A team name means one side is exposed. BOTH LIVE means the game is a knife fight between two failure stories. NONE means the board does not have a side to expose.

The win probabilities next to those labels are the model's, set beside Vegas on purpose. Where they match, the week is quiet. Where they split, the week is the article.

Where the model is louder than Vegas

Four favorites come out of the sauce hotter than the market.

Kansas City is a 10.5-point favorite at Miami, 83 percent in the logistic the market implies, 96 percent in the model. Buffalo is laying 7 to the Chargers, 75 percent versus 97. San Francisco is laying 8.5 to Arizona, 79 versus 94. Seattle is laying 7.5 in Washington, 76 versus 90. Carolina is the quieter version of the same shape: a short road favorite, 60 percent in the market, 80 percent in the model, tagged B-LAG with the Panthers as the exposed side.

These are not four copies of one opinion. Kansas City, San Francisco, and Carolina are B-LAG. The market, in the model's view, still has not caught the team it is already favoring. Buffalo and Seattle are MIXED, both lives. The number moved hard, the public is piled on the favorite, and the tape still says the favorite might be the better team by more than the line. Fading them because the line moved, and backing them because the line did not move enough, are both still legal arguments. The index refuses to pick.

Where the model walks away from the favorite

Five favorites are under 50 percent in the model while still being the Vegas favorite. That is the actual news in the file.

Detroit is laying 6.5 to the Jets. The market has the Lions at 73 percent to win. The model has them at 40, and the Jets at 60. Mode is B-LAG. Attack surface is the Jets. The line opened nearer a touchdown and came backward. The readout treats that retreat as unfinished.

New Orleans is laying 3 to the Raiders. Vegas has the Saints at 62 percent. The model has the Saints at 28 and the Raiders at 72. That is the cleanest identity-lag game on the sheet: a 2–0 team still priced like last year's version of itself, and the attack surface is Las Vegas.

The other three are narrower, and they should be read as narrow. Houston is a 1.5-point favorite at Indianapolis, 57 percent in the market, 48 percent in the model, and the mode is BALANCED with no attack surface. The model is shrugging. Los Angeles is laying 2.5 at Denver, 60 versus 48, B-LAG, attack surface the Broncos. Philadelphia is laying 4.5 at Chicago on Monday night, 67 versus 45, and the tag is MIXED, both live. A hamstring changed the number. Whether the number changed too much is exactly the argument the index will not settle before kickoff.

Green Bay belonged on this list in spirit and did not make it in the file. The model did not walk away from the Packers. It leaned in, harder than Vegas, and Thursday cashed the other side. Anyone using the Sunday sheet as scripture should have to say that sentence first.

The short list of single-sided games

Strip out MIXED and BALANCED, and the board still has a one-sided exposure on ten games. Exposure is not an instruction. It is the name of the assumption.

Game

Mode

If the assumption breaks, it breaks on

Packers–Falcons

B-LAG

Packers. It already broke.

Chiefs–Dolphins

B-LAG

Chiefs

Bengals–Steelers

A-STEAM

Steelers

Lions–Jets

B-LAG

Jets

Panthers–Browns

B-LAG

Panthers

Jaguars–Patriots

A-STEAM

Patriots

49ers–Cardinals

B-LAG

49ers

Vikings–Buccaneers

A-STEAM

Buccaneers

Ravens–Cowboys

B-LAG

Cowboys

Saints–Raiders

B-LAG

Raiders

Rams–Broncos

B-LAG

Broncos

=======================================================

The A-STEAM trio is the public-looking side of the sheet. Cincinnati moved from a short number to 3.5 at Pittsburgh, and the exposed side is the Steelers. Jacksonville is a field-goal favorite against New England, exposed side the Patriots. Minnesota flipped from dog to a short favorite at Tampa Bay, exposed side the Buccaneers. In each case, the claim is modest: the move is the thing that might be early, not a verdict that the favorite cannot win.

Baltimore in Brazil is the smallest B-LAG on the board. The market has the Ravens at 64 percent. The model has them at 53, with Dallas as the exposed side. International games are where home-field habits go to get audited. This one is an audit, not a shout.

Texans–Colts is the game the index cares least about. Two winless teams, a 1.5-point line, mode balanced, attack surface: none. If you need a story there, it will not come from this readout.

