Finding soft pitching matchups isn’t guesswork for me—it’s pattern-spotting. I blend Statcast quality-of-contact trends, pitch mix changes, and bullpen context with AI models to flag when arms are likely to wobble. Then I sanity-check with weather, park factors, and lineup news. The goal: clear, actionable reads you can trust, without noise or hype.
Table Of Contents
- What AI finds when flagging weak pitching matchups
- Data building and labels, fast and careful
- Modeling, validation and interpretability that analysts trust
- Putting it to work today plus safeguards
- Conclusion
- Frequently Asked Questions (FAQs)
What AI finds when flagging weak pitching matchups
Quality of contact that predicts real runs
The fastest way to spot a pitcher in trouble is to watch the batted-ball quality he is giving up, not just ERA. Statcast’s expected weighted on-base metrics anchor this concept beautifully.
We look closely at xwOBA and xwOBACON. These essential metrics summarize the overall quality of contact based on exit velocity and launch angle, completely stripped from defense. While xwOBA includes walks, hit-by-pitches, and strikeouts, xwOBACON focuses solely on balls in play. If a starter’s rolling 3 to 5 game xwOBACON is spiking compared to his 1 to 2 year baseline, that is a massive red flag even if his recent ERAs look perfectly clean on the surface.
We also monitor Hard-hit% and barrel%. A starting pitcher can easily survive loud, hard contact for a week or two based on pure luck. However, when barrels are up and he is consistently losing weak contact in crucial two-strike counts, regression bites incredibly fast. Treat barrels as an accurate pressure gauge on the pitcher’s command and deception.
We love these specific metrics because xwOBA stabilizes quicker than ERA and reacts to small changes in command or sharpness. It also maps directly to expected runs, which is the exact label we ultimately need. For a primary source, you can use the official MLB Statcast leaderboards to analyze real-time data. Pull pitch-level data to compute rolling and season-to-date x metrics.
Plate discipline and count control
Stuff matters, but where a pitcher lives in counts often decides the night. Whiff rate and chase% tell a massive story. Falling whiffs or chases say hitters are seeing the ball better or refusing the chase pitch entirely. If a pitcher loses chase on his best out pitch, that is a weak matchup waiting to happen.
First-pitch strike% and zone% are just as critical because first-pitch strikes set up everything. A dip in F-Strike% increases walk probability, makes fastballs too predictable, and leads to hitter-friendly counts. Meanwhile, Zone% shows if he is nibbling or forced to catch too much of the plate.
Called-strike plus whiff rate (CSW%) is a clean, pitcher-neutral summary of dominance. A 2 to 3 point CSW drop over a few starts, especially when facing patient lineups, is highly meaningful. These elements predict traffic, pitch predictability, and fatigue inside an outing. We fold them into rolling features so the model feels recent command changes.
Platoon splits you can trust
Righty/lefty splits are notoriously noisy month-to-month. To fix this, we smooth them out using career and last 2 years L vs R xwOBA allowed, applying exponential weighting toward the current season.
We also distinguish pitch types. A righty with a great slider might absolutely smash right-handed batters but struggle against left-handed batters if he cannot back-foot it. If tonight’s opponent stacks lefties who hit sliders, the split amplifies. Always compute the pitcher vs the left/right distribution of the projected nine rather than relying on league-average platoon percentages. You can cross-reference these breakdowns by checking out the latest MLB team news to see exactly how managers are stacking their daily lineups.
Pitch-type effectiveness vs hitter strengths
A big piece of ATSwins’ matchup model is pitch-shape and usage vs what the hitters crush. We map the pitcher’s usage and quality by pitch, pulling pitch-level xwOBA and whiff% for four-seamers, sinkers, sliders, curves, changeups, and cutters.
For each opposing hitter, we compute strengths, such as xwOBA vs sliders 84 to 87 mph glove-side, or vs high four-seams above the belt. Even a simple proxy catches a lot. We then simulate pitch-mix exposure. If a pitcher relies 35% on sliders and the lineup holds a collective top-quartile xwOBA vs sliders, his baseline run expectation rises. This is where many hand-cappers miss. Pitch-type matching tells you whether a big-name starter is actually in a bad stylistic spot tonight.
Recent form with rolling windows that aren’t too twitchy
Hot and cold talk often overreacts to variance, so we like rolling windows with memory. We track 3-game and 5-game rolling xwOBA allowed, hard-hit%, whiff%, and F-Strike%.
