Per-Agent Calibration Scores: Spain vs Belgium

Result: ESP 2-1 BEL (90') — Fabián Ruiz 30', De Ketelaere 41' (BEL), Merino 88'. ESP won in regulation, advance to SF vs France.
Auto-scored baseline: winner_90 + advance + scoreline → mm3 25/50 | g4 45/50 | g5 45/50. who_scores 0/20 all (no #1 or top-3 pick scored for any agent).

🆕 g4 = Kimi K2.7 Code (ollama/kimi-k2.7-code:cloud) — DEBUT MATCH. This is the first data point for the new g4 model (replaced glm-5-turbo Jul 9). The 24-match glm-5-turbo era record (avg 58/80, Band C) is now closed; Kimi starts a fresh calibration record here. Result: g4 = A (90%) — the highest of the three, driven by being the ONLY agent to correctly project Spain's possession dominance (~62%) instead of applying the R32-5c flip cap. One match is a small sample; do not over-extrapolate.

Headline calibration learning: All three correctly picked ESP to advance and win in 90', and all three staked the right model-breakers (no GK change, Oliver lenient, F11 neutralized, BEL Plan-A restoration, Courtois open-shape diluted). The spread (mm3 47 / g5 67 / g4 72, a 25pt gap) is driven by two genuine reasoning differences: (1) the scoreline modal (mm3 draw-hedged 1-1 → 0/25; g4/g5 ESP 1-0 with 2-1 in-list → 20/25), and (2) the possession projection (g4 dominance ✅ 5/5; mm3/g5 flip-cap ❌ 0/5). Both are honest signals, not noise.

mm3 (MiniMax-M3)

Findings: 13/20

#	Finding staked	Direction	Magnitude	Score	Note
1	ESP 0 GA compound = load-bearing edge (+5pt momentum cap); Simón "clean-sheet form real but ×0.25 weight — untested vs sustained pressure"	✅	❌	3	Defensive edge held (ESP won, best block at the WC), but the 0 GA magnitude broke at 41' (De Ketelaere). The "untested vs sustained pressure" caveat was prescient — exactly what materialized. Direction right, magnitude wrong.
2	R32-5f: Courtois confirmed-form but in OPEN 4-2-3-1 (not low block) → draw-forcing muted to ×0.25-0.3 (not ×0.5-0.6)	✅	✅	5	Courtois conceded 2 from ESP's ~2.0 xG / 17 shots. Hot-GK diluted in open shape validated (3rd QF after Bounou FRA-MAR). Exact call.
3	F11 weather NEUTRALIZED — 22.8°C, fixed ETFE canopy, no heat penalty	✅	✅	5	Brief confirms 22.8°C, canopy, F11 neutral. Clean.
4	F8/R32-5c possession flip cap — "Cap ESP possession at 48-52% per R32-5c" (favorite_possession=52)	❌	❌	0	Actual ESP 67.8% / BEL 32.2% — NO flip; classic possession-dominant win. Brief: "the KO-favorite possession-flip pattern did NOT fire here." Note: a defensible ~50% default that lost the coin-flip, not a reasoning failure — but as staked (cap applies), it did not apply.

(Additional staked findings, not in top 4: F12 conservation 0% = 5/5 ✅ both desperate, no conservation; BEL bench cavalry (Lukaku 14.3% sub-impact as BEL's "main path to victory") = 3/5 — direction right (it's genuinely BEL's edge) but magnitude wrong, Lukaku 0 goals, cavalry deployed-but-ineffective, brief learning #5 "bench is a threat, not a guarantee.")

Model-breaking: 9/10

All material model-breakers caught. GKs confirmed pre-match (Simón + Courtois, no change) ✅. Oliver behaved exactly as projected: 4 yellows (projected 3-4 ✅), 0 pens (projected 10-15% → 0 ✅), 0 reds ✅. The ESP Fabián-Ruiz-for-Pedri rotation was anticipated AND Fabián Ruiz was listed at 6% in Who Scores (he scored the opener). The 🚨 BEL Plan-A restoration (KDB+Doku→START, 4-3-3→4-2-3-1) was explicitly called ("USA low-block was a surprise; QF = Plan A 4-2-3-1"). Weather + formation (both 4-2-3-1) correct. Only soft miss: Merino as the sub-scorer (mm3 had him in the bench list but not weighted). Deduction is conservative — no agent is perfect; the 0 GA prior was treated as near-immutable (mm3 at least caveated it).

