Per-Agent Calibration Scores: Brazil vs Norway

Result: 1-2 at 90' → 1-2 final | NOR win in 90' (advance route: 90')
Scorers: NOR Haaland 79', 90' (brace) / BRA Neymar 90'+10' (pen)
Headline: NOR upset — H2H curse holds (NOR unbeaten vs BRA in 5 meetings ever). BRA out-xG'd NOR 2.73 to 0.84 but lost (4 big chances missed; Haaland 2 from 0.12 xG = clinical).
Auto-scored baseline: winner_90 + advance + scoreline (mm3 10/50 · g4 30/50 · g5 30/50) | who_scores separate (mm3 10/20 · g4 12/20 · g5 12/20)

Scoring note. All three writers staked positions on the same four upset-driver findings, so they are scored on a comparable set: H2H curse, storm × NOR aerial, Haaland milestone (R32-5g), and F13 sterile-domination. The brief flags two clear calibration misses (H2H under-weighted; F13 should have fired) and two correct calls (storm directionally right; Haaland milestone = 6/6 in knockouts). Differentiation between agents comes from how aggressively each weighted the two NOR-edge factors (H2H + storm).

mm3

Findings: 11/20

#	Finding staked	Direction	Magnitude	Score	Note
1	H2H curse = −2pt to BRA ("NOR unbeaten ever; 1998 scar"; §6.5 synthesis row)	✅	❌	3	Recognised NOR's psychological edge — direction right. But −2pt vs brief's "should have been −5pt." Under-priced the load-bearing factor.
2	Storm × NOR aerial = "direction-neutral, 0pt shift" + wet-condition mismatch +2pt NOR (net +2pt NOR)	✅	❌	3	Net wet position (+2pt NOR) was direction-right and the wet mismatch materialised (Haaland header 79'). BUT mm3 explicitly staked the storm is "NOT directional / variance only" — wrong: the storm was directional and decisive. Magnitude under-called.
3	Haaland milestone (R32-5g) = NOR #1 at 32%, "record-chasing → goals, not minutes"	✅	✅	5	Bullseye. Haaland brace (79', 90'). Milestone rule now 6/6 in knockouts.
4	F13 sterile-domination = `finding_13_fires = false` ("no possession-dominant favorite likely")	❌	—	0	Direction wrong. BRA was possession-dominant (66%) and profligate (0 open-play goals from 2.73 xG). The textbook F13 outcome fired; mm3's screen ("no possession-dominant favorite") missed that BRA dominates territory by design. Brief: "F13 should have fired more strongly."

Model-breaking: 5/10

Partial. Weather caught (direction right via +2pt wet mismatch; but mm3 denied storm directionality — see finding 2). Haaland milestone caught (NOR #1). GK / ref / lineup all correctly anticipated — no GK change (Alisson + Nyland both confirmed), 1 pen within Elfath's 10–15% band, exact-XI match to projection. MISS: BRA profligacy / NOR clinical finishing — the xG reversal (2.73 vs 0.84 → 1-2) is the story of the match and no agent anticipated it; mm3 projected BRA xG 1.5 (actual 2.73 — BRA created more than expected and still lost). The decisive model-breaker (finishing variance) was unanticipated.

TOTAL: 26 / 80 → 33% → Band F

(auto 10 + findings 11 + model-breaking 5)

g4

Findings: 13/20

#	Finding staked	Direction	Magnitude	Score	Note
1	H2H curse = −3pt to BRA ("H2H hex: NOR unbeaten vs BRA ever"; §6.5)	✅	❌	3	Direction right (most aggressive H2H weight of the three). Still under-priced vs brief's −5pt target, but closest of the trio.
2	Storm × NOR aerial = +3pt NOR ("SIGNIFICANT; rain favours NOR's aerial/set-piece; Ødegaard → Haaland/Ajer/Heggem")	✅	✅	5	Best call of the three. g4 weighted the storm the most (+3pt) and the mechanism materialised exactly as staked (Haaland aerial header 79'). Brief: "storm directionally right."
3	Haaland milestone (R32-5g) = NOR #1 at 30%, "record-chaser → goals expected, not just minutes"	✅	✅	5	Bullseye. Brace.
4	F13 sterile-domination = `finding_13_fires = false` ("NOT flagged; BRA has multiple finishers (Vini Jr + Cunha + Martinelli)")	❌	—	0	Direction wrong. g4's reasoning ("multiple finishers → no sterile domination") misread the issue: the problem was poor finishing on the day, not lack of finishers. BRA scored 0 from open play despite 2.73 xG.

