Per-Agent Calibration Scores: England vs France

Result: ENG 6-4 FRA (90') → ENG win bronze in 90' (no ET, no pens)
Auto-scored baseline: winner_90 + advance + scoreline = varies (see per-agent) | who_scores tracked separately

🚨 CALAMITOUS ENSEMBLE MISS. All three writers were UNANIMOUS: FRA 2-1, FRA advance ~55%. Actual: ENG 6-4 — wrong winner (all 3), wrong scoreline by a factor of ~3× (10 goals vs modal 3). The two theses that carried every prediction — "Mbappé retained = decisive edge" and "ENG attack gutted (F4a)" — both INVERTED. The 3rd-place context produced the opposite of the modeled dynamic: the deflated SF-loser (FRA) underperformed while ENG's hungry rotated backups (Saka ×2, Rice, Konsa) + a benched-star sub winner (Bellingham 90'+8') ran riot. Heat (F11, 30-32°C) fired 3-4× harder than modeled.

⚠️ SCORELINE AUTO-SCORE IS A PARSER ARTIFACT — read this before interpreting the bands.

g4 / g5 scoreline = 15/25 is spurious. Their unlabeled "2-1" (intended = France 2-1, FRA win) is regex-parsed by the validator as teamA(England) 2-1 → "win" → matches the actual ENG win → 15. g4 even labels "2-1_ENG" separately, proving unlabeled "2-1" meant FRA. The validator cannot read that intent.

mm3 scoreline = 0/25 is spurious for the opposite reason. mm3 used a capitalized "Other" key (18%), and the validator's scoreline filter only drops lowercase "other" (k !== "other"), so "Other" survived as the modal → no digit-match → 0. mm3's top specific scoreline was "FRA 2-1" at 12% — the identical prediction g4/g5 made.

Artifact-corrected "true" picture: all three modal-predicted FRA 2-1 with the WRONG winner (FRA) and WRONG margin (1 vs actual 2). Honest scoreline for all three = 0/25 → auto = 10/50 → all three ≈ 27-28/80 ≈ 34% → all Band F. The mm3(F)/g4(D)/g5(D) split below is purely a parser artifact, not a quality difference. The findings + model-breaking scores (which I control) are identical/symmetric across all three because the reasoning was genuinely uniform.

mm3

Findings: 11/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F4a underdog-side INVERSE — ENG gutted (Kane+Bellingham+Pickford) → FRA gains ~3pt (§6.5 Step 6)	❌	❌	0	The most sophisticated gutting framing of the three (mm3 recognized F4a is usually favorite-side and tried to invert it) — but direction WRONG. ENG's backups scored 6 and WON. Brief learning #4: "Rotation gutting ≠ F4a… backups out-performed stars."
2	F11 heat fully applies — both Tier-2 (rotated) → full penalty; open-game force, no draw compound (§6.5 Steps 3-4)	✅	❌	3	Direction clean: 30-32°C confirmed, no draw (ENG won in 90'), open game. But magnitude catastrophically under-modeled — mm3 projected a "2-3 goal scenario" / "3+ goal finale"; actual 10 goals (~3-4×). mm3 was the closest to anticipating a high-scoring finale, but still 3× short.
3	F12 = N/A, 0% discount — dead rubber, no conservation (§6.5 Step 2)	✅	✅	5	Confirmed: brief F12 = N/A (dead rubber, no conservation). Clean.
4	Henderson GK = SWE-negative / ENG defensive liability — untested (0 KO min), capped confidence (§6.5 Step 6 + F14)	✅	❌	3	Direction RIGHT — Henderson conceded 4 (liability confirmed). Magnitude wrong: the liability was symmetric (Maignan also conceded 6) and didn't flip the match to FRA; ENG outscored it. "ENG-only liability" thesis was half the story.

Model-breaking: 6/10

Flagged the 🚨 MODEL-BREAKING GK CHANGE (Pickford→Henderson) as the headline — exactly the model-breaker lineups.md flagged with 🚨. Caught the 8/11 ENG + 6/11 FRA rotation fully. Staked F11 heat applies fully. Predicted 1-3 yellows vs actual 0 — direction right on low intensity, but the extreme "most lenient match of the tournament" (dead rubber = no tension) was under-modeled. Did NOT anticipate the 3rd-place motivation inversion (a reasoning-direction error, scored in finding #1) nor the 10-goal magnitude. The catastrophe was a direction/thesis failure, not a detection failure — all structural breakers (GK, lineup, weather) were caught.