We employ exponentially weighted means where the last start gets the most weight, but the last 8 to 10 starts are not ignored. Velocity deltas are also vital. A 1 to 2 mph drop relative to a 30-day baseline is rarely nothing. Tie velocity dips to rising xwOBA and lower chase, and the risk compiles. This blend keeps the model responsive to short-term health and command shifts without throwing out true talent.
Park factors and weather that move lines
You need to normalize a pitcher’s recent contact to the parks he was in and the one he’s going to. Park factors include home runs, doubles, triples, the overall run environment, and handedness splits. Some parks boost HR to the pull side while others kill carry to the opposite field.
Weather features air density, wind angle, wind speed, and temperature. Wind blowing out to center at 12+ mph and 80°F boosts carry significantly. Conversely, 55°F with dense air dampens flight. Weather turns a mediocre matchup into a weak one fast, especially for fly-ball pitchers. Model these as additive or multiplicative adjustments on expected run conversion of contact quality. Keep it simple at first. A signed wind-to-center variable and a temperature variable catch a lot.
Bullpen fatigue and leverage behind the starter
Even a decent starter turns weak if the pen is torched. We look at days of rest for top 3 or 4 relievers, back-to-back usage, and pitches thrown over the last 3 days.
We also track the team leverage index over the last 7 days to estimate stress, since high leverage innings cost more recovery. We assess skill tiers for the pen by checking expected xwOBA allowed for the first two waves behind the starter. If the second wave is league-worst and overworked, late runs spike. We feed a bullpen availability index into the model so the starter’s leash and support are realistic.
Umpire zones and strike expansion
Some umpires consistently shrink or expand the zone on the edges. A tight low-and-away zone hurts sinker/slider types, whereas a wide inside corner favors command-first righties.
Add an umpire zone profile if announced early enough. A few public resources summarize expansion tendencies and accuracy. If there is no confirmation by midday, use team-level expected zone averages and keep an uncertainty buffer. The point isn’t to outsmart the market on every plate assignment. It is to boost confidence when an extreme zone aligns against a pitcher’s plan.
Baselines and context we normalize to
Weak matchups are relative. They must be evaluated against the pitcher’s long-run true talent, which means looking at 1 to 2 year xwOBA allowed.
They must also be weighed against the opposing lineup, not league average. Compute a projected lineup xwOBA vs handedness with park and weather adjustments. Finally, consider the league run environment this month, as MLB run-scoring drifts within seasons. Re-benchmark every two weeks so your flags aren’t stale. We use these baselines to avoid labeling a mid-rotation arm weak simply because he’s pitching in a hitter-friendly environment. Normalize first, flag after.
Data building and labels, fast and careful
Windows that carry enough history without stale bias
Use the last 1 to 2 MLB seasons of Statcast pitch-level data, plus the current season to date. Engineer rolling features that look back 14, 30, and 60 days with exponential decay favoring recency. Clip or winsorize outliers on EV and launch angle to protect the model from noisy tracking. This set balances history for true talent and recency for tonight’s form.
Labels that match betting decisions
Define weak in ways that map to real actions. For a regression label, look at expected runs allowed over the first 5 innings and the full game. You can compute expected runs via a run-conversion model on xwOBA and contact mix. For a classification label, separate weak vs not-weak. Threshold examples include the top 20% highest expected runs allowed among projected starters that day, or an xwOBA allowed percentile above the 80th after park and weather normalization. Alternative labels include the probability of allowing 3+ ER in the first 5 innings or the probability of over 1.5 HR allowed. Having both regression and classification lets us power multiple bet types, from F5 totals to pitcher HR props.
Feature engineering that pays the bills
Start with what is predictive and measurable. Track times-through-the-order (TTO) splits, evaluating performance by the first, second, and third time through, plus team-level hook tendencies to gauge managerial leash.
Use pitch location heatmaps to summarize by quadrant and edge percentage, measuring the drift from the baseline heatmap using simple distance metrics. Track velocity and movement deltas per pitch type vs 30-day and 1-year baselines. Monitor rest days and travel, noting days since the last start, consecutive road games, and time-zone changes. Incorporate catcher framing and game-calling proxies, using team receiving metrics and pitcher-specific strike-gained estimates if available. Finally, monitor schedule density, bullpen readiness, and injury flags like IL stints, recent exits with discomfort, or downscaled pitch counts. Engineer these at the per-start level so the model can reason about tonight, not just season aggregates.
