detect
stage 1 of 5
watch the play-by-play
It polls the MLB Stats API for every game in progress. When a runner
is retired and an outfielder started the throw, that play gets pulled
out of the feed — nothing else counts as an outfield assist.
The raw play description is what triggers
everything downstream.
real play · from plays table
“Hao-Yu Lee singles on a sharp line drive to left fielder Cody
Bellinger. Riley Greene out at home on the throw, left fielder Cody
Bellinger to catcher Austin Wells.”
parsed to → LF → C ,
target base Home , Cody Bellinger throws out Riley Greene.
| v
tier
stage 2 of 5
score the throw
Each play earns points and lands in one of three tiers. Target base
pays the most (Home 4, 3B 3, 2B 1).
A direct throw adds 2; a relay with
3+ fielders subtracts 2. Available video adds 1,
and an out that only stood because of a
replay overturn subtracts 2.
A throw clocked at 95+ mph adds 1 more,
on the rare play where Statcast velocity is on hand.
Totals map to
real credit chain · scores medium
RF → SS → C target base
Home video yes
target base: Home +4
relay chain (3 segments) −2
video available +1
score 3 → = 3
A throw home would tier high on its own, but the cut
through the shortstop is a relay — the −2 penalty drops it
to medium.
| v
post
stage 3 of 5
drop it in slack
The play goes to the channel with its Baseball Savant video. When
Savant carries more than one camera, every angle is posted —
home, away, and high-home — not just the first one it finds.
janitor-bot APP · 7:14 PM
🔥 CF → C · Home ·
Ceddanne Rafaela guns down Edouard Julien at the plate (MIN @ BOS, top 7)
react 🔥 if it rips · 🗑 if it’s a nothing play
| v
vote
stage 4 of 5
let the channel judge
Teammates react 🔥 for a real gem and 🗑 for a dud. Votes
are snapshotted per play into a fire count, trash count, and net score.
When the crowd and the tier disagree — a low play people love, a
high play they ignore — the play is flagged for the weekly review
to look at.
real tally · from vote_snapshots
LF → C · Elly De La Cruz
throws out Matt McLain at home
🔥 fire 3
🗑 trash 0
net +3
Not every play lands this cleanly. A separate throw
home this season pulled 0🔥 / 2🗑 (net −2)
against a medium tier — flagged
channel_disagrees_high_or_medium for review.
| v
review
stage 5 of 5
audit the week
Once a week an agent reads every detection against its votes and the
prior weeks’ history. It files findings about where the scoring
and the channel disagree, each tagged to a suspected rule area.
A human marks each finding confirmed or rejected . The
confirmed ones become the next change to the tiering rules — the
relay penalty above came out of exactly this loop.
real finding · from agent_findings
severity: info
trend: recurring
outcome: confirmed
“Play 333 (LF→SS→C→3B, medium tier, CWS@PHI) is
the only play this week to receive a trash vote (net score −1).
The credit chain involves four hops — the longest relay chain in
the week’s dataset — yet the play was classified medium
rather than low.”
rule area: new_tunable_needed
· directionally consistent with a prior confirmed finding on
relay length → fed the relay penalty now in ranking.ts .