=========================================================

How to read a score

PMV Meaning

0–24 Market, efficiency, public, and injury agree. Do not invent a story.

25–44 One team is live. Wait for Saturday.

45–64 A named failure is load-bearing. Read its falsifier before doing anything with it.

65–100 Several terms are stacked. Tuesday either supports the index or kills the weights.

Moneyline rule. 

PMV only becomes a win bet inside a field goal. Beyond 3.5 points, the story is about the cover, so the moneyline is a pass.

Action  When

Team ML: B-LAG, PMV ≥ 45, the lagged side is the favorite, spread is 3.5 or less

Dog ML: B-LAG, PMV ≥ 45, the lagged side is the dog, spread is 3.5 or less. Same action on A-STEAM if the dog is 3.5 or less and PMV ≥ 45, because missing the cover and losing are the same bet

Pass

MIXED, Balanced, PMV under 45, or any spread past 3.5

Deeper Discussions

The Premortem Value number is the failure-information level. It is not the chance that either team wins.

A 73 means a large share of that line is still a named, breakable assumption. A 14 means almost none of it is. Reading 73 as "73 percent this team wins" is the wrong question.

Win probability answers who is likely to win the game. Premortem Value answers how much of the price can still be specifically wrong. Those move apart on purpose.

Game shape

Win probability

Premortem Value

Tape and price agree, and the favorite is clearly better

High for the favorite

Low. Little left to attack.

Favorite is clearly better, and the price still lags that tape

High for the favorite

High. The assumption is that the market has caught up. It may not have.

Favorite is the Vegas side, but the failure story sits on the dog

Can be under 50 for the favorite

High. The information is about the exposed side, not about crowning the favorite.

Both a steam story and a lag story are alive

Either side can lead

High, and the mode is mixed. The level is real. It does not pick a winner.

So Kansas City can be very likely to win and still carry a high Premortem Value, because "likely winner" and "the line still contains a lag" are different claims. Detroit can be the favorite in the market while the failure information is sitting on the Jets. The level does not flip into a win probability just because it is written as a number from 0 to 100.

The other percentages in the file, Vegas and the tape-implied chances, stay win probabilities. Do not add them to Premortem Value, and do not relabel them as failure levels.

The moneyline and the spread are not the same claim, and Premortem Value applies only to the second.

Kansas City laying 10.5 at Miami is a useful case because the two claims are easy to pull apart. Both the market and the tape already agree on the game's sign. The Chiefs are the better team by enough that a loss is the tail, not the story. That is what a win probability near the mid-80s or higher is saying. Once a team is that far to one side of zero, another point of confidence barely changes the moneyline. Most of the remaining argument is about the margin.

The lag is a claim about where the distribution sits, not about which side of zero it is on. The market has a number. The recent body of work has a different, larger gap. The difference is the assumption still inside the price: that two weeks have already been absorbed, and that this Chiefs team and this Dolphins team are the ones the number describes. Premortem Value is high when that assumption is still nameable and still the kind of assumption that has broken before. It does not rise because Kansas City might lose. It rises because the price may still be an older, smaller Chiefs, or an older, better Miami, or both.

That is why the same game can carry a high chance of a Kansas City win and a high failure-information level at the same time. The winner can be obvious while the margin is stale. Three different Sundays all count as "the Chiefs were likely to win":

• They win by 3. The win-probability claim was right. The lag claim was wrong. The market's smaller gap was the better description, and the high Premortem Value was failure information that did not pay.

• They win by 20. The win-probability claim was right, and so was the lag. Same winner, opposite verdict on the assumption the index actually scored.

• Miami wins outright. Both claims fail. That is the tail. A high Premortem Value was never a warning about it, and it was never protection against it.

The third one is what Thursday already demonstrated with Green Bay. The model was more sure the Packers would win than Vegas was, and the failure information was a lag, not a doubt. Collapsing those into one moneyline bet is how a margin argument gets asked to survive a blowout loss. It cannot. The information was about respect missing from the number. It was not about the identity of the winner.

A second split exists inside the Chiefs game, and it matters. Part of a stale price is the favorite. Part of it is the opponent. A double-digit line is a statement about Miami as much as about Kansas City. If the Dolphins' side of the number still carries last year's competence, the lag can live entirely in the dog while the Chiefs remain an obvious winner. Fading Kansas City, or backing them on the moneyline because the index is "high," answers a question the score did not ask.