Build matchup-level rows: pitcher x projected lineup
For each scheduled game and probable starter, join a projected batting order using team news feeds and lineup projection services. Aggregate hitters’ strengths against the pitcher’s pitch mix and handedness.
Compute the expected contact profile, looking at the share of PA ending as fly balls vs grounders, or the share that see 2-strike sliders. A simplified approach still helps greatly, such as lineup xwOBA vs pitch types weighted by pitcher usage. The result is a row per pitcher-game with both pitcher and opponent features tuned to tonight’s context.
De-noise with exponential weighting so small samples don’t bite
Use exponential moving averages for rolling metrics with half-lives of 7 to 21 days, depending on stability. Velocity stabilizes quickly, while platoon splits take much slower.
Blend career and recent performance with a shrinkage estimator. Recent form gets more weight when the sample size is decent, but it never gets 100% early in April. Cap per-start feature impact to avoid one outlier start overwhelming the signal. This prevents the classic small-sample trap while staying responsive.
Prevent leakage with date and series-aware splits
Information from the future must not leak into training. Split train, validation, and test data by calendar date with a rolling origin, never using a random shuffle.
Keep an entire series on the same side of the split, because lineups and bullpen states carry over heavily within series. Use only pregame information available by your timestamp cutoff. Hold out final lineup confirmations for a second-stage update if you support intraday refresh. Strong leakage discipline preserves real-world performance.
Modeling, validation and interpretability that analysts trust
Frame the problem two ways
Classification focuses on weak vs not-weak. Output a probability so we can rank flags and set thresholds by expected value, not vibes. Regression targets expected runs allowed for F5 and the full game, alongside expected HR allowed. This enables totals, team totals, and player props. Running both increases the surface area for bets and lets the models sanity-check one another.
Models that perform and calibrate well
Start with tree ensembles like gradient boosted trees, including XGBoost or LightGBM, or Random Forest for baseline robustness. They handle interactions like pitch type x weather incredibly well.
Add calibrated logistic regression for the classifier. This is excellent for probability calibration and monotonic constraints, and it gives analysts a clean baseline. You can also utilize quantile regression forests or gradient boosting for distributional estimates like the P50, P75, and P90 of runs allowed. Keep the stack small, because overfitting is the enemy on daily slates.
Nested cross-validation by week to simulate real time
For the outer loop, roll weekly. Train up to week N-1, validate on week N, and test on week N+1. For the inner loop, tune hyperparameters with a time-aware split, such as a 4-week block, to avoid peeking ahead. Refresh features weekly and run a daily micro-update for lineup and weather data if you can. This approximates how ATSwins runs models in production, meaning always forward, never backward-looking.
Metrics that match what bettors need
For flags classification, Precision-Recall AUC beats ROC-AUC when weak matchups are rare. Watch precision at K, like the top 3 flags a day, and track the Brier score.
For regression, use mean absolute error for interpretability, and pinball loss for quantile models so you can price tails. For totals and props alignment, track the correlation between predicted runs allowed and market closing lines. Monitoring this delta flags edges and drift. You can review standard modeling metrics and calibration tools in popular data science libraries to implement this properly.
Calibration and probability honesty
Plot reliability curves to bucket predictions into deciles and compare predicted probabilities vs observed frequencies. Use Isotonic or Platt scaling, fitting on validation data only and locking it before the test phase. Recalibrate monthly as the run environment shifts. Set thresholds by target precision. If you are aiming for a 60% hit rate on weak tags, solve for the probability cutoff that gives that exact result on validation. Well-calibrated probabilities let us set profit-oriented thresholds and avoid overtrading.
Make the “why” visible with SHAP and constraints
SHAP values show the top 5 drivers for each flagged pitcher, such as velocity down 1.7 mph, chase% down 4.2%, the lineup holding a +0.040 xwOBA vs sliders, wind out 10 mph, or high bullpen fatigue. Analysts trust transparency.
Enforce monotonic constraints for sanity. Make sure higher velocity deltas reduce risk, more rest days reduce risk, and higher wind-out increases run risk. Constraints fight weird shortcuts the model might take. Interpretability shortens the loop from model output to confident action.
