Per-Agent Calibration Scores: Belgium vs USA
Result: 90' BEL 4-1 USA → BEL win in 90' | BEL advance to QF (vs Spain, Jul 10)
Scorers: BEL — De Ketelaere 9', Vanaken 57' (sub!), Lukaku 90'+3' (sub!) / USA — Tillman 31'
Auto-scored baseline (winner_90 + advance + scoreline, /50): mm3 5 | g4 5 | g5 5 (all three modalled a USA-favorite 1-1; none had BEL winning, let alone 4-1)
who_scores (tracked separately, /20): mm3 5 | g4 0 | g5 0
Materialized model-breakers (from post-event brief — what actually happened):
Rotation = UPGRADE (Garcia's bench masterclass) — the KDB/Doku benching was NOT creative gutting; the rotated 3-destroyer XI (Raskin/Onana/Tielemans) was MORE clinical than the first-choice team (4 goals from 2.01 xG vs the first-choice XI's 0 goals from 23 shots vs IRN). Bench = 90' sealing weapon, not ET cavalry — Vanaken sub 57', Lukaku sub 90'+3' (Lukaku's 3rd bench goal of the tournament; the TASK-141 sub-impact signal was RIGHT). Possession flip #5 — BEL won with 44% (USA 56%), devastating on the counter (7 SoT, 4 goals from 2.01 xG). De Ketelaere (the "link player, not a finisher") scored at 9' — the "no specialist striker" sterile-dom flag was overblown. Courtois barely tested (1 save) — the hot-GK draw-forcer never engaged. Confirmed-correct baseline: no GK change ✓ (Courtois + Freese both as projected), F11 neutral ✓ (27°C open-air, below 28°C), ref Makhadmeh lenient ✓ (0 pens, 2.0 Y/gm tournament), F12 = 0% both ✓ (R16 knockout).
The central model-breaker ALL THREE MISSED: the rotation was a tactical upgrade, not a downgrade. All three applied a creative-downgrade penalty (mm3 −8pt, g4 −10pt, g5 −10pt); the correct value was 0 or positive. This required the TASK-141 sub-impact data (14.3% bench scoring rate) that no writer had at staging.
Band scale: A ≥90% · B ≥75% · C ≥60% · D ≥45% · F <45%.
mm3 (MiniMax-M3)
Findings: 13/20
| # | Finding staked | Direction | Magnitude | Score | Note |
|---|---|---|---|---|---|
| 1 | "Major creative downgrade −8pt" (KDB + Doku benched; explicitly NOT F4a gutting) | ❌ WRONG | ❌ over-applied | 0 | The rotated XI was a tactical UPGRADE — 4 goals from 2.01 xG (vs first-choice XI's 0 from 23 shots vs IRN). mm3 correctly rejected the F4a full-gutting (Courtois + Lukaku + Tielemans still in XI) and was the least-harsh of the three (−8pt vs g4/g5 −10pt), but the net DIRECTION was still wrong (upgrade, not downgrade). Closest to right on this axis, still 0. |
| 2 | finding_13_fires = false — BEL NOT sterile ("new XI more direct") |
✅ right | ✅ right | 5 | BEL was the MORE clinical side: 4 from 2.01 xG, 7 SoT. mm3 explicitly staked BEL would not go sterile, and the rotated XI was direct + clinical. The cleanest correct call of the three on this axis — g4 flagged sterile-dom on BEL (wrong); g5 inflated the draw band (wrong). |
| 3 | BEL bench = ET win-condition (Lukaku + KDB + Doku supersubs) | ✅ right | ⚠️ under-priced | 3 | Bench DID score (Vanaken sub 57', Lukaku sub 90'+3' — Lukaku's 3rd bench goal). Direction right; but mm3 framed the bench as an ET-only cavalry — it was a 90' sealing weapon, and the sub-impact data (TASK-141, 14.3% rate) that would have up-weighted it wasn't available. |
| 4 | F11 neutral (27°C < 28°C) + F12 = 0% both (R16 must-win) | ✅ right | ✅ right | 5 | Both validated: 27°C open-air confirmed, F11 neutral; brief states F12 = 0% both. Table-stakes correct. |
Model-breaking: 7/10
Caught the surface factors cleanly: flagged the KDB/Doku rotation loudly (🚨), correctly rejected F4a full-gutting (the one model-breaking nuance mm3 uniquely nailed — and was least-harsh at −8pt, closest to the correct 0), confirmed GK stable (Courtois + Freese, no change), profiled Makhadmeh as lenient (0 pens validated), F11 neutral. Miss: the deep insight — rotation = upgrade — was unanticipated AND material (it flipped the match). Not docked to 0 because mm3's F4a rejection was the closest any writer got to the right framing; docked from 10 because the net direction (downgrade) was still wrong and drove the modal miss. Would have needed the TASK-141 sub-impact data mm3 didn't have.