Markets do this unevenly. They update fast when a result confirms a strong prior, and slowly when a result says the prior was too small. A famous favorite who keeps winning can be under-updated precisely because the market already believed in them and only nudges. The public often does the opposite with stars and over-updates. Both are stories about price. Neither one reopens the question of who is likely to win.

So the deep relation is this. Win probability is the mass on one side of zero. Premortem Value is the amount of named, breakable assumptions left in the location of that mass. Kansas City can have almost all the mass and still have a stale location. Until you keep those separate, a high Premortem Value on a big favorite will keep getting misread as a hot pick or a trap. It is neither. It claims the margin hasn't finished catching up, and Sunday can confirm that or kill it without the winner changing.

A breakable assumption is a specific belief still priced into the line that Tuesday can prove false without rewriting the game.

It has three parts.

It is named. Not "someone might be wrong." A named thing: the market already paid for this steam, this 2–0 team is still the old team, this public pile is the right side, this injury is already in the number.

It is still in the price. If the line has already moved to match the tape, the belief is no longer an assumption. It is the market.

It has failed before in this shape. A 3-point steam through a key number, a statement win still priced like a dog, a quarterback ruled out on a line that barely moved. The history is what makes it breakable rather than merely uncertain.

"The Chiefs might lose" is not one. Every game has that. "Kansas City's margin is already fully updated" is one. They can win by 3, and the sentence is false; or they can win by 20, and the sentence is true. The winner never had to change.

If you cannot say the sentence before kickoff, and you cannot say what result kills it, it is not a breakable assumption. Premortem Value only scores those.

Each game can carry up to five breakable assumptions: the move went too far, the market still has the old team, the tape and the price disagree, the public is piled on one side, or a regime-change injury is not in the number. A class counts only when it is still a full sentence. A small line wobble, a ticket lean, or a skill-player tag does not.

Game

Count

What is still breakable

Packers–Falcons (final)

3

Old identity, tape versus price, public pile

Chiefs–Dolphins

4

Steam, old identity, tape versus price, public pile

Giants–Titans

3

Old identity, tape versus price, quarterback not in the number

Bengals–Steelers

2

Steam, public pile

Seahawks–Commanders

5

Steam, old identity, tape versus price, public pile, quarterback not in the number

Bills–Chargers

4

Steam, old identity, tape versus price, public pile

Lions–Jets

3

Old identity, tape versus price, public pile

Texans–Colts

0

None

Panthers–Browns

3

Old identity, tape versus price, public pile

Jaguars–Patriots

1

The favorite flip

49ers–Cardinals

3

Old identity, tape versus price, public pile

Vikings–Buccaneers

1

The favorite flip

Ravens–Cowboys

1

Old identity, and it is thin

Saints–Raiders

2

Old identity, tape versus price

Rams–Broncos

1

Old identity, and it is thin

Eagles–Bears

5

Steam, old identity, tape versus price, public pile, quarterback not in the number

Sixteen games, 41 live assumptions. Seattle and Philadelphia are the only games where all five are still open. Houston has none. The count is how many ways the price can be wrong. It is not a win probability, and a 5 is not a stronger side than a 1.

Each one is a belief still sitting in the price. Tuesday can make that belief false even if you already knew who was likely to win.

Steam. The line already jumped toward the favorite, often through 3 or 7, and the jump itself may be the mistake. The belief is that the market paid the right amount for a bad loss, a 0–2 start, or a pile of tickets. It breaks if the dog covers and the move looks like an overshoot. It also breaks the other way if the favorite wins by more than the new number, because then the move was the market catching up.

Old identity. A team has already changed the story, with a 2–0 start or a statement win, and the price still describes last year's roster. The belief is that the prior has been updated. It breaks if the team the market will not fully respect covers. It dies before kickoff if that team loses the quarterback or the line finally moves to the new version of them.

Tape versus price. The recent body of work and the closing number disagree on the margin. The belief is that the line has the gap right and two weeks of tape are noise. It breaks if the final margin lands with the tape instead of with the number. It is a false assumption if that tape was a blowout, garbage time, or a backup, and the rest of the season does not look like it.

Public pile. One side has the tickets, and the money often does not. The belief is that the crowd and the price are the same information. It breaks if the crowded side fails to cover. It is not a pile when the handle is on the other side. That is someone disagreeing with the tickets, not the public being right.

