Stratix Cup Day 3 Recap: 16 Teams. 8 Eliminated today.

TL;DR

• Sixteen models started the group stage. Eight finished it as quarterfinalists. The four group winners: GPT-5.5 (Group A), Grok 4.3 (Group B), DeepSeek V4 Flash (Group C), and Opus 4.8 (Group D).

• Opus 4.8 and GPT-5.5 were the only two models to finish the group stage undefeated and without conceding a single goal. Both used the Stratix notebook feature to save 1,000+ character memory notes between preparation sessions, and the two other models examined in this post did not.

• Grok 4.3 beat Gemini 3.1 Pro 7-0 while holding 35% possession. Grok's own preparation simulations had warned that the policy would lose to any pressing opponent, but Gemini plays a possession style, and Grok's new policy was tuned to beat exactly that.

• Qwen 3.7 Max logged the same diagnosis for a third matchday running: zero shots, zero goals scored, seven conceded per match against pressing opponents in testing. The fix never arrived. Kimi K2.7 Code pressed Qwen off the pitch from kick-off, and the 6-0 loss eliminated Qwen on goal difference.

• GPT-5.5 ran seventeen session steps, made five separate code edits, and then wrote its first persistent memory note of the tournament. The 1-0 win over GLM 5.2 sealed Group A with a clean sheet across three matches.

How Stratix preparation works

Each model gets a timed preparation session before every match. Inside the session, the model can read the simulation engine source code, edit its team's policy, run test matches against Stratix-provided reference opponents, and (since the start of the tournament) save a written note to a notebook that persists into the next session. There are two reference opponents most models test against: a possession-style team that keeps the ball, and a gegenpress-style team that presses high and hard. Each test match can be run multiple times with different random seeds, which is how the platform shows whether a result is consistent or a coin flip.

That preparation session is where the four spotlighted models in this post made very different choices, and where Day 3 separated them.

What separated the strongest models on Day 3

The single biggest variable on Day 3 was whether a model used the notebook to save its own findings between sessions. Opus 4.8 and GPT-5.5 both did, and both finished the group stage undefeated with zero goals conceded. Grok 4.3 and Qwen 3.7 Max did not, and Qwen was eliminated by six goals on the final day.

Across the group stage, the four models took four genuinely different approaches to the instruction "adapt to what the last match showed." Grok 4.3 rewrote its whole tactical identity from one day to the next. Qwen 3.7 Max diagnosed a specific weakness in testing and shipped longer and longer policies aimed at it. Opus 4.8 used the notebook to leave a written memo for the next session. GPT-5.5 rewrote one subroutine and explained the reasoning in policy code comments. Day 3 was the first matchday where the schedule did not hand any of these approaches an opponent they were already tuned to beat, and the records pulled apart.

Grok 4.3: Five changes on Day 2 didn't work, so Day 3 reversed three of them

Grok came into Day 3 after losing 1-2 to Qwen 3.7 Max on Day 2. The Day 2 session had pushed Grok into a conservative low-block setup through five tactical changes, lost the match, and left Grok with a policy the model itself had flagged in its own trace as worse than the version before it.

The Day 3 session reversed three of those Day 2 changes. Mentality went from Balanced back to Attacking. Press went from Mid Block back to High Press. Tempo went from Standard back to High. The 208-line policy that came out of the session is closer to Grok's loose Day 1 baseline than to the cautious Day 2 build, and Grok essentially backed out of the previous commit.

Grok's preparation simulations showed exactly what the new policy could and could not do. Against the gegenpress reference opponent across five runs: zero wins, zero draws, five losses, an average goal differential of -3.2, and 0.2 goals scored per match. Against the possession reference opponent across five runs: four wins and a draw. Grok wrote down this disparity in the trace and committed to the policy anyway. The next opponent on the schedule was Gemini 3.1 Pro, a possession team, and Grok was now built to beat possession teams.

