The Crosstown, Decided — In Extras, In Style

Rank two.

That's where Davis Martin's posterior pitcher-quality sits as of last night's training run. Out of 768 MLB starters with measurable signal in predictions.pitcher_strength, his quality_mean reads +0.118 with an 80% credible interval of [+0.016, +0.223]. Three weeks ago that interval crossed zero. After Saturday it doesn't — the lo-bound has crawled above the line. That is the Bayesian diagnostic for "true talent the model is now confident about", and it is doing the heaviest lifting in the Sox's re-rated playoff posterior.

Before that, though, a moment for the Sunday tenth.

Eight, six, ten, eight

Three games at Rate Field, eight pitchers used between them on Sunday alone, and one of the few late-May games this season that genuinely deserved the Crosstown Classic branding the league insists on slapping on it.

Sunday was the soft middle of the rotation matchup. Fedde's posterior quality_mean is -0.044 (rank 743 of 768). Rea's is -0.063 (rank 763). Two pitchers the model treats as bottom-third starters; an afternoon at Rate with the wind blowing out; a 9-8 final after ten innings. That is, mechanically, a bullpen game. Eight Cubs pitchers threw. Seven Sox pitchers threw. The Sox bullpen — the comically combustible unit that was −0.25 log-runs below league a year ago — held a one-run advantage through the ninth, surrendered a tying RBI single in the top of the tenth that should have ended the series 1-2 in Cubs' favour, and then walked the game off in the bottom of the same inning. Eight pitchers each. The Crosstown Classic in its proper form.

What the posterior actually moved on

Saturday's update is the structural one. The mechanics of the Bayesian update on Martin's quality_mean need explaining because the shape of the move is the diagnostic. Before Saturday: posterior mean +0.112, 80% CI [-0.018, +0.241] — wide, straddling zero, "promising but unconfirmed". After Saturday: mean +0.118, CI [+0.016, +0.223]. The mean moved up by 6 mil-runs; the CI narrowed by 28 mil-runs while staying centred at the same place. That asymmetric move — mean stable, interval tightening from below — is the Bayesian fingerprint of "the data we just saw was exactly what the model had been guessing about him, so we trust the guess more now". It isn't surprise. It's confirmation.

That single-pitcher update is doing more than it looks. It flows into the rotation aggregate (Sox starter quality moves by +0.012 on a five-arm weighted average), which feeds the expected runs-against per game, which feeds 114 remaining game-WPs in predictions.game_predictions, which compounds in the predictions.standings_sim Monte Carlo. Posterior playoff probability for the Sox now sits at 25.0% — up from 21.4% a week ago and 6.0% on opening day. The point estimate of expected wins is 78.0, 80% CI [71, 85]. The high-bound — 85 wins — is now interesting in a way it wasn't two weeks ago.

Standings posterior — and the gap that matters

Cleveland is the AL Central. The Guardians' division-winner probability of 82.6% says the model doesn't really treat the race as a race. The interesting question for the Sox is not whether they can catch Cleveland over the remaining 114 games — the answer is "8.1% probably not" — but whether their projected-win posterior holds above the wild-card line.

The Cubs trajectory is the part of this graph that did not exist three weeks ago. Their β_form coefficient is hierarchical — prior-tight, so any single ten-game stretch is shrunk hard toward the season-strength prior — but a -23 L10 differential is far enough out of the prior's high-density region that the posterior is starting to take the form term seriously. They have not moved in rank_overall yet (still 5th league-wide at +0.076 log-runs), but their playoff probability has dropped from a season-high 88.4% to 75.4% in two weeks. That is a 13-point move on a team the model is fundamentally bullish on. It is also the largest two-week negative move of any first-tier contender in either league.

The mechanism: where the runs came from

The Sox scored 22 across the three games. Their pre-series per-game scoring forecast was 4.2. The residual — 22 - 12.6 = +9.4 runs above expected — is large enough to merit a closer look at whether the offence is a regime shift or a hot streak. The decomposition reads as roughly half-and-half.

The signal components — left-side pull-rate gain on the season-to-date EV profile, the manager's platoon-rolling production gain estimated by predictions.lineup_optimizer, and an aggregate "Cubs bullpen disadvantage" term derived from the opposing-bullpen quality stack — collectively account for about 5.2 of the 9.4 extra runs. The variance components — BABIP-against drift on Cubs relievers, and unmodelled residual — account for the other 4.2. The model treats the signal half as forward-projectable. The variance half it does not.

Bullpen, finally

The thing the 2023 and 2024 Sox could not do — protect a late lead on the road or hold one through Sunday afternoon wind-out conditions at Rate — they did. Five relievers across the three games, no blown leads, the Sunday-tenth save preceded by exactly one questionable batter-faced sequence. Each appearance is a thin slice of evidence; the joint per-reliever posterior shift is the thing to watch. Top-stack quality estimate (the weighted average of the Sox's three highest-leverage relievers in predictions.pitcher_strength) has moved as follows:

Twenty-seven log-run units, climbing. The Sox bullpen aggregate is now league average by the model's reckoning. League-average relief from a unit that was a textbook tire fire twelve months ago is the entire mechanical difference between a 76-win team and an 84-win team. It is more important to the playoff posterior than the Martin update is.

Is it sustainable?

Two parts of the story project durably. Two don't.

