700 CLARK — RESEARCH

Everyone with a better than average third time through the order pitching staff step forward – not so fast, St. Louis

St. Louis starters give back 107 points of OPS the third time through the order, more than three times what the league gives back, and the stat most fans check can’t see it. That is what came back when we rebuilt times-through-the-order numbers one plate appearance at a time, for every start in the majors this season — 3,928 starts across 1,964 games.

The league-wide result cuts in two directions at once. Measured fairly — each pitcher’s third pass against his own second pass in the same games — the third-time penalty is about 30 points of OPS. That is nearly four times the 8-point step the season split shows between those passes. And it is nothing like the wholesale collapse the stat’s reputation suggests.

Matthew Liberatore’s fall from his second time through to his third is the fifth-biggest in baseball, and his steady-looking season line hides it. Among Cardinals starters, Andre Pallante (now on the IL) is the counterweight: 23 third passes in 24 starts, and the same test that flags Liberatore finds nothing wrong with him.

If you take one thing to the group chat, take this: a starter’s times-through-the-order split blends two different kinds of game — the days he earned a long outing and the days he got pulled early. His third-time line comes only from his good (or not terrible) days, while his second-time line mixes both. So the split can make a pitcher who gets worse late in games look steady, and one who gets better late look scary. The question is what a pitcher’s third pass did compared to his own second pass in the same games. Ask it that way and the answer is about 30 points league-wide. That’s more than the raw split shows but far less than a collapse. At the same time a handful of real cliffs, one of them in St. Louis, stand out.

In the last research piece we showed the blend using Cardinals. This one is for all starts in MLB. We look at who genuinely falls off, why the standard split misses them, and who falls off the most.

Start with what the season split blends together. A starter’s second-pass number is an average of two different games. In the starts where he stayed in long enough to face the order a third time, hitters posted a .717 OPS against the league’s second pass. In the starts where he was gone before that, they posted .956. That gap is no 2026 fluke — we measured it at .212 across all of last season, too. Managers pull starters who are getting hit in the middle innings, so the second-pass line you see in a season split is a blend of the good days and the disasters, and the third-pass line only ever comes from the good days.

Now the number at the center. Compare a starter’s third pass to his own second pass in the same games — the only apples-to-apples version of this question — and the league gives back 30 points of OPS; in games that were still close, 21. Read that against the season split’s own step: .739 on the second pass to .747 on the third, a step of just 8 points. The raw split understates the third-pass cost. The cost is nearly four times the raw step — and still far from the wholesale collapse the stat’s reputation suggests. Both corrections are real.

There could be several additional ways to look at these findings. We split the season three ways — by pitch count, by score, and by opponent quality — and asked whether the decline disappears under any of them.

One caution on that control: comparing two passes at the same pitch count quietly swaps one selection problem for another — inside a band, a second-pass plate appearance tends to come from a laboring outing while a third-pass one comes from an efficient one — so read the 67 points as the decline surviving the pitch count, not growing under it.

Pitch count, score and opponent each get their turn, and the decline outlives all three. The third-time-through penalty is real.

We need to add a word about how long starters actually last, because starters are being pulled earlier than they once were. That shift is well documented, and we will put our own numbers on it in a coming piece. In 2026 starters have reached a third time through in 84 percent of starts. The median start is 23 batters faced — about two and a half turns. Whatever has happened to the length of a start, the third time through has not become rare.

Two units are getting mixed together here. “Going deep” usually means innings, and this measures turns through the order. A pitcher who retires 18 in six clean innings never faces the order a third time, while one who labors through 27 in five does.

Managers also pull starters partly because they believe in the third-time penalty. Everything we just measured comes from the starts they chose to extend. Hook earlier, and the third passes that survive come only from the best days. That would make the penalty look smaller than it is. It cuts against our own number: if hooks keep coming earlier, the true penalty is probably somewhat bigger than the 30 points we measured.

How to read the small samples here. We are using a single season of plate appearances. Across all pitchers for a season, that is statistically significant. When it narrows to a single pitcher and a handful of starts that end early, the sample sizes get small. We will treat actual outcomes however small as the truth of what happened. What a small sample can’t tell you on its own is whether a number is a pattern or a hot stretch because a starter’s third-pass line has about 100 plate appearances.