Unpriced injury. A starting quarterback is out, and the close has not moved like a new quarterback. The belief is that the absence is already priced in. It breaks if the team that lost the quarterback fails to cover. It goes away if he is cleared, or if the line finally moves the full amount. A skill-player tag is not this.

A favorite flip means the team that opened as the favorite is no longer the favorite. The sign of the game changed, not just the number.

A move from -3 to -7 is steam. The same team was the favorite the whole time. A flip is when the other team takes the favorite's spot. The market did not nudge a price. It changed its mind about who should win.

Two games on this board are flips.

• Jacksonville opened as a short underdog to New England and closed laying 3. The Patriots were the morning favorite. The Jaguars are the favorite now.

• Tampa Bay opened laying 1.5 to Minnesota and closed as a 1.5-point underdog. The favorite changed sides.

That is why each one was counted as a single breakable assumption. The price reflects the belief that the flip was the right correction. It breaks if the team that lost the favorite's spot covers,

Steam is a move of 2 or more points toward the favorite after the open. 

Crossing 3 or 7 matters more than the same move between 4 and 6, because those are the numbers bettors actually pay for. 

A half-point is noise. 

A flip is a different event: the favorite changes sides. A reverse is the line moving toward the dog.

Six games moved at least 2 points toward the side that closed as the favorite.

Game

Open

Close

Move

What the move did

Eagles–Bears

Bears −1.5

Eagles −4.5

6

Flip. Through pick'em and through 3.

Bills–Chargers

Bills −3

Bills −7

4

Off 3, landed on 7.

Seahawks–Commanders

Seahawks −3.5

Seahawks −7.5

4

Through 3 and through 7.

Vikings–Buccaneers

Bucs −1.5

Vikings −1.5

3

Flip. Through pick'em.

Chiefs–Dolphins

Chiefs −7.5

Chiefs −10.5

3

Off 7, through 10.

Bengals–Steelers

Bengals −1.5

Bengals −3.5

2

Through 3, and only just.

The total splits those six into two kinds of steam.

Opinion steam. The side moved, and the total did not collapse. Buffalo went from −3 to −7, and the total rose from 48.5 to 50.5. That is money on the Bills, not a missing player. Kansas City went from −7.5 to −10.5 and the total barely moved. Cincinnati went from −1.5 to −3.5. The total did fall 5 points there, but the spread only crossed 3, and the tickets are piled on the Bengals while the handle is not. That is the cleanest "the crowd moved a short favorite" case on the board. Minnesota's flip is the same family, smaller: 3 points through zero, total down 2.

Injury steam. The side and the total moved together, which is what a new quarterback looks like. Seattle went from −3.5 to −7.5 and the total fell from 46.5 to 40.5, with Daniels out. Philadelphia's 6-point flip off the Bears and a total drop from 46.5 to 41 is the hamstring, not a sudden opinion about either roster. Those numbers may still be short, or they may already be too far. They aren't the same as Buffalo moving to 7 while the total went up.

Three favorites were steamed the other way, by 3 points each.

Game

Open

Close

Move

Packers–Falcons

Packers −7.5

Packers −4.5

3 toward Atlanta. Already settled.

Lions–Jets

Lions −9.5

Lions −6.5

3 toward the Jets. Off 10 and off 7.

49ers–Cardinals

49ers −11.5

49ers −8.5

3 toward Arizona. Off 10, stopped at 8.5.

A reverse is not steam. The market is taking points off a big number. Green Bay's came off 7, and the Falcons won outright. Detroit's and San Francisco's are still open. In both, the retreat is toward the side the recent tape prefers, so the live question is whether 3 points back was enough.

The rest of the board is not steam.

• Carolina −1.5 to −2.5, Baltimore −2.5 to −3.5, Jacksonville −1.5 to −3. A point or a point and a half. Jacksonville is the arguable one: some books opened them at +1.5, which would make it a real flip of 4.5. The number stored here already favored them, so on this sheet it is a small steam onto 3, not a flip.

• Houston did not move off −1.5.

• The Raiders gained a half-point, to the Saints' 3. The Rams lost a half-point, off 3.

• The Giants went from −3 to −2.5 while the total fell from 45.5 to 38.5. That is the tell. The total priced Dart. The spread did not. No steam, and that absence is the assumption.