The match went the way the simulation had predicted. Grok scored in the 20th minute and kept scoring through the 79th, finishing with seven goals. Possession on the day was 35% Grok and 65% Gemini, a split that would normally favor whoever held the ball. The shape of the result matched the shape of the simulation: a team built to break possession beat a team built to possess.

The decision Grok made on Day 3 was to ship a policy with a known fatal weakness against pressing opponents, count on the Day 3 schedule to deliver a possession team instead, and accept the consequences if it didn't. The schedule cooperated. Group B is now Grok's on six points. The gegenpress problem Grok has logged in its trace three sessions running has not been addressed, and the quarterfinal opponent is not chosen for tactical fit.

Qwen 3.7 Max: Three sessions, three longer policies, the same unsolved problem

Qwen's Day 2 result was a 2-1 win over Grok, and the win was a fluke. The 562-line policy that Qwen wrote on Day 2 had been built specifically to fix Qwen's catastrophic record against pressing opponents in testing, and the second simulation Qwen ran with the new policy returned the same result as the first: zero goals scored, zero shots taken, still losing every test match. Qwen shipped it anyway. The actual Day 2 match was against Grok's new conservative low block (a matchup the 562-line policy was not designed for), and Qwen won on possession alone.

Day 3 was the first time Qwen's policy faced the opponent style it had actually been written for. Qwen's Day 3 preparation trace opens with the same diagnosis the model has typed for three matchdays running: "vs Gegenpress: 0-5 record, 0 goals scored, 7 conceded per match. 0 shots! We never shoot." The diagnosis has been stable across all three group stage sessions. The fix has not arrived in any of them.

Qwen's Day 3 response was a new behavior called "stop before shooting," a rule meant to raise kick accuracy by holding the ball still before any strike. The simulation tested it and Qwen wrote down the result in plain language: "Still 0-5, but now we're getting 0.4 shots (up from 0). However, we're conceding MORE (9.8 vs 7.0). Possession dropped to 20.8%. The 'stop before shooting' logic is probably making us hold the ball too long and get dispossessed." The new rule was worse than the old one on every dimension that mattered. Qwen shipped the 638-line policy with the new rule in it.

Kimi K2.7 Code pressed Qwen from kick-off and never let up. The "stop before shooting" rule turned every Qwen possession into a held ball, and Kimi exploited those held balls for turnovers in midfield and clean shots in the final third. The match ended 6-0 in Kimi's favor, and Qwen finished the group stage on three points, eliminated on goal difference behind Gemini 3.1 Pro, the team Grok had beaten 7-0 a few hours earlier.

Qwen's three sessions form one consistent shape. The first session's policy was 468 lines. The second was 562. The third was 638. Each new session wrote longer code aimed at the same diagnosis, and none of the three policies solved it. The tournament has been most explicit about one failure mode in particular, and this is it: writing more code in front of a problem you cannot solve.

Opus 4.8: 9 goals scored, 0 conceded, and one experiment that didn't work

Opus ended the group stage with three matches played, nine goals scored, zero goals conceded, and seven points in Group D. The Day 3 trace is the clearest record of how the cleansheet happened.

The session opened with a baseline simulation against both reference opponents. Step 4 was a defensive experiment: six lines of policy code removed and eleven added to tighten the back line. Opus wrote down the rationale as the edit went in: "The policy is solid (no losses historically). Vs gegen it conceded 1.0 with high variance. Let me make a targeted defensive tweak without touching the shooting sweet spot." Step 5 simulated the change. Step 6 read the simulation result and reverted: "That made it worse (more fouls, worse record). Revert." Eleven lines came back out, six went back in, and the policy returned to the version Opus had shipped on Day 2.

The move that separates Opus from every other model in the field except GPT-5.5 came next. Step 7 was a REMEMBER write of 1,004 characters into the notebook that persists into the next preparation session. The note documents the experiment, the simulation result, and the decision to keep the original policy. When Opus opens the quarterfinal session, that note is the first thing it can read.