TOTAL: 47/80 → 59% → Band D

(auto 25 + findings 13 + MB 9. Slipped to D on two genuine misses: the 1-1 draw-hedged modal cost the auto scoreline, and the possession flip cap cost a finding. The draw band itself — 30% — was a reasonable QF hedge, not docked here; the modal-winner was right (38% ESP win ✅).)

g4 (Kimi K2.7 Code — DEBUT)

Findings: 18/20

#	Finding staked	Direction	Magnitude	Score	Note
1	ESP 0 GA compound = "the single biggest edge," +5pt (confirmed-form ×0.5-0.6)	✅	❌	3	Edge held (ESP won), clean sheet broke at 41'. Direction right, magnitude wrong. g4 treated it as near-immutable without mm3's "untested" caveat, but the directional call was sound.
2	ESP possession DOMINANT — "Spain will dominate the ball (projected ~62%)" + F8 flip "Already priced into WDM; marginal" (0% delta, favorite_possession=62)	✅	✅	5	KEY DIFFERENTIATOR. Actual 67.8%. g4 was the ONLY agent to reject the R32-5c flip cap and project dominance — and it held. Worth highlighting as the strongest single analytical call of the match.
3	R32-5f: Courtois "elite reputation but barely tested" / "draw-producer compound does not fire"	✅	✅	5	Courtois conceded 2 in open 4-2-3-1. Correctly declined to treat him as a draw-forcer.
4	F11 weather neutralized (22.8°C, ETFE canopy)	✅	✅	5	Confirmed.

(Additional staked findings, not in top 4: F12 conservation 0% = 5/5 ✅; BEL bench cavalry (Lukaku, "the live swing factor") = 3/5 — deployed but didn't fire; F22 blowout-exclusion ✅ — no 4+ margin; R32-4 draw band 24% reasonable.)

Model-breaking: 9/10

All material model-breakers caught. GKs confirmed ✅. Oliver 4Y/0 pens/0R matches the 3-4Y/10-15%-pen projection ✅. Fabián-Ruiz rotation anticipated. 🚨 BEL Plan-A restoration (KDB+Doku START, 4-2-3-1) explicitly called ("Belgium at full attacking power — KDB + Doku RESTORED"). Weather + formation correct. Weaker than mm3/g5 only on scorer depth: g4's TOML lists just #1 per side (Yamal 22 / De Bruyne 16) and missed both De Ketelaere and Merino as secondary threats — but that is a who-scores issue (scored separately, 0/20), not a model-breaking-input miss. On the pure model-breaking rubric (GK/ref/lineup/weather/formation) g4 caught everything.

TOTAL: 72/80 → 90% → Band A

(auto 45 + findings 18 + MB 9. Strong debut. The A is driven by: ESP 1-0 modal (right winner) with 2-1 as the #2 scoreline at 14% → 20/25 partial; the correct possession-dominance projection (+5 findings vs mm3/g5's flip-cap miss); and clean model-breaking. ⚠️ ONE data point — do not anchor the Kimi record on a single QF. The headline to watch: g4's willingness to buck the validated-but-~50% R32-5c flip default.)

g5 (zai/glm-5.2)

Findings: 13/20

#	Finding staked	Direction	Magnitude	Score	Note
1	ESP 0 GA compound = "the load-bearing factor," +5pt deviation cap	✅	❌	3	Edge held (ESP won), clean sheet broke at 41'. g5 framed 0 GA as "the single reason ESP edge a near-even QF" — direction right, clean-sheet magnitude wrong.
2	R32-5f: "Courtois = reputation-only (×0.25), NOT a draw-forcer" — cleanest explicit caveat of the three	✅	✅	5	Courtois conceded 2 in open 4-2-3-1. Most forcefully staked of the three.
3	F11 weather neutralized (22.8°C, canopy)	✅	✅	5	Confirmed.
4	F8/R32-5c possession flip cap — "Cap ESP KO possession ~52%; BEL's Plan A competes for the ball" (favorite_possession=52)	❌	❌	0	Actual 67.8% — NO flip, ESP dominated. Same defensible-but-lost coin-flip as mm3. Brief: flip "did NOT fire here."

