Source: arXiv

Paper

Auditing Contextual Bias in Human Ball-Strike Calls Using KBO’s Automated Umpiring Transition — Kichang Lee, JeongGil Ko (arXiv 2609.03786, Sep 2026).

What it claims

Baseball fans have long argued that human umpires shrink or expand the strike zone by count, inning, catcher, and home context — but confounds make that hard to prove. The Korean Baseball Organization’s Automated Ball-Strike (ABS) rollout from 2024 is a natural experiment: same league, machine-drawn zone.

Lee and Ko model called-strike probability on taken pitches near the boundary using KBO pitch data from 2021 through available 2026 games. 2022–2023 are the human-umpire baseline; ABS seasons are the benchmark. The headline is not “umpires are corrupt everywhere” but a map of where human judgment was most context-sensitive.

The strongest signal is count pressure. Relative to 0–0, humans called fewer strikes at 0–2 and more at 3–0. In the main 0.25-ft boundary band, effects were −17.17 and +6.61 percentage points. Under ABS, those effects neared zero and failed false-discovery-rate correction. Inning effects were smaller — fewer early strikes, more in 7–9+ late-close games — while ABS seasons were flat. Salary and catcher proxies show human-period heterogeneity that disappears under ABS; home-context evidence is mostly null, with one FDR exception and an exploratory umpire–team gap treated as an audit lead.

The breakdown

This is an audit paper, not a crooked-officiating hot take. ABS is a fixed ruler for borderline pitches — where discretion lives. KBO went league-wide, unlike piecemeal MLB trials, yielding enough volume to test subtle count and inning effects.

Count results match broadcast intuition — protect the hitter on 3–0, expand for the pitcher on 0–2 — but quantified in percentage points and shown to vanish under automation. Inning drift is smaller. Broad home favoritism and stable catcher-framing advantages did not show up strongly once ABS removed the human receiver.

Why readers outside the lab should care

Overseas KBO followers and anyone building Korean baseball media or analytics should era-tag pre-2024 strike stats: count drift on borderline pitches was real and sizable; ABS flattened it. League-wide adoption also shows how governance plus tech rollout exposes hidden judgment — narrowing fan debate from “always biased” to where context actually moved the zone.

What builders and Korea-touching teams should watch

• Do tag KBO pitch datasets as human-umpire (≤2023) vs ABS (2024+) before modeling strike probability or framing value.
• Don’t treat pre-ABS count splits as if they still describe today’s called-strike environment — 0–2 and 3–0 effects near −17 and +6.6 pp largely disappear under ABS.
• Expect inning and late-close human-era drift to shrink post-ABS; game-state features trained on old seasons may misfire.
• Re-check home-field and catcher-framing storylines against ABS seasons — heterogeneity that mattered with humans may not survive automation.
• Treat exploratory umpire–team gaps as audit leads, not proof of systematic favoritism.

Context

Read this as KBO’s ABS transition used as a league-scale lie detector for strike-zone context, not as proof that every umpire suspicion was wrong. Korelay’s frame: count pressure was the human strike zone’s biggest bend; automation straightened it — and the paper shows where the evidence was weaker than the broadcast booth assumed.

Source

Primary: arXiv:2609.03786 (abstract and framing cited; open the OA PDF for methods and full results). Do not republish the PDF.