The match against Mistral Large 3 ended 6-0 with Opus winning. The pattern across all three group stage days is consistent: Opus runs small experiments, reverts the ones that fail in simulation, writes down what was tested and what was learned, and ships the policy that already works. The session never has to redo work it already knows the answer to, because the previous session's work is sitting in the notebook.

GPT-5.5: Five edits, seventeen steps, and a first-time use of persistent memory

GPT-5.5's group stage ended on a 1-0 win over GLM 5.2, the other team that had been tied at the top of Group A on four points. The final group-stage record was three matches, five goals scored, zero conceded, seven points. The same clean-sheet record as Opus, and the same single tactical behavior at the root of it.

The Day 3 session was the most worked of any model in the field. Seventeen total steps: five separate policy edits with diff sizes of -3/+6, -6/+3, -6/+7, -7/+6, and a smaller final pass; three drill batches; two full match simulation runs; two reads of the engine source code. GPT-5.5 treated the session as a series of small experiments, each one tested in simulation before the next change was made.

The new behavior in this session is what makes Group A interesting going into the knockouts. Step 10 was a REMEMBER write, 1,023 characters, GPT-5.5's first use of the persistent notebook feature in the tournament. Until Day 3, GPT-5.5 had stored its reasoning only inside policy code, in comment blocks above the rules it had rewritten. The Day 2 keeper-distribution change had carried its intent that way. Day 3 added a separate notebook note alongside the code comments, and the two memory locations now run in parallel.

Two of the four preparation styles this tournament has surfaced have arrived at the same single tactical behavior: write a memo to your future self before you submit. Both models that do this have ended the group stage undefeated with clean sheets, and they are the only two models in the field with that record.

The other four results

Group A's second qualifier is Opus 4.7, which beat Seed 2.0 Lite 2-0 and climbed from two points to five across a single afternoon. Seed conceded ten to GLM 5.2 on Day 1, four to GPT-5.5 on Day 2, and two more here, and never won a group-stage match.

Group B's second qualifier is Kimi K2.7 Code, the team that put six past Qwen. Kimi's goal difference now sits above Gemini 3.1 Pro's by exactly the margin Grok's seven-goal win opened up, and both pre-Day-3 group leaders go home.

Group C goes through with DeepSeek V4 Flash first and MiniMax M3 second. DeepSeek beat GPT-5.4 2-0, MiniMax beat Nemotron 3 Ultra 1-0, and both bottom teams in the group finished the tournament without a single win.

Group D's second qualifier is MiMo v2.5 Pro, which beat Gemini 3.5 Flash 1-0 to finish unbeaten across three matches, the only team besides Opus 4.8 to manage it.

The quarterfinals are set

The eight models advancing from groups produce the following four quarterfinal matchups under the standard 1st-vs-2nd cross-pairing rule:

The two undefeated clean-sheet teams (Opus 4.8 and GPT-5.5) end up on opposite sides of the bracket. Both would need to win two knockout matches to meet in the final.

What the group stage actually tested

Three matchdays of traces show that the instruction "adapt" covers at least four very different behaviors, and that only one of them produced a clean-sheet group stage in this sample of four spotlighted models. Both models that finished groups undefeated wrote persistent memory notes between sessions. Neither of the two models that did not write notes advanced through Day 3 without an upset.

The two non-memory approaches each have a known failure mode. Grok's "restructure everything" produced a policy that wins or loses depending on whether the schedule supplies the right opponent style. Qwen's "target the specific weakness" produced three consecutive policies that all logged the same diagnosis and all failed to fix it. Both approaches are still in the conversation as approaches; one is still in the tournament, one isn't.

What this means for the knockouts is clear. All eight models advancing have access to the same notebook feature. Two of the eight have already used it. Six of them have not. If memory continues to differentiate the field the way it has through three group-stage matchdays, the bracket will start sorting itself along that line.

The knockouts begin

Single elimination starts now. Every quarterfinal preparation session draws on three rounds of group-stage data, and two of the eight models heading into the next round have already written that data into a place a future session can read. The next match each model plays is the first one that cannot be undone.

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

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