Per-Agent Calibration Scores: Morocco vs Canada (R16, WC 2026)

Result: 3-0 (90') → 3-0 (final, MAR win in 90') | MAR advance to QF via 90'
Auto-scored baseline: winner_90 + advance + scoreline | who_scores tracked separately

mm3: 25/50 (w15+a10+s0) | who_scores 0/20
g4: 35/50 (w15+a10+s10) | who_scores 0/20
g5: 25/50 (w15+a10+s0) | who_scores 0/20

Headline: All 3 writers correctly called MAR as the modal 90' winner + advance — but all three mis-reasoned the mechanism (a coin-flip decided by Bounou's shootout pedigree) when the actual story was MAR's dominant 90' win (3-0, 55% possession, 4 shots → 3 goals). The flagship "hot-GK pen edge = THE model-breaker" thesis was mooted — pens never happened. Who Scores missed hard (Ounahi brace was consensus #3, not in any writer's TOML top-3; #1 Saibari blanked). g4 leads only because its 1-0 modal scoreline partially credited the 3-0 (winner-right), vs mm3/g5's 1-1 draw modal (wholly wrong). The 1pt advance spread — flagged as suspiciously tight in the consensus — produced near-identical qualitative scores: genuine correlation, not insight.

mm3

Findings: 16/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F11 NEUTRALIZED — NRG roof closed, no heat/press tax	✅	✅	5	Brief confirms roof CLOSED; F11 neutralized exactly as staked.
2	F12 = 0% both — R16 must-win, no conservation	✅	✅	5	Brief confirms F12=0% both sides; no conservation materialized.
3	Bounou confirmed-form hot-GK ×0.5-0.6 = THE model-breaker (58% pen edge, "pushes MAR over the coin-flip line")	✅	❌	3	Direction held (Bounou kept a clean sheet, 1 save) but the load-bearing pen-shootout mechanism was MOOTED — MAR won in 90', pens never happened. Brief: "hot-GK compound didn't fire." Flag loudly.
4	R32-3 reactive blueprint applies — cap MAR advance ~60%, respected at 55%	✅	❌	3	Direction right (MAR favorite + advance) but cap too conservative — MAR won 3-0 dominantly at 55% possession; the reactive-block framing was inverted (MAR was the dominant possession side, not a counter-attacking block). Brief: "R32-3 cap conservative for this matchup."

Note: mm3's finding_13_fires=false (sterile domination does NOT fire) was also CORRECT — MAR was clinical (3 from 0.79 xG), the opposite of sterile — but it's a minor TOML stake, not in the top 4. mm3 also staked F18 CAN-bench explosion IF level/leading late — conditional never triggered (CAN trailed from 50').

Model-breaking: 8/10

Caught: The single most material model-breaker — the manager/formation identity question (scout said Regragui reactive; canonical says Ouahbi possession). mm3 flagged the discrepancy and trusted canonical (Ouahbi possession 4-2-3-1) — and MAR was possession-dominant (55%). Getting this right prevented a worse reactive-block over-application. Also confirmed both GKs (no change — Bounou/Crépeau as projected), F11 weather correct, lineup gap noted honestly (lineups.md missing, used scout brief + canonical; no surprise materialized).
Gap: Referee TBD flagged as a process gap (not mm3's fault; ref wasn't material — 0 pens, CAN 24 fouls from frustration). Did NOT anticipate MAR winning dominantly in 90' (predicted pens path) — but that's an outcome/reasoning miss (scored in findings), not a model-breaker-factor miss. No actual model-breaking event (GK change/lineup surprise/weather/formation shift) materialized that mm3 failed to flag.

TOTAL: 49 / 80 → 61% → Band C

(auto 25 + findings 16 + MB 8) | who_scores 0/20 (separate)

g4

Findings: 16/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F11 NEUTRALIZED — Houston roof closed	✅	✅	5	Brief confirms roof CLOSED; F11 neutralized exactly as staked.
2	F12 = 0% — KO, no conservation	✅	✅	5	Brief confirms F12=0% both; no conservation materialized.
3	Bounou confirmed-form hot-GK ×0.5-0.6 = THE model-breaker (62% pen edge)	✅	❌	3	Clean-sheet floor held; pen-shootout mechanism MOOTED (won in 90'). Brief: "hot-GK compound didn't fire." Flag loudly.
4	R32-3 reactive blueprint analog (CAF template) — cap MAR ~60%	✅	❌	3	Direction right (MAR favorite/advance); cap conservative — 3-0, 55% possession, reactive framing inverted.

Notes: g4's finding_13_fires=true (sterile domination fires MILDLY, ×2.0 empirical weight, −1% MAR) was WRONG — MAR was hyper-clinical (3 goals from 0.79 xG, 4 shots/4 on target), the opposite of sterile. Partially offset by g4 staking CAN low-block vulnerability (CAN "struggled vs low blocks MD1 vs BIH") — VALIDATED (CAN press bypassed, 45% possession, 0 goals, 0.85 xG). Both are tactical nuances below the top-4 framework findings.

Model-breaking: 8/10

Caught: Same core as mm3 — flagged the Regragui-vs-Ouahbi discrepancy, trusted canonical (Ouahbi possession), confirmed GKs, F11 correct, lineup gap noted. The CAN-low-block-vulnerability insight was a genuine tactical positive that materialized (CAN's press bypassed). The formation/manager model-breaker (the key one) caught correctly.
Gap: Ref TBD flagged (not material — 0 pens). F13 sterile flag was a reasoning miss (above). Davies sub-impact stake (F21/F18) was conditional on CAN level/leading — never triggered. No material model-breaker factor (GK change/ref/lineup/weather/formation) was missed.