TOTAL: 27 / 80 → 34% → Band F

(auto 10 = w90 10 + adv 0 + scor 0[artifact] + findings 11 + MB 6) · who_scores 12/20 (Mbappé #1 FRA ✓ brace + Saka #2 ENG ✓; Rogers #1 ENG ✗). Artifact-corrected "true" band = F (uniform with g4/g5).

g4

Findings: 11/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F4a fires for England — gutted creative spine (Kane main striker + Bellingham main creator + Pickford GK) + known-unknown in goal	❌	❌	0	Standard gutting framing — direction WRONG. ENG's "gutted" attack scored 6. Brief learning #4 explicit: the F4a framework does not apply to voluntary 3rd-place rotation.
2	F11 heat applies — Hard Rock open-air, heat-index 33-38°C; pressing degrades after 60′	✅	❌	3	Direction right (heat fired, pressing collapsed). Magnitude badly wrong: projected_xg 1.4 / modal 2-1 (≈3 goals) vs actual 10. ~3× under.
3	3rd-place = open/goals — lower draw band (~20%), wider scoreline bands, no clean sheet, profligacy > clean sheet	✅	❌	3	Direction right (open game, no draw, no clean sheet for either). But g4 used this to lock a 2-1 modal — catastrophically under-weighting the goal TOTAL despite correctly calling "open." Left 19% on "other" but modal stayed low.
4	F12 = 0 — conservation_discount = 0 (dead rubber)	✅	✅	5	Confirmed. Clean.

Model-breaking: 5/10

Caught the GK change but bundled it into F4a ("known-unknown in goal") rather than flagging it as a discrete model-breaker — weaker model-breaker emphasis than mm3/g5. Caught the rotation + heat (TL;DR "Heat applies"). Predicted 1-3 yellows vs actual 0. Same structural detection as the others, but the most generic framing — no explicit model-breaker callout, and the "goalkeeping gap = swing variable" thesis swung the wrong way (it was symmetric, not FRA-favoring).

TOTAL: 41 / 80 → 51% → Band D

(auto 25 = w90 10 + adv 0 + scor 15[artifact] + findings 11 + MB 5) · who_scores 10/20 (Mbappé #1 FRA ✓ brace; Gordon #1 ENG ✗). ⚠️ The D band is entirely the spurious scoreline=15 (parser read "2-1" as ENG 2-1). Artifact-corrected "true" total = 26/80 = 33% = Band F (uniform with mm3/g5).

g5

Findings: 11/20

#	Finding staked	Direction	Magnitude	Score	Note
1	"Model-breakers STACK FOR FRANCE; motivation asymmetry IS the match" — Deschamps retained Mbappé/Olise/Doué/Rabiot (wants to win); Tuchel benched stars (development fixture)	❌	❌	0	The most explicitly-committed central thesis of the three — and the most decisively INVERTED. Brief learning #1: "3rd-place context INVERTS the retained-stars thesis." The asymmetry cut the other way: ENG's hungry backups wanted it more; FRA's retained stars (fresh off losing a final spot) underperformed.
2	F11 heat applies (MEDIUM confidence — flagged missing NWS forecast) — pressing degrades, FRA amplifier	✅	❌	3	Direction right (heat fired). The honest "MEDIUM confidence / no archived NWS" flag was a transparency plus — but the magnitude was 3× under (actual 10 goals), and it amplified ENG as much as FRA, not FRA-only.
3	ENG GK change = SWE-negative (R32-5f), not draw-forcer — tilts xG-conversion toward FRA	✅	❌	3	Direction right (Henderson conceded 4 — SWE-negative confirmed) and the cleanest GK framing of the three (correctly rejected the draw-forcer read). Magnitude wrong: the tilt was symmetric (Maignan conceded 6); didn't produce a FRA win.
4	F12 = 0 — conservation_discount = 0 (dead rubber)	✅	✅	5	Confirmed. Clean.

Note: g5 also explicitly staked F22 blowout EXCLUSION ("no pressing-side opener… blowout tail ~12% IF Henderson collapses"). Direction defensible on MARGIN — ENG 6-4 is a 2-goal margin, not a 4+ blowout — so F22-as-margin didn't fire. But g5 used the exclusion to cap scorelines, contributing to the 10-goal under-weight; and the Henderson-collapse conditional it down-weighted actually fired (conceded 4). Not separately scored (5-slot cap); noted as a near-miss that leaned the wrong way on goal-total.