So every within-game change we print carries confidence bounds: the range the number could reasonably be, given the starts that produced it. The bounds are plain arithmetic on his actual starts. (The math is on the methods page below.) The league’s 30-point figure is built on 3,285 starts, so its bounds are tight: 9 to 51 points. One pitcher’s figure is built on 20-odd starts, so his bounds are wide — often 150 points in each direction. When a pitcher’s whole range sits on one side of the league figure, the difference is statistically significant. The nine pitchers named next are all significant.

This makes the exceptions worth naming. Nine starters this season decline by so much more than the league that their whole bounds range sits above the league’s 30 points. Their fall is statistically significant, not just large. Keep in mind this measures how much a pitcher changes from his second pass to his third — not how well batters do against him overall. In fact, two of the nine, Christian Scott and Dylan Cease, still hold batters below the league average on the third pass. They just fell further than anyone else to get there. The leaders are Cincinnati’s Nick Lodolo, the Athletics’ Jacob Lopez, Seattle’s Bryce Miller and Cleveland’s Slade Cecconi. Then, fifth-worst in baseball, the Cardinals’ Matthew Liberatore. His second pass in survival starts is a very good .648. From the third pass on in those same games it is 1.091. Tarik Skubal, Taj Bradley, Christian Scott and Dylan Cease round out the nine. Some big falls miss the list: Shota Imanaga (+.316), Trey Yesavage (+.404) and Paul Skenes (+.266) all fell hard, but their seasons swing too much from start to start, so their bounds cross the league line.

And here is why the standard split can’t find them. The standard card shows OPS allowed the first, second and third time through. Thirteen starters have cards where all three numbers look at or better than league average — and this analysis found them. Nine of the thirteen get caught because of the step between passes. Batters gain far more on them from the second pass to the third than they gain on the league. Dylan Cease is the clearest. His line reads .489, .465, .693, against a league of .697, .739, .747 — every pass better than average. But looking at the step instead of the levels we see from the second time through to the third, hitters gain 229 points on him.

And unlike Liberatore, his stats are not a blend of games without a third pass. Cease reached a third pass in all 24 of his starts, so his season line and his survival-start line are the same numbers. Cam Schlittler, Chris Sale, Reid Detmers and Gavin Williams show smaller versions of the same shape — clean-looking lines, a bigger step inside them — though none of the four is statistically significant.

The mirror case is still Brady Singer, whose splits we picked apart last time. In the games he pitched into a third pass, hitters had him at .911 on the second pass and .714 from the third on — better late by 196 points of OPS. His bounds run from −.448 to +.056. The top of that range sits past the league’s +.030, so the improvement just misses on statistical significance. Read it as a strong lean, not a settled pattern. Either way, the misleading number in his season split is the ugly second-pass one, not the third-pass one.

Why there’s no list of the pitchers who hold up best. We tried to build one. That list is what we set out to write. Did we find such pitchers? Yes — our drafted list had nineteen names on it. But we could not trust the list enough to print it. Here is the problem. Take the Astros’ Peter Lambert, the middle name on that list. He cleared his qualifying bar by 90 points of OPS. His margin of error was 246 points — nearly three times his edge. Most of the nineteen look like that; only two of them beat their own margin of error at all. When the error is that much bigger than the edge, the names are mostly noise: run the same season again and you would get a different list. The nine collapses above are different. Their falls clear the league by hundreds of points, far more than their error. That is why we can name them and not the others. Making the bar stricter made the list worse, not better. We would rather give you no list than one we can’t stand behind.

The St. Louis staff, in detail. St. Louis starters reach the third time through about as often as anyone — 115 of 132 starts, against a league rate of 84 percent. The staff’s second pass in survival starts is .692 and its third pass is .799, a slide of 107 points against the league’s 30. But a staff’s third-pass season is still only a few hundred plate appearances, so the bounds on that 107 are wide: +.001 to +.213. The low end dips under the league’s 30, so we stop short of calling the staff slide statistically significant.