Backtest right, guard against survivorship bias
Run rolling-origin backtests to simulate the season week by week, using only what was known at that exact moment in time. Include all scheduled games and probable starters, not just those who actually started. This avoids survivorship bias, which happens when you only see healthy, confirmed starters.
Compare your outputs to simple baselines like market closing lines and pitcher ERA over the last 30 days. If you don’t beat them out-of-sample, fix features before swapping models. We also track stability across months as league run-scoring and ball travel change.
Putting it to work today plus safeguards
A practical workflow you can copy
Below is the step-by-step pipeline we run when we build weak pitching flags for ATSwins MLB slates:
Data ingest (early AM)
Pull prior-day Statcast pitch-level data and update rolling features. Update park and weather forecasts for each venue, noting temperature, wind angle, and speed. Refresh bullpen usage and rest counters from game logs.
Probables and projected lineups
Load probable starters and preliminary projected lineups. Build pitcher x lineup matchup rows with platoon and pitch-type exposure features.
First model pass (late morning)
Run classification and regression models on pre-lineup features. Produce preliminary weak-matchup probabilities and expected runs allowed.
Analyst review and early price shopping
Surface top-10 risk flags with SHAP attributions. Analysts sanity-check velocity dips, pitch-mix changes, weather shifts, and bullpen states. Compare to opening markets for totals and team totals to identify early EV.
Lineup confirmation updates (90 to 120 minutes pre-first pitch)
Incorporate confirmed batting orders and any umpire assignments, then rerun the model quickly. Apply exposure caps if weather uncertainty rises.
Final signals and bet selection
Convert probabilities to bet decisions using expected value thresholds. Write notes into ATSwins dashboards with the key drivers, then push picks and props.
Postgame logging
Record outcomes and model residuals. Tag misses by driver, such as an unexpected pitch-mix change, and add them to a retraining buffer to review weekly.
This cadence keeps us fast and careful. It also meshes perfectly with how bettors actually make decisions throughout the day.
Setting action thresholds by expected value, not vibes
A simple EV framework to gate which weak flags become bets looks like this. For a team total over X at -110, convert -110 to an implied break-even of 52.38%. Use your model to estimate P(over X). If that probability is 57%, then EV equals 0.57 minus 0.5238, giving you +4.62%. Set a house threshold, say +3% for pregame and +2% for live, and only take bets above that threshold.
For F5 opponent team total overs, tie P(3+ runs allowed in 5 IP) and the distribution shape to the price. For pitcher props like over 1.5 HR allowed or under outs recorded, use the regression model’s HR expectation and the distribution tail to approximate prop probabilities. Automate EV calculations and let the analyst decide which edges to keep based on market liquidity and portfolio balance.
Blend AI outputs with human context that models miss
Lineup quirks can change everything, such as veterans getting a day off, a call-up with raw power, or a platoon guy who punishes one specific pitch. Travel and schedule spots matter too, including Sunday getaway days, cross-country flights, or day-game after night-game battery changes.
Look out for health notes like throwing program tweaks, blisters, or minor injuries that won’t hit the IL but sap command. This is where a professional desk adds value fast. Model first, context second, bet third.
Track drift when the league changes under your feet
The run environment drifts with ball composition, weather, and tactics. Monitor calibration and error by month. If predicted runs are consistently low compared to observed runs, re-center park and weather multipliers and retrain.
Watch feature importances over time. If the league chases less or attacks heaters up more, the model may shift away from certain pitch-type interactions. Recalibrate probabilities monthly, because small fixes here save big edges. ATSwins also looks across sports. We use this exact same discipline to manage drift in MLB as we do for other sports, ensuring our long-term accuracy.
Cap exposure when weather risk spikes
If forecast variance is high due to storm cells or shifting winds, cut stake sizes or delay your wagers to live betting. Use a weather risk score. When it rises above a threshold, halve your bet size or cap per-game exposure to avoid correlated losses. Simple sizing rules tame variance.
Log misses and turn them into features
If a pitcher’s pitch mix shifted 10 points after a rough first inning, add a pitch-mix volatility feature based on prior starts. If the catcher changed and framing collapsed, add catcher identity to the feature set earlier in the day. If an umpire’s zone had an unusual squeeze, log it. If it is persistent, include him among zone-profile umps. A tight logging loop converts pain into profit the next week.