Model-breaking: 5/10

Partial. Weather caught best — g4's +3pt storm weight was the strongest of the three and the swing factor materialised (Haaland aerial in the rain). Haaland milestone caught. GK / ref / lineup correctly anticipated (no changes, no surprises, 1 pen in-band). MISS: BRA profligacy / NOR clinical finishing — same blind spot as the others; g4 projected BRA xG 1.6 (actual 2.73) and staked F13 N/A, so the xG→goals reversal was unanticipated. g4 is the strongest of the three on model-breaking direction (weather) but shares the finishing-variance miss.

TOTAL: 48 / 80 → 60% → Band C

(auto 30 + findings 13 + model-breaking 5)

g5

Findings: 8/20

#	Finding staked	Direction	Magnitude	Score	Note
1	H2H curse = 0pt ("Soft; mechanism already encoded above" — §6.5 synthesis row, all three columns 0/0/0)	❌	—	0	Direction wrong. g5 explicitly dismissed the H2H curse as non-load-bearing, choosing to encode the tactical mechanism in the storm + Haaland rows instead. The brief contradicts this sharply: "H2H curse is REAL and load-bearing… should have been −5pt." g5 staked it doesn't apply; it did. This is the most distinctive analytical miss of the match.
2	Storm × NOR aerial = +1pt NOR ("I weight it modestly because both teams play in the rain — it is a real edge, not a decisive one")	✅	❌	3	Direction right (storm favours NOR aerial). Magnitude wrong — g5 explicitly hedged "not decisive"; the storm was decisive (produced the 79' header, the swing goal). Under-weighted.
3	Haaland milestone (R32-5g) = NOR #1 at 32%, "record-chasing (Golden Boot, Norway WC scoring record) → goals, not just minutes"	✅	✅	5	Bullseye. Brace.
4	F13 sterile-domination = `finding_13_fires = false` ("BRA attacking, not sterile-dominating; NOR efficient counter")	❌	—	0	Direction wrong. Same shared miss: BRA was sterile in outcome (0 open-play goals from 2.73 xG). g5's framing ("BRA attacking") confused creation with conversion.

Model-breaking: 5/10

Partial. Weather caught weakly — direction right (+1pt NOR) but g5 explicitly called it "not decisive," under-calling the match's swing factor. Haaland milestone caught. GK / ref / lineup correctly anticipated (g5 even correctly resolved the Ryerson T-51 override of the T-3.5h injury report). MISS: BRA profligacy / NOR clinical finishing — same xG-reversal blind spot (projected BRA xG 1.6 vs actual 2.73). g5 is the weakest of the three on the weather catch (magnitude) but otherwise on par.

TOTAL: 43 / 80 → 54% → Band D

(auto 30 + findings 8 + model-breaking 5)

Cross-agent notes

The two shared misses (H2H magnitude + F13 direction) cap everyone below Band B. No agent anticipated BRA scoring 0 from open play on 2.73 xG — the finishing-variance reversal is a genuine model-limitation, not a writer failure per se, but it is the material model-breaker and went uncaught.
Differentiation is real and honest. g4 (13/20) was the most alert to NOR's edges (storm +3pt, H2H −3pt — both the most aggressive). g5 (8/20) was the most dismissive (H2H 0pt, storm +1pt "not decisive"). mm3 (11/20) sat between — recognised both NOR edges but hedged the storm's directionality ("direction-neutral"). The 5pt findings spread mirrors genuine analytical disagreement, not correlation.
Don't double-penalize the outcome. All three modal-picked a BRA win (wrong) and the auto-script already docked winner_90 (10/15 each) + advance (0/10 each). The findings scores above judge reasoning quality on the staked factors, separate from the wrong-side outcome — e.g., g4's storm call was excellent analysis even though g4 still landed on the wrong winner.
Haaland milestone = 6/6 in knockouts (Messi, Ronaldo, Mbappé, Kane, Haaland ×2). All three writers caught it as NOR #1 at 30–32%. This is now the most reliable scorer-prediction rule in the dataset.

Scores Summary (machine-readable — DO NOT EDIT)

{
  "match": "bra-vs-nor",
  "result": { "score90": "1-2", "final": "1-2", "outcome90": "loss", "advance": "b", "route": "90" },
  "agents": {
    "mm3": { "auto": 50, "winner_90": 10, "advance": 0, "scoreline": 0, "findings": 11, "model_breaking": 5, "total": 26, "pct": 33, "band": "F", "who_scores": 10 },
    "g4":  { "auto": 50, "winner_90": 10, "advance": 0, "scoreline": 20, "findings": 13, "model_breaking": 5, "total": 48, "pct": 60, "band": "C", "who_scores": 12 },
    "g5":  { "auto": 50, "winner_90": 10, "advance": 0, "scoreline": 20, "findings": 8, "model_breaking": 5, "total": 43, "pct": 54, "band": "D", "who_scores": 12 }
  }
}