Per-Agent Calibration Scores: Portugal vs Spain

Result: POR 0-1 ESP (90') → ESP advance in 90' (Merino 90'+1' sub winner) → ESP to QF
Auto-scored baseline: winner_90 + advance + scoreline, tracked per-agent below | who_scores 0/20 for all (separate)
Match shape: Predicted 50/50 coin-flip, modal 1-1 (unanimous). ESP won 1-0 in 90' via a 90'+1' sub winner. The 0-0 draw shape held for 90 minutes, broken at the death. ESP xG 1.77 vs POR 0.60 — ESP were the better side.

mm3

Findings: 20/20

Note (minor): mm3 also added a "+1pt POR milestone amplifier" nudge in the synthesis table — a small probability lean that did NOT fire (Ronaldo scored 0). This is a minor wrong probability nudge, not a core staked finding; the modal POR-win outcome it contributed to is already docked in winner_90 (5/15). Not double-counted here.

Model-breaking: 10/10

All structural model-breakers caught and confirmed pre-match: ✅ GK (Costa + Simón both confirmed, no change), ✅ referee (Taylor pen band 15–25%, correctly regressed HIGH-WIDE → 0 pens, within band), ✅ weather (roof closed, F11 neutralized), ✅ lineup (Félix-over-Leão noted, accounted in Who Scores). No unanticipated structural break occurred. The decisive event (Merino 90'+1' sub) is a Finding 19 late-sub tail — anticipated as a category (mm3 flagged Ramos super-sub + Williams-sub, F18/F19) but Merino himself wasn't named; this is a Who Scores tail miss (scored separately, 0/20), NOT a structural model-breaker. Not docked here.

TOTAL: 5 (auto) + 20 (findings) + 10 (MB) = 35 / 80 → 44% → Band F

Honest read: mm3 had the strongest qualitative reasoning of the three (findings 20/20 — the two marquee findings of the match, Simón-real + F16-Ronaldo, both validated by the brief). The F-band is driven entirely by outcome luck: the modal POR-win (40%) + 1-1 scoreline missed, and the TOML scoreline list didn't contain the exact "0-1" string the auto-matcher needed (so scoreline 0 vs g5's 20 despite near-identical §7K tables). The findings layer correctly distinguishes mm3's sound reasoning from the unlucky outcome.

g4

Findings: 20/20

Bonus context: g4 also staked F4a "single-absence, NOT gutting" for ESP wings (Williams+Pino out, −2pt only) — held (ESP won, spine intact, scored). g4 was the most ESP-leaning agent on every material call (Simón-real credit +1 ESP, finishing-gate against POR, momentum +2 ESP), which is why it was the only agent to get winner + advance + scoreline-margin right.

Model-breaking: 10/10

All structural model-breakers caught: ✅ GK (Costa + Simón confirmed, no change), ✅ referee (pen band 20–25% → 0 pens, within band), ✅ weather (roof closed), ✅ lineup (Félix-over-Leão explicitly profiled in §1 + accounted). g4's Match Flow §9 even flagged the 75–90' window as "high drama... late subs... golden goal" — the exact window Merino scored in — though Merino wasn't named (Who Scores tail, scored separately). No structural break missed.

TOTAL: 40 (auto) + 20 (findings) + 10 (MB) = 70 / 80 → 88% → Band B

Honest read: g4 was the clear winner on this match — the only agent to land winner + advance + scoreline-margin (auto 40/50, top of the three). Its qualitative reasoning was equally strong (findings 20/20), with the decisive edge being that it gave ESP real credit for the 0 GA block (×0.25 open-shape hot-GK → +1% ESP) and applied the POR finishing-quality gate. This is g4's documented strength profile firing: "best on clear favorites + scoreline precision" — here, best on correctly identifying the marginal ESP edge in a coin-flip.

g5

Findings: 10/20

Note: g5's two marquee findings (the very ones it leaned on hardest, repeated 5× each) were BOTH wrong, and they're exactly the two findings the brief calls out as validated learnings. g5 framed this match as "the 0-GA breaks vs first real attack" — the defining prediction was falsified. g5 also staked "sterile-domination risk returns" (ESP attack) — didn't fire (ESP xG 1.77, scored 1).