Dashboards that surface the right red flags
For analysts and subscribers, fast visuals beat walls of numbers. Build lightweight tiles for each probable starter featuring velocity deltas by pitch, rolling xwOBA allowed with a league baseline overlay, and hard-hit% and barrel% trend lines.
Include whiff%, chase%, and F-Strike% changes with heat colors, alongside a pitch-mix pie vs an opponent strengths list. Add a park and weather card with a wind arrow and temperature, plus a bullpen fatigue dial showing reliever availability. This is what we show inside ATSwins MLB props and picks areas to give context with one glance.
Useful tools and a simple build template
Core tools we recommend are nothing fancy. For data, use the official resources for pitch-level events and park context, alongside industry-standard leaderboards for plate discipline, platoon splits, and dashboards. For programmatic access, use pybaseball to fetch Statcast and standings data into Pandas.
For modeling, use scikit-learn for logistic and linear models, XGBoost or LightGBM for tree ensembles, SHAP for interpretability, and statsmodels for quick baselines. For ops, a Postgres or BigQuery warehouse, a daily cron, and a small Flask or Streamlit app is plenty.
A barebones template to structure your pipeline includes data tables for games, pitchers, hitters, lineups, and the bullpen. Follow this with feature joins to combine pitcher and lineup data, add weather adjustments, and include the bullpen availability index. Define your labels, train your models with time-aware cross-validation, and serve your daily batch inference. If you want a one-sentence rule, start small, ship daily, and improve the feature store every week.
Practical do’s and don’t's that keep you profitable
Do normalize everything to the park, weather, and league month. Don’t compare raw xwOBA in Denver to San Diego without adjustments. Do weight recency, but don’t let a single bad start redefine true talent. Do check bullpen context, and don’t assume a starter is on a normal leash after a 110-pitch outing.
Do price in lineup confirmations, and don’t finalize until you see how many lefties or righties are playing and who is resting. Do use probabilities and EV, and don’t chase narratives without prices moving your way. Do log and revisit misses, and don’t let bad beats turn off a profitable process. These aren’t fancy rules, but they compound advantages day after day.
Turning weak matchup flags into specific bet types
First five overs and opponent team totals are primary targets when the model shows elevated early damage and a short leash for the starter. Full-game overs are better when bullpen fatigue aligns with starter weakness and weather is friendly.
Pitcher under outs recorded is strong when walk probability is up and the opponent’s patience is high. Pitcher over ER allowed or HR allowed props should be tailored to barrel% spikes and fly-ball tilt, especially on wind-out days. Use same-game parlays sparingly, but if the model shows correlated value, combine an opponent team total over with a home run hitter against his worst pitch. Allocate stakes across a small set of the best edges to avoid overexposure to one game’s variance.
Examples of weak-matchup patterns you’ll see often
You will frequently see a four-seam heavy righty whose velocity is down 1.5 mph facing a lefty-leaning lineup that hammers high heaters, with 82°F weather and the wind blowing out. In this scenario, the model will push F5 opponent team total overs and likely HR props for the top two lefty bats. You can check the current MLB standings and team splits to see which squads match this description perfectly.
Another pattern is a command-first sinkerballer with a tight-zone umpire and a patient lineup. Expect early walks, long counts, and elevated run expectation despite a low barrel% year-to-date. Finally, watch for a slider-dependent starter vs a team that demolishes sliders, especially if his chase% has dropped and first-pitch strikes are down. If the bullpen behind him is thin, full-game overs come into play. Keep a notebook of these recurring shapes to help you choose between similar edges when markets move.
How this plugs into ATSwins products and tracking
We publish the top weak-matchup flags with the core reasons, attaching recommended bet types with clear EV ranges. When public money steams the other way but our model is steady, that is often where value sits. We note it on the slate page so you can get ahead of late moves. Every flagged bet logs pregame probability, EV estimate, and model deltas. Over weeks, you see which drivers print and which are noisier.
We keep everything transparent so you understand what you’re betting and why. You can find deeper strategic examples in our guide on Stanley Cup Finals value or read our breakdown of MLB betting edges to see how these methods look in practice.
Troubleshooting: when the model and market disagree
If your model flags a weak matchup but the total drops, re-check weather and lineups because you may be using stale inputs. If you’re consistently high on run expectation for one park, recalibrate park factors and review carry vs weather assumptions.