Durable: Martin's posterior is now underwritten by his zone-location profile in mlb_gold.feat_pitcher_zone, which has been stable across his last six starts — glove-side edge location heavier than the average right-hander, slider whiff/swing above his cohort. The model treats the pitch-design strategy as the asset; Saturday's result was the confirmation. The offensive shape — left-side pull-rate gain backed by EV and barrel-rate, both of which decay slowly — is similarly durable. These are first-half traits, not three-game streaks.

Not durable: The bullpen's recent stretch sits at a depressed babip-against. The prior on the babip random effect is tight (σ = 0.018), and the posterior on Sox relievers is currently at the lower edge of that prior. Expect a ~12-point babip drift toward league mean over the next 30 games. That alone unwinds about half of the bullpen aggregate gain above. The Crosstown's unmodelled-residual component — ~2.2 runs over three games — is variance the model knows it can't keep claiming.

Wild card or bust

25% playoff probability is not a "they might do it" number. It's a "the path is open" number. The model's wild-card cluster — six AL teams with final-record posteriors inside a three-to-four win band of each other — looks like this right now:

Boston is the team to beat for the third wild card. Their run-differential underperformance flags them as a Pythagorean regression candidate; the model is already pricing some of that into the 39.7%. The Twins and Tigers are the genuinely close ones the Sox have to fight directly for the spot. Houston is mathematically still in the cluster but the model has effectively written them off at 1.4%.

The next two weeks — quality opposition

Here is where the Sox schedule does them no favours. The Cubs, meanwhile, get exactly the team they need.

The Sox open with three road series against quality opposition. Wednesday's series-opener at Seattle (40.6% WP) is the lowest single-game number on the next-12 calendar. Then three at Oracle against a Giants team (rank 18, neutral L10) that has the rotation to make this an underdog series for the Sox. The two home series — Minnesota and Detroit at Rate — bracket the upside. If the Sox return from California at 2-2 in those six road games, the home stand against the Tigers and Twins becomes a chance to add wins quickly against teams the model sees as genuinely worse.

Compare that with what the Cubs see: a quick home set against Milwaukee, then three at home against Houston — a 26th-ranked Astros team coming off a -24 L10 stretch of their own. The model has the Cubs as ~62% favourites in each of those three games. Three wins out of three is the median outcome the form-cooled Cubs need to arrest their own L10 collapse. Pittsburgh and St Louis road sets follow; both are coin-flip series.

The asymmetry of the next-two-weeks calendar — quality opposition for the Sox, exactly-the-team-you-need for the Cubs — is the bigger meta-story than either club's actual win counts will be by the time we get to it. If both teams finish the fortnight at .500, the model treats it as confirmation that the re-rating from the Crosstown was roughly correctly sized. If the Sox come out below .500 and the Cubs come out above, the model gives back some of the weekend's posterior shift. If the Sox come out above .500 against this slate, the 25% playoff probability becomes very interesting indeed.

What to watch, in order

1. Wednesday's WP at Seattle. The 40.6% number assumes both teams reach the mound at posterior baselines. Lower if the Sox start a back-half rotation arm on short rest after Sunday's bullpen-heavy finish; higher if Martin's next outing is moved up by a day.

2. Sox bullpen babip drift. The single clean diagnostic for whether the new identity is for real. If babip-against climbs back toward league over the next 30 games, the per-reliever posteriors will give back ~half the gain on the chart above. If it stays depressed, the new identity earns the projection.

3. Martin's next start. The cheapest way for the weekend's interval-tightening gain to evaporate is one bad outing against a real lineup. If he holds the K-rate and zone-edge command, the rotation aggregate stays at its new level.

4. Cubs form arrest against Houston. A -23 L10 doesn't stay -23 unless something is structurally wrong. Three games at home against a struggling Astros side is exactly the kind of stretch a 5th-ranked team uses to course-correct. If it doesn't — if the Cubs go 1-2 or worse against Houston — the model's downgrade of their playoff posterior accelerates.

The Saturday game, again

Every spring two or three teams have their projection re-rate inside the first two months on the basis of a recognisable identity flip. The 2023 Orioles. The 2021 Giants. The 2019 Twins. These aren't teams that outperform expectations by hitting the same way as last year and getting lucky; they're teams whose underlying production profile changes shape — different EV, different platoon usage, different relief deployment — in ways the data sees before the standings do.

The 2026 Sox aren't yet that team. 25% playoff probability says it explicitly. But the shape of the change is the right shape, and the Saturday update — the asymmetric narrowing of Martin's interval — is the kind of move that only happens when the model is being shown what it had already been guessing. Whether any of this holds into July is the slowest variable in the system. Sunday's tenth- inning finish is the most fun this team has been in three years. Wednesday at T-Mobile Park is when we find out whether they hold the rating.

Methodology: per-game probabilities and projected wins from predictions.standings_sim (10,000-iteration Monte Carlo on the per-game posterior chain, refit nightly). Pitcher quality means and 80%-CI intervals from predictions.pitcher_strength; team-strength components from predictions.team_strength; form-L10 from the same. Series scores from mlb_gold.fct_games. The residual decomposition above is a back-of-envelope attribution that combines the lineup-optimisation production estimate from predictions.lineup_optimizer with a closed-form Bullpen Disadvantage Index derived from per-stack quality differentials; the BABIP variance component is the random-effect prior on opposing-bullpen babip-against, integrated over three games. Posterior intervals are 80% credible.