Matchday 1 Recap: What the Traces Actually Showed

TL;DR

• Five of seven match winners had negative pre-match goal differentials.

• Gemini 3.1 Pro had a -13.4 average GD in testing, the worst of all 16 models, and won 4-1.

• GLM 5.2's session was cut short by an SSL error before it could iterate. It won 10-0.

Sixteen AI models wrote their own soccer strategies. Eight matches. Seven decisive results and one draw.

Before each match, every model trained against the same two test opponents: a "gegenpress" opponent (high pressure, high intensity) and a "possession" opponent (measured, patient). The gegenpress broke everyone. Not a single model solved it cleanly in testing, and the test predicted nothing about what happened next.

Two Models Crashed Before They Could Prepare. Both Won.

GLM 5.2 lost all five of its pre-match simulations. A 603-line policy, a -5.8 average goal differential, and an SSL connection error that killed the session before it could iterate on anything. One draft. No revisions. No simulation feedback loop.

The result: 10-0 against Seed 2.0 Lite, the largest margin of any match on Matchday 1.

Seed had actually done more work. Twelve steps, three simulation rounds, a kickoff-state crash it found and fixed, an improved average goal differential (from -9.4 to -3.0) through iteration. None of it mattered. GLM 5.2 scored from minute 12 through minute 89 while Seed finished with 75% possession and 10 goals conceded.

The same pattern played out in Group B. Grok 4.3 submitted a cold policy with zero simulations run after an SSL error killed its session at step 6, immediately after confirming the policy loaded and generated valid commands for all 11 players. Its opponent, Kimi K2.7 Code, had read sim/world.py three separate times, run 5 drills, and improved its gegenpress deficit from -11.7 to -3.7 before timing out on the final submission.

Grok won 1-0 on a 14th-minute goal. 33% possession.

Two crashed models, two wins. Less preparation produced better results in both matches. But the next set of results complicated that story.

Two Models Found One Bug and Fixed It. Both Won.

"Players l2, l3, l8 all attempting 3 tackles in 1 second."

MiniMax M3 found this in its own trace after the first simulation. At 725 lines, it had written the largest first draft of any model, and the problem was mechanical: three defenders spamming tackles every tick. After renaming the approach from "press-and-possess" to shadow marking, fouls dropped from 13.8 per match to 2.8.

The fix was enough. MiniMax beat GPT-5.4 1-0 on a 62nd-minute goal. GPT-5.4 had iterated in the wrong direction: its initial policy produced a -6.0 average goal differential, and after one round of edits the numbers got worse (zero goals scored, 7.3 conceded on average). 57% possession on the day. A loss.

DeepSeek V4 Flash had a different bug but the same approach. Across 17 session steps, it focused on a single issue: its goalkeeper was not receiving commands during kickoffs. From the trace:

"Aha! The kickoff is treated as a setpiece with ours=True and setup_ticks_left=0. So my code enters _handle_our_setpiece for kickoffs but does not have a kickoff case."

The fix worked. A regression followed (all 11 players dropped to 2). That got fixed too. Then the session timed out before completing its simulation round. Final logged result: -7.3 average goal differential.

Won 5-0. Scored in minutes 5, 9, 18, 28, and 74.

Neither MiniMax nor DeepSeek solved the gegenpress problem. Neither posted good test results. What they did was find a concrete implementation flaw, fix it, and win by wide margins.

Two Models Diagnosed the Problem Perfectly. Neither Could Fix It in Time.

Qwen 3.7 Max wrote the clearest self-diagnosis of any model on Matchday 1:

"Interesting results: vs Gegenpress: 0-5, getting crushed (conceding 7.6 goals, only scoring 0.6)... The gegenpress is overwhelming our midfield and we're not handling pressure well."

The numbers confirmed it. Average ball position: -11.9 (constantly pinned in its own third). 41.8 turnovers. Only 1.4 shots. Qwen knew exactly what was wrong, drafted a fix, and timed out before it was ready. The unresolved version went to match day. Lost 4-1 to Gemini 3.1 Pro.

In Group D, Gemini 3.5 Flash had the opposite timing problem. Rather than writing a fresh policy, it ran simulations first against an existing one. The session timed out mid-analysis in a THINKING step with no output. No new policy ever got submitted. Lost 0-3 to Opus 4.8.

The session budget creates a hard ceiling. Both models identified their weaknesses accurately but ran out of compute before the fix was ready.

Gemini 3.1 Pro Posted the Worst Test Results of All 16 Models. Won 4-1.

At 29 session steps, Gemini 3.1 Pro worked harder than any other model on Matchday 1. Defensive clearance strategies, then short-passing strategies, then more iteration. Its final submitted policy still lost every single simulation with a -13.4 mean goal difference, the worst testing result in the tournament.

Then it won 4-1 against Qwen. Qwen had even scored first to equalize at 1-1 before Gemini put three unanswered past it.

The other Group A match told a quieter story. GPT-5.5 held 68% possession against Opus 4.7 without converting a single chance. Opus 4.7's trace was blunt about its own problems: "We're losing badly to gegenpress (0-7). The issues: way too aggressive, fouling constantly, and 64 shots faced." A full policy rewrite, from 378 lines down to 332, still couldn't produce a goal. 0-0.

One Model Actually Improved Its Test Results

Most models added code when they iterated. Opus 4.8 cut. Its first draft was 189 lines. After two revisions: 160. When the edit tool broke mid-session, it rewrote the entire strategy from scratch rather than patching. The rewrite was shorter.

The starting read was clear: "We beat possession but lose to gegenpress. Need better defense (conceding 3) and possession retention." By submission, goals conceded had dropped from 3 per match to 1.2. No other model in the tournament reached a neutral goal difference against the gegenpress test. Beat Gemini 3.5 Flash 3-0.

MiMo v2.5 Pro took the same instinct in the other direction. Starting from 615 lines, it saw it was "getting crushed by gegenpress (0-5)" with 34 shots faced and halved the policy to 351. Its opponent, Mistral Large 3, had written a full strategy without reading any source files, starting from a "placeholder that chases the ball with all players." Mistral scored first at minute 5. MiMo equalized at 8 and won 8-1.

The Pattern from Matchday 1

Two crashed models won. Two bug-fixers won. Two accurate diagnosticians lost because they ran out of time. Gemini 3.1 Pro posted the worst simulation results in the entire tournament and won anyway. Opus 4.8 actually solved the test and went on to win the whole thing.

For six of seven decisive matches, the traces told a different story than the scores. Opus 4.8 was the only model where preparation aligned with result.

Matchday 2 starts Tomorrow at 7am PT/10am ET

Watch every match at layerlens.ai/stratix-cup/season-1

All trace data and reasoning quotes sourced directly from the Stratix Cup match viewer.