TOTAL: 59 / 80 → 74% → Band C

(auto 35 + findings 16 + MB 8) | who_scores 0/20 (separate) — leads only because 1-0 modal scoreline (winner-right) beat mm3/g5's 1-1 draw modal

g5

Findings: 16/20

#	Finding staked	Direction	Magnitude	Score	Note
1	F11 NEUTRALIZED — Houston roof closed + AC	✅	✅	5	Brief confirms roof CLOSED; F11 neutralized exactly as staked.
2	F12 = 0% both — R16 must-win, winner→QF	✅	✅	5	Brief confirms F12=0% both; no conservation materialized.
3	Bounou confirmed-form hot-GK ×0.5-0.6 = "the single variable most likely to decide who goes to the QF" (65% pen edge)	✅	❌	3	Clean-sheet floor held (1 save, barely tested); the "single variable" pen-shootout thesis was MOOTED — MAR won in 90'. Brief: "shootout pedigree was moot." Flag loudly — g5's staking was the most emphatic of the three.
4	R32-3 reactive blueprint — MAR advance "capped ~60%"; 56% sits under cap; "CAN is a live 44% underdog"	✅	❌	3	Direction right (MAR favorite); cap conservative — 3-0 dominant, 55% possession. CAN was not a "live 44% underdog" (battered 3-0, 0.85 xG, 0 goals).

Notes: g5 staked Koné single-absence "−3 to −5pt, NOT gutting per rule 13" (Finding 4a) — HELD (CAN created 0.85 xG, just didn't finish; not a 25-30pt gutting). g5 also flagged Finding 17 mild sterile-risk for MAR but correctly noted "CAN presses high, not low block, so MAR's chances still convert" — CORRECT (MAR clinical). F7 CB set-piece ~5-6% flagged — didn't fire (no CB scored). CAN travel fatigue −2pt staked. Most thorough model-breaking analysis of the three, but the top-4 framework stakes are identical.

Model-breaking: 8/10

Caught: Same core — Regragui-vs-Ouahbi discrepancy flagged + trusted canonical (possession), confirmed GKs, F11 correct, lineup gap noted. The Koné-not-gutting (rule 13) application was a clean Finding-4a call that held. Most nuanced factor coverage (Koné, travel, Davies uncertainty, F7 CB set-piece). Formation/manager model-breaker (the key one) caught correctly.
Gap: Ref TBD flagged (not material — 0 pens). Did not anticipate the dominant 90' win (predicted pens path) — outcome/reasoning miss, scored in findings. No material model-breaker factor missed.

TOTAL: 49 / 80 → 61% → Band C

(auto 25 + findings 16 + MB 8) | who_scores 0/20 (separate)

Cross-Agent Notes

The 1pt advance spread was correlation, not insight. All 3 writers staked the identical framework (F11/F12/Bounou/R32-3), so qualitative scores converge at 16/20 findings + 8/10 MB. The consensus doc itself flagged the 1pt spread as "tightest in R16 so far" — this match confirms it was artificial anchoring on the shared pens-path thesis.
The pens-path thesis was the systemic miss. All 3 built advance probability on a draw→ET→pens→Bounou path (e.g., g5: "42 + 25×0.40 ET + 25×0.40×0.65 pens"). MAR won in 90' dominantly. The Bounou "model-breaker" was mooted. Future: weight the 90'-dominant path higher when the favorite has a possession identity (Ouahbi) + the underdog's press is bypassable.
Who Scores was a clean whiff. Ounahi (brace, MOTM) was consensus #3 but not in any writer's TOML top-3; #1 Saibari blanked; CAN #1 David blanked. Rahimi (sub, 90+8') was in bench lists only. Consistent with the #1-misses-secondary-delivers pattern.
The manager/formation catch is the calibration win. All 3 correctly resolved the Regragui-vs-Ouahbi discrepancy (trusted canonical = possession) — and MAR was possession-dominant (55%). This prevented a worse reactive-block over-application and is the model-breaking highlight.
g4's edge is auto-scored only. g4's 74%/C vs mm3/g5's 61%/C is entirely from the 1-0 modal scoreline (winner-right, margin-off-by-2 → 10/25) vs the 1-1 draw modal (0/25). On qualitative reasoning, all three are tied (g4's F13-sterile-miss offsets its CAN-low-block validation).

Scores Summary (machine-readable — DO NOT EDIT)

{
  "match": "mar-vs-can",
  "result": { "score90": "3-0", "final": "3-0", "outcome90": "win", "advance": "a", "route": "90" },
  "agents": {
    "mm3": { "auto": 25, "winner_90": 15, "advance": 10, "scoreline": 0, "findings": 16, "model_breaking": 8, "total": 49, "pct": 61, "band": "C", "who_scores": 0 },
    "g4":  { "auto": 35, "winner_90": 15, "advance": 10, "scoreline": 10, "findings": 16, "model_breaking": 8, "total": 59, "pct": 74, "band": "C", "who_scores": 0 },
    "g5":  { "auto": 25, "winner_90": 15, "advance": 10, "scoreline": 0, "findings": 16, "model_breaking": 8, "total": 49, "pct": 61, "band": "C", "who_scores": 0 }
  }
}