Liberatore is most of it. Take his starts out and the rest of the staff slides just 40 points, with bounds of −.076 to +.156 — statistically no different from the league. His own fall clears the bar comfortably. Andre Pallante is the counterweight. He has reached a third pass in 23 of 24 starts, held hitters to .613 on the second pass in those games and .684 from the third on — a change of +.071 with bounds of −.116 to +.258, indistinguishable from the league. Below is the same unblended view for every starter on the staff.

What changes in 700 Clark. First, when we quote a third-pass number we’ll keep showing you how many starts produced it and how the second pass looked in those same games — the practice we recently started. Second, and new this week: when we print a within-game change, we’ll say whether it is statistically significant, right where the number appears. The number always prints either way; the label tells you how hard to lean on it. Liberatore’s is significant; Singer’s just misses, and the piece says so in the same sentence.

HOW WE MEASURED IT

A companion to the piece above. Everything the piece claims is derived here, from raw counting stats, with the arithmetic shown. If you want to check us, this is the page to check.

What counts

Times through the order: for a batter facing a starting pitcher, his number on any given plate appearance is how many times he has now completed a plate appearance against that starter in that game. Appearances against relievers don’t advance the count.

A completed plate appearance is one that ends with a batter outcome: a hit, an out made at bat, a strikeout, a walk, a hit batsman, a sacrifice fly or bunt, or catcher’s interference. Plays that retire a runner while the batter stays at the plate are excluded — caught stealing, pickoffs, wild pitches, and the plays the data files label “Runner Out.” Across the 2026 season that exclusion covers 218 plays, against 84,776 completed plate appearances by starters.

A start belongs to the pitcher who throws the first play of his team’s defensive half. Starts facing fewer than nine batters are treated as openers; there were 176 league-wide.

A survival start is one where the starter faced some batter for a third time. Third and fourth passes are pooled throughout, since fourth passes are rare — 349 plate appearances in the whole league.

The formulas

For any group of plate appearances, add up the counting stats, then:

OBP = (hits + walks + hit-by-pitch) / (at-bats + walks + hit-by-pitch + sacrifice flies)
SLG = total bases / at-bats
OPS = OBP + SLG

Sacrifice bunts stay out of the on-base denominator, which is the standard convention.

Three derived measures carry the piece:

attainment      = starts reaching a third pass / starts with at least one completed PA

conditional gap = OPS on the 2nd pass in starts he did NOT survive
                  minus OPS on the 2nd pass in starts he did

within-game     = OPS on the 3rd pass and later, minus OPS on the 2nd pass,
change            both measured inside survival starts only

The last one is the piece’s central statistic. Both of its terms come from the same games, which is what makes it a fair comparison.

One floor governs everything. Any group under 50 plate appearances is reported as a count, never as a rate. A within-game change is withheld when either side falls under the floor, because the difference of two figures we won’t quote is itself not quotable.

Worked example: one start

Matthew Liberatore, March 26 against Tampa Bay, 24 batters faced. Walking the plate appearances in order and counting each batter’s meetings with him:

Pass	PA	AB	H	TB	BB
1st time through	9	8	3	3	1
2nd time through	9	8	1	1	1
3rd time through	6	6	3	6	0

Nine plate appearances is one turn through a nine-man order. The 24th batter he faced was the sixth man seeing him a third time, so this start reached a third pass and joins the survival group. The damage that night landed on the third pass, but one start is an anecdote; the season totals below are the evidence.

Worked example: a full season

Liberatore made 25 starts and survived into a third pass in 22 of them. Pooling the plate appearances inside those 22:

Second pass — 198 PA, 181 AB, 46 H, 64 TB, 9 BB, 3 HBP, 4 SF

OBP = (46 + 9 + 3) / (181 + 9 + 3 + 4) = 58 / 197 = .294416
SLG = 64 / 181                                     = .353591
OPS                                                = .648008   ->  .648

Third pass and later, same 22 starts — 96 PA, 79 AB, 28 H, 50 TB, 16 BB, 0 HBP, 1 SF

OBP = (28 + 16 + 0) / (79 + 16 + 0 + 1) = 44 / 96 = .458333
SLG = 50 / 79                                     = .632911
OPS                                               = 1.091245  -> 1.091

within-game change = 1.091245 - .648008 = +.443237  -> +.443

Both groups clear the 50-PA floor, so both are quotable and so is the difference.

