Per-Agent Calibration Scores: Norway vs England

Result: 1-1 at 90' → ENG 2-1 AET (Bellingham 45'+2', 93' brace / Schjelderup 36' Ødegaard assist). ENG advance to SF2 vs Argentina.
Stage: QF | Venue: Hard Rock Stadium, Miami (open-air, fixed canopy) | Ref: Clément Turpin
Weather (actual): 34.4°C — F11 heat APPLIED ✅ (as predicted by all 3 agents)

Auto-scored baseline (winner_90 15 + advance 10 + scoreline 25 = 50):

mm3: 30/50 | who_scores 5/20 (separate)
g4: 37/50 | who_scores 0/20 (separate)
g5: 37/50 | who_scores 5/20 (separate)

mm3

Findings: 15/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F11 heat FULLY APPLIES (>33°C, R32-5a) + asymmetric pressing degradation	✅	✅	5	34.4°C confirmed in brief; pressing compressed (NOR 4 SoT, ENG 8 SoT — modest xG); no possession flip (ENG only 52.4%); heat degraded ENG press exactly as staked
2	Nyland ×0.5-0.6 confirmed-form + compact block = draw compound (R32-5f)	✅	✅	5	Block+GK held for 90', forced the 1-1 draw (mm3's modal); conceded 1 before HT + broke in ET — textbook R32-5f "90-min only" system product
3	R32-5g milestone → both #9s (Haaland + Kane) score	❌	❌	0	Both #1 picks BLANKED. Brief flags this as a new boundary: R32-5g breaks vs elite Tier-1 defense. Shared miss with g4/g5
4	Coin-flip framing / fav_share 0.642, ENG NOT a runaway favorite	✅	✅	5	1-1 at 90', decided only in ET (93') — genuine coin-flip. Rejecting the "clear favorite" tag was correct

Model-breaking: 7/10

Strong catch of the pre-match material factors: F11 weather flagged prominently (the match's biggest swing — validated); both lineup changes explicitly noted and adjusted (Step 6b Schjelderup, 6c Stones-at-RB); lightning-suspension window flagged as a genuine model-breaker consideration (30% PoP); GK correctly handled (no change — both confirmed). Hedged the lineup-change scorer by listing Schjelderup at 7% (NOR #4). Missed: did not flag the risk that both milestone #1 picks could blank (staked "both likely scorers"), nor cite the emerging "lineup-change scorer" QF pattern (3rd-straight). Partial-plus.

⚠️ PROCESS FLAG (advance label bug): mm3's [knockout] TOML lists advance_a = 51 # ENG and advance_b = 49 # NOR — it swapped the team convention. Per the nor-vs-eng code, team A = NOR, team B = ENG, so the auto-scorer read mm3's advance as NOR (advance_a 51 > advance_b 49) and scored 0/10. mm3's intended call (per TL;DR "Advance ~51% ENG" and the comment) was ENG advance — i.e., a correct team call mechanically mislabeled. This cost ~7 pts and is a TOML-execution error, not a reasoning miss. Reflected as-is (0) in the JSON footer per the auto-scorer contract.

TOTAL: 30 (auto) + 15 (findings) + 7 (MB) = 52/80 → 65% → Band C

(who_scores 5/20 tracked separately — Bellingham #2 pick 12% scored the brace)

g4

Findings: 15/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F11 heat equalizer (compresses pressing after 60', widens draw band)	✅	✅	5	34.4°C applied; pressing compressed; the draw DID fire at 90' — g4's "heat = equalizer" was the right directional read
2	Nyland confirmed-form hot-GK (×0.25 empirical → +3-4% draw band)	✅	✅	5	Nyland held 90' behind compact block; draw band (g4 staked 30%, the widest of the 3 agents) fired correctly. ×0.25 scale not under-priced here
3	F12 = 0 / milestones override conservation (both #9s play full minutes)	✅	✅	5	Both #9s started & played; F12 = 0 confirmed in brief. Conservation-override call held
4	R32-5g milestone → both #9s (Haaland + Kane) score	❌	❌	0	Both blanked; R32-5g boundary vs Tier-1 defense (shared miss)

Model-breaking: 6/10

Caught the headline factors: F11 weather (TL;DR "heat is the equalizer"), lineup changes noted (Schjelderup "first tournament start replacing Nusa"; Stones makeshift-RB depth crisis), GK correctly weighted (no change). Hedged the lineup-change scorer (Schjelderup 5% NOR #4) and the secondary ENG scorer (Bellingham 14% — the actual match-winner). Flagged RB depth crisis + sterile-domination risk. Missed: did not flag the both-#1-blank risk (staked both as scorers), nor the lineup-change-scorer QF pattern. Less distinctive model-breaker framing than mm3/g5 — g4's "Watch" was the heat-wall/Haaland-transition, both of which were weather/finding items rather than true breakers. Partial.

⚠️ PROCESS FLAG (who_scores TOML truncation): g4's [who_scores] TOML lists only team_a_1=Haaland and team_b_1=Kane — the full prose table (which DID include Bellingham #2 14%, Madueke #3 10%, Schjelderup #4 5%) was not serialized into the TOML. The auto-scorer reads the TOML, so g4 scored 0/20 on who_scores despite hedging Bellingham correctly in prose. TOML-serialization gap; noted, reflected as-is (0).