Long-run, favorites who get steamed 3 or more points have covered worse than a coin flip. That prior is about opinion steam. It does not automatically indict an injury move, and it does not turn Buffalo, Seattle, or Kansas City into dogs, because in those three the tape still wants a bigger margin than the close. There, the steam may be the market catching up. 

Cincinnati is the one where the margin and the tape already agree, so the 2-point cross through 3 has nothing underneath it but the tickets.

The file has two different win probabilities. Neither one is the Premortem Value score.

The Vegas column is only the closing spread, run through the logistic you specified.

Take the absolute spread, add 0.35, and divide by 6.8. That number is the log-odds. Turn it back into a percent:

100 / (1 + e^(−z)), then round.

A pick'em is essentially a coin flip. Each point of spread adds less than the one before it, so a 3-point favorite is about 62 percent and a 10.5-point favorite is about 83 percent, not 50 plus a flat 3 points per point. The favorite is always the side above 50, because the input is the absolute spread. Kansas City −10.5 is (10.5 + 0.35) / 6.8 = 1.60, which is 83 percent. Miami is the other 17.

The PMV column is the tape's margin, not the market's, using the same logistic family.

The margin comes from the blended efficiency already in the workbook. Current DVOA gets 65 percent of the weight, the preseason-blended figure gets 35, so two weird weeks cannot own the number. The gap between the two teams is converted at 3.5 efficiency points per spread point, then home field is added. Home field is 2 points, and it is taken off for Brazil.

That margin is signed for the team in question. It is negative when the tape likes the dog. Log-odds go up by 0.157 for each point of margin, which is the same as dividing the margin by about 6.37, and the same curve turns it into a percent. No extra 0.35 is added. Nothing forces the Vegas favorite to stay above 50.

Detroit is the clean example. The close says the Lions are 6.5 points better. The blended tape says they are about 2.5 points worse than the Jets at home. −2.5 × 0.157 is the log-odds, and that is 40.5 percent for Detroit and 59.5 percent for the Jets. Kansas City's tape margin is about 20 points, and 20 × 0.157 is 96 percent. Same curve, opposite situation.

So the Vegas percent answers what the closing number implies. The PMV percent answers what the blended tape implies. The Premortem Value score is neither. It still counts how much named failure remains in the price.

The tape is the body of work, not the betting line and not the film room.

In this file, it means each team's blended efficiency through Week 2: what the offense and defense have already done, pulled partway back toward the preseason view so two odd games cannot write the whole story. That blend becomes a point margin. Whichever team the margin favors is the team the tape likes.

If that team is the dog, the tape and the price disagree. The line still has the favorite. The results so far have the other side. Detroit is the case. The close says the Lions are 6.5 points better. The tape says the Jets have been about 2.5 points better when they play on the road in Detroit.

What would make the sauce look foolish by Tuesday?

The point of writing the labels down on Friday is that Sunday can kill them without a rewrite.

The identity-lag favorites, Kansas City, San Francisco, and Carolina, are the market's own teams with the model shouting louder. If those teams lose outright, the sauce over-trusted two weeks of tape. Thursday already put one loss on that side of the ledger.

The identity-lag dogs, Jets and Raiders especially, are the other blade. If Detroit and New Orleans win comfortably, the "market hasn't caught up" claim was nostalgia.

The MIXED games can't be claimed by either story after the fact. Buffalo, Seattle, and Philadelphia were marked both-live before the ball was snapped. A win by the favorite does not prove the steam was smart. A win by the dog does not prove the lag was smart. Both stories were on the card.

What would kill the model?

Each term has a falsifier, written before the game:

• Saturday retraces 1.5 or more off a steamed favorite, or a new structural injury hits the steamed side.

• The “disrespected” team loses a quarterback, WR1, or OT1 before kickoff.

• Week 4 opponent-adjusted DVOA reverses the two-game sample.

• Handle flips while tickets stay crowded. That is sharp money, not a public fade.

• Home-field of 2.0 is wrong for Brazil or a short week. Those rows already carry an adjustment.

Tuesday rule, already written down: if the high-PMV attack locations go 1–5 or worse, the index over-weighted two-week DVOA and public piles. Raise the DAVE weight toward 0.50 and cut public to 8. If they go 5–1, do the opposite. Do not rewrite the cards after the fact.