His season split says .757 on the second pass instead of .648. The gap is the three starts he didn’t finish: 22 plate appearances from early hooks, averaged in with the 198 from the games he survived. Blending is what hides the size of the fall.

The mirror case: when nothing is blended

Dylan Cease is worth stating carefully, because it is easy to get backwards. He reached a third pass in all 24 of his starts. He has no early-hook plate appearances at all. So for him:

season split, 2nd pass  = .4647 (225 PA)     2nd pass in survival starts = .4647 (225 PA)
season split, 3rd on    = .6934 (143 PA)     3rd on in survival starts   = .6934 (143 PA)

The two views are the same numbers, necessarily. Presenting the conditional figures as though they uncovered something his season line concealed would be wrong, because there is nothing concealed.

A different property of the same display is what catches him. His three passes read .489, .465 and .693 against a league of .697, .739 and .747 — every level better than average. The step between passes is the abnormal part:

Cease       .6934 - .4647 = +.2287   ->  229 points
the league  .7474 - .7175 = +.0299   ->   30 points

So the standard three-way split can fail a reader two ways. It can blend different kinds of game together, as it does with Liberatore, which conditioning fixes. Or it can be entirely accurate and still invite a comparison against the league rather than against the pitcher’s own earlier innings, as it does with Cease, which only reading the step fixes.

Worked example: the league

All 3,928 starts, 3,285 of which reached a third pass.

Second pass, survival starts — 29,543 PA, 26,586 AB, 6,563 H, 10,757 TB, 2,326 BB, 323 HBP, 208 SF

OBP = (6563 + 2326 + 323) / (26586 + 2326 + 323 + 208) = 9212 / 29443 = .312876
SLG = 10757 / 26586                                                    = .404611
OPS                                                                    = .717487  -> .717

Second pass, starts he was pulled from — 2,966 PA, 2,548 AB, 812 H, 1,408 TB, 340 BB, 43 HBP, 30 SF

OBP = (812 + 340 + 43) / (2548 + 340 + 43 + 30) = 1195 / 2961 = .403580
SLG = 1408 / 2548                                              = .552590
OPS                                                            = .956170  -> .956

conditional gap = .956170 - .717487 = .238683   ->  239 points

Third pass and later, survival starts — 17,402 PA, 15,614 AB, 3,968 H, 6,628 TB, 1,491 BB, 151 HBP, 116 SF

OBP = (3968 + 1491 + 151) / (15614 + 1491 + 151 + 116) = 5610 / 17372 = .322933
SLG = 6628 / 15614                                                     = .424491
OPS                                                                    = .747424 -> .747

within-game change = .747424 - .717487 = +.029937   ->  30 points

The two figures the piece contrasts come from the same 3,285 starts: 239 points of difference between kinds of game, and 30 points of decline within them.

The confidence bounds, and when a difference gets called significant

A within-game change is the difference of two rates, each built on roughly 100 to 200 plate appearances, so it carries both of their sampling errors. Before printing one, we put bounds on it: the range the number could reasonably be, given the season that produced it.

The bounds come from the pitcher’s own starts, with no simulation and no random numbers. We recompute his pooled within-game change once for each of his starts, each time leaving that one start out. How much those recomputations scatter tells us how much any single start swings his season figure. A standard formula — the jackknife — turns that scatter into a standard error. The start is the unit, not the plate appearance. A manager’s hook decision operates on whole starts. And the plate appearances inside one start share the same night: the same opponent, the same stuff, the same score.

for each of his n starts:
    recompute the within-game change with that start left out

SE     = the jackknife standard error of those n values
bounds = change ± 1.645 × SE          (a 90% margin of error)

a change is SIGNIFICANT when its entire bounds range sits on one side
of the league figure; otherwise it prints with its sample and no label

The test runs in both directions, since a pitcher who is clearly better late is as worth printing as one who collapses.