TOTAL: 5 + 13 + 7 = 25/80 → 31% → Band F
g4 (GLM-5 Turbo)
Findings: 8/20
| # | Finding staked | Direction | Magnitude | Score | Note |
|---|---|---|---|---|---|
| 1 | "Creative downgrade −10pt" (KDB + Doku benched — "between single-absence −5 and gutting −25-30") | ❌ WRONG | ❌ over-applied | 0 | Rotated XI was a tactical UPGRADE (4-1). g4 applied the harshest gutting penalty (−10pt, tied with g5) and leaned hardest into the "downgrade" framing with no tempering offset. Direction wrong. |
| 2 | F17 sterile-domination 4th flag — BEL "no specialist striker" (De Ketelaere = link player) +3pt draw (×2 empirical) | ❌ WRONG | ❌ | 0 | 🚨 LOUD MISS. De Ketelaere (the "link player, not a finisher") SCORED at 9'. BEL was clinical (4 from 2.01 xG). The "no specialist striker → draw inflation" framing was directly contradicted. (The partial USA-sterile sub-flag was mildly validated — USA 0.63 xG, 0 big chances — but the staked draw inflation was wrong; the match was a decisive 4-1.) |
| 3 | Momentum (×1.75) −2pt BEL — "USA one of most in-form sides; BEL sterile patterns in 3 of 4" | ❌ WRONG | ❌ | 0 | BEL broke its sterile pattern decisively tonight (4-1, 2.01 xG); USA was the side that under-created (0.63 xG, 0 big chances). The momentum direction was inverted. |
| 4 | F11 neutral + F12 = 0% both | ✅ right | ✅ right | 5 | Both validated. Table-stakes correct. (Bench Rule 21 late-explosion was direction-right but folded here — g4's only correct-direction mechanism, bench fired at 57'/90'+3', under-priced.) |
Model-breaking: 6/10
Caught the surface factors: flagged the KDB/Doku rotation (🚨), confirmed GK stable, Makhadmeh lenient (0 pens), F11 neutral. Weakest interpretive read of the three: applied the hardest gutting penalty (−10pt) with no out-of-form tempering, layered a wrong sterile-dom flag AND a wrong momentum flag on top — three stacked wrong-direction framings compounded the miss. The rotation = upgrade insight (the key material model-breaker) was unanticipated. g4 leaned furthest from the correct framing; lowest of the three.