If your weak flags hit but props keep losing, ensure you’re translating the regression to the correct prop distribution. If precision on top-3 daily flags falls below target for a week, lower the probability threshold until you retrain and stabilize. Short feedback loops keep you both honest and profitable.
A few closing reminders about primary sources
Always anchor on primary data. Use Statcast for pitch-level truth via Baseball Savant, and rely on official leaderboards for definitions and consistent aggregates.
Save your own rolling features daily to prevent hindsight bias. Finally, implement lightweight weather and park models that you thoroughly understand rather than trusting a black box. When your inputs are clean, weak pitching matchups reveal themselves, and ATSwins can turn them into measured, repeatable bets.
Conclusion
We showed how AI spots weak pitching by analyzing contact quality, platoons, park, weather, and bullpen context. Your key moves are to trust calibrated models, blend news, and manage risk instead of relying on vibes. Your next step is to track a few slates, note outcomes, adjust, and then scale.
ATSwins is an AI-powered sports prediction platform offering data-driven picks, player props, betting splits, and profit tracking across NFL, NBA, MLB, NHL, and NCAA. Free and paid plans give bettors insights and guides to make smarter, more informed decisions.
Frequently Asked Questions (FAQs)
What is the single most important Statcast metric for flagging a weak pitcher?
There isn't one single metric that operates in a vacuum, but if forced to choose, xwOBACON (Expected Weighted On-Base Average on Contact) combined with Barrel% gives the truest look at a pitcher's current vulnerability. ERA can hide a lot of luck if a pitcher has an elite defense behind him or plays in a massive park. xwOBACON strips away the defense and the park context, evaluating purely the exit velocity and launch angle of the balls thrown. When a pitcher's rolling 3-game xwOBACON spikes significantly above his 2-year baseline, it indicates he is losing command, missing spots, and giving up dangerous contact that will inevitably turn into real runs once his luck runs out.
How many starts does it take for a drop in velocity to become a real model signal?
A velocity drop is one of the unique features that does not require a large sample size to be predictive. A drop of 1 to 2 mph in a single start is an immediate red flag that triggers an automated alert in our pipeline. Unlike plate discipline metrics which can fluctuate based on an umpire or a patient opponent, a sudden loss of velocity usually points directly to physical fatigue, a mechanical issue, or an underlying injury. When that single-start velocity drop aligns with a decrease in first-pitch strike percentage and a drop in chase rate within the same outing, the model heavily penalizes the pitcher's baseline for his next scheduled appearance.
Why do you prefer exponential moving averages over simple seasonal averages?
Simple seasonal averages carry a major flaw known as stale bias. If a starting pitcher had an incredible April and May, his full-season ERA and metrics will look excellent in August, even if his arm is completely spent and he has been blasted in his last three starts. Exponential moving averages (EMAs) solve this by assigning progressively higher weights to the most recent performances while still retaining the historical context of his true talent baseline. By utilizing a half-life of 7 to 14 days for fast-stabilizing metrics like velocity and strikeout rates, the model quickly captures a pitcher's current form without completely forgetting his long-term capability.
How does the model account for late lineup changes after the opening lines are set?
Our system handles this through a strict two-stage inference process. In the morning, the model runs a preliminary pass using highly accurate projected lineups based on team news and historical platoon tendencies. This allows us to spot early expected value against opening totals. However, we do not finalize our heavy positions until 90 to 120 minutes before the first pitch, when managers release their official, confirmed batting orders. The pipeline automatically runs a second-stage micro-update that swaps the projected hitters for the actual starters, recalculating the pitch-type exposure and platoon splits instantly to ensure our final signals are perfectly accurate.
Can an elite bullpen completely nullify a weak starting pitching matchup flag?
An elite bullpen does not nullify the starting pitcher's individual weakness, but it completely changes how we choose to bet that specific matchup. If a starter is flagged as weak but has an elite, fully rested bullpen behind him, betting the full-game over or the full-game opponent team total becomes highly risky because the manager will likely pull the struggling starter early and suppress late-inning runs. Instead, the AI adapts by directing value toward the First 5 Innings (F5) market. This allows us to isolate the vulnerable starter during his exact pitch-mix exposure before the elite relievers can enter the game to clean up his mess.