Pitcher	within-game change	bounds	significant
Matthew Liberatore	+.443	+.207 to +.680	yes
Dylan Cease	+.229	+.068 to +.390	yes
Shota Imanaga	+.316	+.014 to +.619	no — the lower end crosses the league line
Brady Singer	−.196	−.448 to +.056	no — the upper end crosses
Robbie Ray	−.000	−.209 to +.209	no

Imanaga and Singer are the instructive rows. Both changes are large. But both seasons swing hard from start to start, so both ranges touch the other side of the league line, and neither gets the label. Of the 134 starters with enough work to qualify, seventeen are distinguishable from the league: the nine collapses named in the piece, and eight starters whose third pass is significantly better. The other 117 fall short.

The same construction applies to pooled groups. The league’s own +.030 is built on 3,285 surviving starts, so its bounds are tight: +.009 to +.051. The Cardinals staff’s +.107 carries bounds of +.001 to +.213. That range is wide because even a staff’s third-pass season is only a few hundred plate appearances. It crosses the league line, so the staff-level slide is reported without the significant label. Pooling the staff’s 107 non-Liberatore starts the same way gives +.040, bounds −.076 to +.156 — the “without him” figure the piece quotes.

The three controls

Each control recomputes the same within-game change inside a subset of plate appearances. The 50-PA floor applies within every band, so thin bands report counts and no rate.

Control	Band	2nd pass	3rd on	change
Pitches thrown entering the PA	50–74	.691 (14,735 PA)	.758 (6,538 PA)	+.067
Score, from the pitcher’s side	behind by 4+	.686 (1,384)	.802 (1,676)	+.116
Score, from the pitcher’s side	within 3 runs	.728 (25,105)	.749 (12,757)	+.021
Opponent, by their own OPS vs starters	weakest third	.691 (9,931)	.719 (5,846)	+.028
Opponent, by their own OPS vs starters	middle third	.718 (9,748)	.748 (5,753)	+.030
Opponent, by their own OPS vs starters	strongest third	.744 (9,864)	.776 (5,803)	+.032

Reading the pitch-count rows correctly matters, and it cuts against us. Holding the pitch count roughly fixed trades one selection problem for another: inside a band, a second-pass plate appearance tends to come from a laboring outing while a third-pass one comes from an efficient outing. The band shows that the decline persists under the control. It does not separate batter familiarity from fatigue, and no version of this design can, because the two move together by construction.

What this does not settle

The within-game change records what happened. It does not explain why, and familiarity, fatigue and pitch count cannot be pulled apart here.

It also stops short of certifying anyone as good. A pass-or-fail screen at these sample sizes is unstable: in the screen we drafted and rejected, the median member (Peter Lambert: edge 90 points, margin of error 246) carried an error nearly three times his edge, and only two of the nineteen members (Logan Webb and Michael Soroka) beat their own margin at all. The piece names collapses, which sit far from any threshold, and ranks nobody as reliable.

There is one bias we should name before anyone else does, because it runs through the central statistic. Managers pull starters partly because they believe in the third-time penalty, and the within-game change is computed only over the starts they chose to extend. Pull earlier and the third passes that still happen come from the best days only, which would make the decline look smaller than it is. So the figure partly reflects how a league uses its starters, not only how starters pitch. Two weak signals point that way — across the 30 teams, those who let starters go deeper show a slightly larger decline, and from 2025 to 2026 attainment and the decline fell together — but neither is close to strong enough to lean on, and untangling it needs a multi-season series we have not built yet. It is the next thing we are looking at.

Year-over-year comparisons need two separate margins kept apart, and it is easy to quote the wrong one. We checked our 2026 data against an independent source for the 2025 season and the two agreed to about a thousandth, so a difference between seasons is not an artifact of which source we used. That is a statement about the instrument. The binding limit is the sample: the jackknife puts the league’s own standard error near thirteen thousandths, roughly twenty times coarser. The 2025-to-2026 move of nineteen thousandths is therefore under one and a half standard errors — a difference we can describe, not a change we can claim.

And the season is partial: 1,964 games of roughly 2,430, weighted toward April.