Agent Lab — a live forecast exhibit · Applied AI
← Schedule
Group A · Matchday 14 · 2026-06-24

South Africa vs South Korea

Guadalupe, Mexico · neutral venue

Full time RSA 1–0 KOR
The agent called South Korea (48%) — Worse than the model call wrong

The forecast below is what the agent published before kickoff.

Most likely scoreline
0–0 17.7%
26.7%
RSA win
31.4%
Draw
41.9%
KOR win
Top scorelines
  • 0–0 17.7%
  • 0–1 16.0%
  • 1–0 12.1%
  • 1–1 10.9%
Total goals
0 to 10 goals · most likely → tallest

The agent’s call

3-agent ensemble · deepseek-chat
RSA win
22.3%
Draw
29.3%
KOR win
48.3%

The model already favours South Korea (45% away win), and the supplied intel reinforces that lean. South Africa's recent form is poor (-0.60 pts/match under expectation), they are missing key midfielder Mokoena, and their squad is significantly less valuable (rank 40 vs 32). South Korea's key players (Son, Lee Kang-In, Kim Min-Jae) operate at a higher club level. I increase South Korea's win probability slightly from 45% to 50% to reflect these advantages, while keeping a non-trivial draw probability (28%) given the model's 30% draw prior and the fact that South Africa are not completely outclassed. The model's expected goals (RSA 0.77, KOR 1.19) suggest a low-scoring match with South Korea likely to score once or twice. South Africa's defensive solidity (recent results low-scoring) and South Korea's attacking quality point to 0-1, 0-2, or 1-1 as the most probable scorelines. I keep the shape similar to the model but slightly shift probability toward away wins.

What the agent weighed
What the agent chose to investigate
The agent decided what to check on tavily — 6 sources kept:
  • “South Korea 2026 World Cup squad injuries lineup June 25” KOR · 3 hits
    Check for late fitness issues or lineup changes that could affect South Korea's strength.
  • “South Africa vs South Korea neutral venue Guadalupe weather forecast June 25 2026” match · 3 hits
    Weather conditions like heat or rain could impact performance at the neutral venue.
Agent scorelines
generated 2026-06-24
Last 5 — RSA
-0.65 pts/match vs model expectation · opp. avg #75
  • PAN #72 -1.7
  • NIC #151 -1.5
  • JAM #91 -0.6
  • MEX #20 -0.7
  • CZE #43 -0.1
GF / GA last 10: 10 / 13
Last 5 — KOR
+0.24 pts/match vs model expectation · opp. avg #73
  • AUT #23 -0.9
  • TRI #141 +0.5
  • EL #140 +0.5
  • CZE #43 +1.6
  • MEX #20 -0.9
GF / GA last 10: 13 / 12
Head-to-head (last 5)

No prior meetings in the dataset.

▸ Full score grid (7×7)
P(home goals × away goals), %
0
1
2
3
4
5
6
0
17.7
16.0
8.7
3.7
1.3
0.4
0.1
1
12.1
10.9
5.9
2.5
0.9
0.3
0.1
2
4.9
4.5
2.4
1.0
0.4
0.1
0.0
3
1.6
1.4
0.8
0.3
0.1
0.0
0.0
4
0.4
0.4
0.2
0.1
0.0
0.0
0.0
5
0.1
0.1
0.1
0.0
0.0
0.0
0.0
6
0.0
0.0
0.0
0.0
0.0
0.0
0.0

Rows = RSA goals; columns = KOR goals. Above-diagonal cells are home wins; diagonal = draws; below = away wins. Truncated at 6 — the joint mass past that is < 1 %.

Expected goals (Poisson rate, λ)
0.79
RSA
1.11
KOR

Independent-Poisson rates. W/D/L and the score grid integrate the full 13×13 joint under these two parameters.

News context — injuries, suspensions, projected XI — is not part of the model, which sees only past results.