TOTAL: 37 (auto) + 15 (findings) + 6 (MB) = 58/80 → 73% → Band C

(who_scores 0/20 — TOML truncation; prose had Bellingham #2)

g5

Findings: 15/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F11 heat fully applies (>33°C, asymmetric — degrades ENG pressing, favors NOR counter)	✅	✅	5	34.4°C applied; ENG couldn't dominate (52.4% poss, no flip); modest xG. Asymmetric read correct
2	Nyland confirmed-form hot-GK + MODERATE block (not a 5-4-1 bus) → "partially amplifies" draw compound (R32-5f, ×0.5)	✅	✅	5	Sharpest GK read of the three: "partially amplifies" was EXACTLY right — block held 90' (forced 1-1), then broke in ET (Bellingham 93'). Moderate-not-bus nuance predicted the ET collapse
3	R32-5g milestone → both #9s (Haaland + Kane) score	❌	❌	0	Both blanked; R32-5g boundary vs Tier-1 defense (shared miss)
4	ENG attacking depth tells in ET → advance B (57%)	✅	✅	5	ENG won in ET via Bellingham 93' — g5's "ENG depth is the better ET chance-creation unit; one moment more likely from them" was the exact mechanism that materialized

Model-breaking: 6/10

Nailed its lane's headline call: explicitly labeled Nyland as "Model-breaker" (the crux) and the "moderate block, not a bus" nuance — the match-deciding read (Nyland drew 90', broke ET). Caught F11 weather + the moderate-block system-product framing. Missed: the lineup-change scorer — g5's NOR scorers were Haaland/Ødegaard/Sørloth/CB-set-piece/Berge; Schjelderup was NOT listed at all (the actual NOR goalscorer, a genuine miss for a model-breaker lane, and a contrast to mm3/g4 who hedged him). Also staked both #1 picks as sure scorers (missed the both-#1-blank risk). So strong on the GK/block finding-detection but the lineup-surprise scorer (a true model-breaker) slipped through. Partial.

TOTAL: 37 (auto) + 15 (findings) + 6 (MB) = 58/80 → 73% → Band C

(who_scores 5/20 — Bellingham #2 16% scored the brace)

Cross-agent notes

All 3 shared one finding miss: R32-5g "milestone → goals" applied to both #9s (Haaland + Kane). Both blanked. The brief identifies this as a new boundary: R32-5g breaks when a record-chaser meets an elite Tier-1 defense (Konsa/Guéhi held Haaland; NOR block held Kane). This is a framework learning, not a per-agent error — hence 0/5 each, but no extra demerit.
The win was in the draw band + ET depth: mm3 (modal 1-1, 25% draw) and g5 (modal 1-1, 25% draw) hit the exact 90' scoreline; g4 staked the widest draw band (30%) and the most aggressive ENG-bull (68% advance). All three correctly identified that heat + Nyland/block → draw at 90' → ENG edge in ET. g4's modal-90' was ENG-win (wrong) but its final-scoreline path (ENG 2-1, 17% modal) was the eventual final.
Who Scores was the shared weak spot: 0/3 #1 picks scored (Haaland ❌, Kane ❌). The actual scorers (Schjelderup, Bellingham) were secondary/deeper — mm3 + g4 + g5 all had Bellingham as ENG #2; only mm3 + g4 hedged Schjelderup (g5 missed him). g4's TOML truncation turned a prose-hedge into a 0/20.
No classic model-breaker materialized: no GK change (both confirmed), no formation shift, Turpin within range (1 Y, 0 pens vs 2-3/25-35% projected). The weather (F11) was the only material swing, and all 3 caught it — so model-breaking scores cluster at 6-7 (good, not perfect, on the available factors).
mm3's advance label bug is the single biggest per-agent anomaly: a correct reasoning call (ENG advance) mechanically mislabeled in the TOML, costing 7 auto-pts and dropping mm3 from a would-be 59/80 (74%) to 52/80 (65%). Flagged for calibration but scored as the auto-scorer reads.

Scores Summary (machine-readable — DO NOT EDIT)

{
  "match": "nor-vs-eng",
  "result": { "score90": "1-1", "final": "1-2", "outcome90": "draw", "advance": "b", "route": "aet" },
  "agents": {
    "mm3": { "auto": 30, "winner_90": 5, "advance": 0, "scoreline": 25, "findings": 15, "model_breaking": 7, "total": 52, "pct": 65, "band": "C", "who_scores": 5 },
    "g4":  { "auto": 37, "winner_90": 10, "advance": 7, "scoreline": 20, "findings": 15, "model_breaking": 6, "total": 58, "pct": 73, "band": "C", "who_scores": 0 },
    "g5":  { "auto": 37, "winner_90": 5, "advance": 7, "scoreline": 25, "findings": 15, "model_breaking": 6, "total": 58, "pct": 73, "band": "C", "who_scores": 5 }
  }
}