Model-breaking: 10/10

Structural model-breakers caught: ✅ GK (Costa + Simón confirmed, no change), ✅ referee (pen band 20–28% → 0 pens, within band), ✅ weather (roof closed), ✅ lineup (Félix LW noted). No real structural break missed — the model-breaking category scores whether real breakers were CAUGHT, and none materialized. Important nuance: g5's headline "model-breaker" flag — the ESP 0 GA as a breakable ENG-MEX-style record — was a FALSE POSITIVE (g5 over-flagged a break that didn't happen), not a missed breaker. That reasoning error is scored under findings (#3, 0pts) to avoid double-counting; it does not dock model-breaking, which g5 otherwise completed cleanly (no real breaker unanticipated). g5 did NOT name Merino (Who Scores tail, scored separately).

TOTAL: 30 (auto) + 10 (findings) + 10 (MB) = 50 / 80 → 63% → Band C

Honest read: g5's Band C is propped up by outcome luck — it alone listed the exact "0-1" scoreline string (scoreline 20/25) and had loss as #2 at 35% (winner_90 10/15) — despite staking the WRONG side on both marquee findings. The findings layer (10/20, the lowest of the three) correctly captures that g5's core reasoning was wrong on the two questions that defined this match. g5 is the project's default + "best model-breaker detector," but here its signature call (ESP 0 GA opponent-inflated) was the match's biggest analytical miss. The auto-score's generosity (exact scoreline) masks a genuinely weaker reasoning performance — which is exactly what the findings layer exists to expose.

Cross-agent summary

Key takeaways:

1. g4 was the right agent for this match type (marginal favorite identification in a coin-flip) — it gave ESP real 0-GA credit + applied the POR finishing gate, landing winner + advance + scoreline-margin. Its documented "favorite precision" strength fired.
2. The findings layer did its job. mm3 (auto 5, Band F) and g5 (auto 30, Band C) had inverted reasoning-vs-outcome luck: mm3's reasoning (findings 20/20) was far stronger than g5's (findings 10/20) — both got the marquee findings (Simón-real + F16-Ronaldo) that g5 got wrong — but g5's exact-scoreline luck pulled its total above mm3's. The findings column makes the reasoning-quality gap legible.
3. g5's signature call failed. The "ESP 0 GA opponent-inflated / ENG-MEX warning" — staked 5× — was the match's biggest analytical miss and the direct cause of g5's POR-win modal. Brief learning #1 is a direct falsification.
4. Merino (90'+1') beat all three — a Finding 19 late-sub winner, anticipated as a category (late-goal scenarios flagged by all) but not named (all listed Ramos/Williams/Ferran, not Merino). who_scores 0/20 across the board.
5. No structural model-breakers materialized (GK confirmed, ref as profiled, roof closed, lineup as projected) — model-breaking 10/10 for all three; the decisive event was a tactical late sub, not a structural break.

Scores Summary (machine-readable — DO NOT EDIT)

{
  "match": "por-vs-esp",
  "result": { "score90": "0-1", "final": "0-1", "outcome90": "loss", "advance": "b", "route": "90" },
  "agents": {
    "mm3": { "auto": 5, "winner_90": 5, "advance": 0, "scoreline": 0, "findings": 20, "model_breaking": 10, "total": 35, "pct": 44, "band": "F", "who_scores": 0 },
    "g4":  { "auto": 40, "winner_90": 15, "advance": 10, "scoreline": 15, "findings": 20, "model_breaking": 10, "total": 70, "pct": 88, "band": "B", "who_scores": 0 },
    "g5":  { "auto": 30, "winner_90": 10, "advance": 0, "scoreline": 20, "findings": 10, "model_breaking": 10, "total": 50, "pct": 63, "band": "C", "who_scores": 0 }
  }
}