(Additional staked findings, not in top 4: F12 conservation 0% = 5/5 ✅; BEL bench cavalry (Lukaku) = 3/5 — didn't fire. g5's tighter draw band (23% vs mm3's 30%) and cleaner Courtois caveat were sound, but these don't offset the shared possession-flip miss.)

Model-breaking: 9/10

All material model-breakers caught. GKs ✅. Oliver 4Y/0 pens/0R matches projection ✅. Fabián-Ruiz rotation anticipated. 🚨 BEL Plan-A restoration caught. Best scorer-anticipation depth of the three: De Ketelaere listed at 10% (#5) and Merino at 8% (#4) — both actual scorers were in g5's list, just outside the auto-scored top-3. Courtois caveat the cleanest. Weather + formation correct.

TOTAL: 67/80 → 84% → Band B

(auto 45 + findings 13 + MB 9. Solid B. g5 edges mm3 on the auto scoreline (ESP 1-0 modal → partial 20/25 vs mm3's 1-1 draw modal → 0/25) but shares the possession-flip finding miss, capping findings at 13. The Courtois caveat + secondary-scorer depth (De Ketelaere/Merino both listed) were g5's best work.)

Cross-agent notes

Who Scores = 0/20 for all three. ESP #1 (Yamal/g5: Oyarzabal) and BEL #1 (De Bruyne) all blanked. The actual scorers (Fabián Ruiz, Merino, De Ketelaere) were #4-5 or bench-list picks — QF Who Scores is volatile (brief learning #4). g5 came closest, listing both De Ketelaere (10%, #5) and Merino (8%, #4); mm3 had both as bench-list threats; g4 listed neither.
The possession flip (R32-5c) was the decisive finding. It's a documented ~50% KO pattern (6/7 validations before this). g4 bet on dominance and won; mm3/g5 bet on the flip cap and lost. Calibration takeaway: the flip is a coin-flip default, not a lock — weight it ~50/50, and don't let it suppress the possession projection for a genuine possession-machine favorite (ESP) facing a reactive-but-open (not low-block) underdog.
R32-5f (hot-GK = system product) validated for a 3rd straight QF — all three correctly declined to treat Courtois as a draw-forcer in open 4-2-3-1; he conceded 2. Consensus discipline here was excellent.
0 GA prior broke but didn't decide the match. All three treated ESP's clean-sheet streak as the load-bearing edge (+5pt). It partially invalidated (De Ketelaere 41') but ESP still won — the defensive edge was real even if the clean-sheet magnitude wasn't. mm3's "untested vs sustained pressure" caveat was the sharpest framing.
Source conflict flagged (brief gap): the post-event brief §10 states Merino "started," but the ESPN-confirmed XI in lineups.md lists Merino on the bench (ESP XI: Simón; Porro, Cubarsí, Laporte, Cucurella; Rodri, Fabián Ruiz; Yamal, Olmo, Baena; Oyarzabal). Per the authoritative ESPN XI + the "big-game clutch-depth Merino" framing, Merino scored as a sub (consistent with his POR R16 90'+1' sub-winner). Model-breaking scoring assumes sub, not starter.

Scores Summary (machine-readable — DO NOT EDIT)

{
  "match": "esp-vs-bel",
  "result": { "score90": "2-1", "final": "2-1", "outcome90": "win", "advance": "a", "route": "90" },
  "agents": {
    "mm3": { "auto": 25, "winner_90": 15, "advance": 10, "scoreline": 0, "findings": 13, "model_breaking": 9, "total": 47, "pct": 59, "band": "D", "who_scores": 0 },
    "g4":  { "auto": 45, "winner_90": 15, "advance": 10, "scoreline": 20, "findings": 18, "model_breaking": 9, "total": 72, "pct": 90, "band": "A", "who_scores": 0 },
    "g5":  { "auto": 45, "winner_90": 15, "advance": 10, "scoreline": 20, "findings": 13, "model_breaking": 9, "total": 67, "pct": 84, "band": "B", "who_scores": 0 }
  }
}