Model-breaking: 6/10

Model-breaker detection lane — flagged the GK change prominently (SWE-negative, R32-5f citation), caught the rotation asymmetry, staked F11 heat, and honestly flagged the MEDIUM-confidence weather gap (missing NWS forecast) rather than bluffing. Predicted 2-3 yellows vs actual 0. Detection of structural breakers was strong (g5's specialty). BUT the model-breaker interpretation — "stack for FRA" — was inverted (scored in finding #1), and the F22 blowout exclusion under-weighted the high-total tail. The decisive material outcome (ENG rout) was unanticipated, but the headline breakers (GK, rotation, weather) were all caught.

TOTAL: 42 / 80 → 53% → Band D

(auto 25 = w90 10 + adv 0 + scor 15[artifact] + findings 11 + MB 6) · who_scores 10/20 (Mbappé #1 FRA ✓ brace; Gordon #1 ENG ✗, Rogers #2 ENG ✗). ⚠️ The D band is entirely the spurious scoreline=15 (parser read "2-1" as ENG 2-1). Artifact-corrected "true" total = 27/80 = 34% = Band F (uniform with mm3/g4).

Cross-agent notes

All three are genuinely at the SAME analytical level. Identical modal winner (FRA), identical modal scoreline (FRA 2-1), identical advance (FRA ~55%), identical miss. The findings (11/20) and near-identical model-breaking (5-6/10) scores reflect this uniformity. The D/F band split in the JSON is 100% a scoreline-parser artifact (g4/g5 unlabeled "2-1" read as ENG 2-1; mm3 capitalized "Other" not filtered). Artifact-corrected, all three = Band F (~34%).
The two load-bearing theses both inverted: (1) "Mbappé retained = decisive edge" — Mbappé scored a brace but FRA lost 4-6; ENG's distributed backup attack (Saka ×2, Rice, Konsa, Bellingham sub) was MORE clinical. (2) "ENG attack gutted (F4a)" — the "gutted" XI scored 6; brief learning #4 is explicit that F4a does not apply to voluntary 3rd-place rotation.
F11 heat: universal direction-right, universal magnitude-fail. All three staked heat applies (30-32°C confirmed) → open game, no draw. All three under-weighted the goal total ~3× (modal 2-1 / xG 1.4-1.7 vs actual 10 goals). The "3rd-place = more open, more goals" pattern fired 3-4× harder than any agent modeled — the draw band was correctly cut (no draw), but the goal TOTAL was never up-weighted to the 30-40% it warranted.
GK liability was symmetric, not FRA-favoring. All three correctly read Henderson as a liability (SWE-negative) — and he DID concede 4 (direction right). But the same dead-rubber defensive collapse hit Maignan too (conceded 6). The "ENG-only liability tilts to FRA" framing was half the story; in a 3rd-place shootout BOTH GKs concede. Brief learning #3.
Mbappé #1 FRA scorer: unanimous HIT (brace, 48'+66'). But the ENG modal scorers (Rogers/Gordon — neither started prominently / neither scored) all missed. Actual ENG scorers: Rice, Konsa (CB!), Saka ×2, Bellingham (sub) — none was any agent's ENG #1. Only mm3 had Saka in its list (ENG #2, 12%) → the lone extra who_scores hit.
Ref: universal partial miss. All three predicted 1-3 (mm3/g4) or 2-3 (g5) yellows → 0 cards. Valenzuela's leniency was extreme in a tension-free dead rubber (FRA 14 fouls, ENG 8). Direction right (low), magnitude off (zero).
Process integrity held. This was flagged pre-match as "the lowest-confidence pre-match prediction of the tournament" (lineups.md); honest spreads were wide-ish (36-39% ENG). The framework correctly signaled uncertainty — but uncertainty is not prescience, and a 10-goal ENG rout lived in a tail none of the three weighted. This is the worst ensemble miss of the tournament and a clean demonstration that the 3rd-place fixture needs its own context rule (motivation inversion + goal-total up-weight) for future cycles.

Scores Summary (machine-readable — DO NOT EDIT)

{
  "match": "eng-vs-fra",
  "result": { "score90": "6-4", "final": "6-4", "outcome90": "win", "advance": "a", "route": "90" },
  "agents": {
    "mm3": { "auto": 10, "winner_90": 10, "advance": 0, "scoreline": 0, "findings": 11, "model_breaking": 6, "total": 27, "pct": 34, "band": "F", "who_scores": 12 },
    "g4":  { "auto": 25, "winner_90": 10, "advance": 0, "scoreline": 15, "findings": 11, "model_breaking": 5, "total": 41, "pct": 51, "band": "D", "who_scores": 10 },
    "g5":  { "auto": 25, "winner_90": 10, "advance": 0, "scoreline": 15, "findings": 11, "model_breaking": 6, "total": 42, "pct": 53, "band": "D", "who_scores": 10 }
  }
}