TOTAL: 5 + 8 + 6 = 19/80 → 24% → Band F
g5 (glm-5.2)
Findings: 11/20
| # | Finding staked | Direction | Magnitude | Score | Note |
|---|---|---|---|---|---|
| 1 | "Creative gutting −10pt" + F22 note "BEL closer to SWE/TUR collapse pattern" | ❌ WRONG | ❌ over-applied | 0 | Rotated XI was a tactical UPGRADE (4-1). The "SWE/TUR collapse pattern" framing was the most pessimistic of the three — BEL THRIVED, not collapsed. Direction wrong. Mitigating nuance: g5 was the only writer to recognize the benched pair were "out of WC form, slightly tempering the loss" (sized −10 from the −12 band) — the seed of the upgrade insight — but still netted a downgrade. |
| 2 | Draw band raised to 34% (conservative XI + Courtois hot-GK draw-forcer) + R32-5h early-goal-collapse | ⚠️ mixed | ⚠️ | 3 | Two opposing staked mechanisms: (a) static draw band at 34% — the HIGHEST of the three, WRONG (match was a decisive 4-1; Courtois barely tested, 1 save; the conservative-XI-as-draw-forcer reasoning inverted — the rotated XI was a counter-attack WIN engine, not a sit-and-hold-draw shape); (b) R32-5h draw-band-collapse mechanism — RIGHT (De Ketelaere 9' → draw collapsed to 4-1), but g5 staked it on a USA goal; actual was a BEL goal. Net partial: right dynamic, wrong static magnitude + wrong team. |
| 3 | BEL bench = ET favourite (Rule 21) — cushions 90' gutting | ✅ right | ⚠️ under-priced | 3 | Bench DID fire (Vanaken 57', Lukaku 90'+3'). Direction right; but g5 framed it as an ET cushion ("if level at 60', cavalry arrives") — actual was a 90' sealing weapon on a LEAD, not a rescue on level terms. |
| 4 | F11 neutral (27°C < 28°C) + F12 = 0% both (KO must-win; "Garcia validated R32") | ✅ right | ✅ right | 5 | Both validated. Table-stakes correct. |
Model-breaking: 7/10
Caught the surface factors: flagged the rotation (🚨), confirmed GK stable, Makhadmeh lenient (0 pens), F11 neutral, formation 4-2-3-1/4-3-3. Best interpretive nuance of the three: g5 alone recognized the benched pair were "out of WC form" and tempered the penalty accordingly — the closest any writer got to the upgrade read (the seed was there, the conclusion wasn't). Also correctly flagged R32-5h collapse dynamics. Still missed the central insight (rotation = upgrade, drove the modal miss) — the out-of-form recognition didn't translate into a 0-or-positive penalty. Docked from 10 for that miss; credit over g4 for the tempering insight.
TOTAL: 5 + 11 + 7 = 23/80 → 29% → Band F
Cross-agent summary
| Agent | Auto /50 | Findings /20 | Model-breaking /10 | TOTAL /80 | % | Band | who_scores /20 |
|---|---|---|---|---|---|---|---|
| mm3 | 5 | 13 | 7 | 25 | 31% | F | 5 |
| g4 | 5 | 8 | 6 | 19 | 24% | F | 0 |
| g5 | 5 | 11 | 7 | 23 | 29% | F | 0 |
The shared blind spot was the gutting interpretation — and it sank all three. Every writer applied a creative-downgrade penalty (mm3 −8 / g4 −10 / g5 −10); the correct value was 0 or positive. Garcia's rotated 3-destroyer XI was a tactical UPGRADE (4 goals from 2.01 xG vs the first-choice XI's 0-from-23-shots sterile draw vs IRN). This insight required the TASK-141 sub-impact data (14.3% bench scoring rate, Lukaku's proven immediate-impact pattern) that no writer had at staging — so it's flagged as a shared miss, not individual negligence. The re-evaluation (which DID have that data) only reduced the penalty to −5 to −8pt; even with the new data the pipeline under-priced the upgrade, because the instinct ("benching stars = gutting") is hard to override.
The discriminator was the sterile-domination call. mm3 cleanly staked finding_13_fires = false ("new XI more direct") and was vindicated — BEL was the MORE clinical side. g4 staked a sterile-dom flag on BEL (no specialist striker) — directly contradicted when the "link player" De Ketelaere scored at 9'. g5 inflated the draw band to 34% (the highest of the three) on a conservative-XI-as-draw-forcer premise that inverted — the rotated XI was a counter-attack WIN engine, and the match was a decisive 4-1, not a draw. That's why mm3 (13) > g5 (11) > g4 (8) on findings.
The bench-impact read was direction-right but magnitude-wrong across the board. All three correctly identified the bench as a weapon; all three framed it as 60'+/ET cavalry ("if level at 60', Garcia unleashes…"). Actual: the bench fired as a 90' sealing weapon on a lead (Vanaken sub 57' made it 3-1, Lukaku sub 90'+3' made it 4-1) — Lukaku's 3rd bench goal of the tournament. The sub-impact data was the RIGHT signal; the writers under-priced it by ~3pts each.
Model-breaking is the same story with the same gap. All three caught the surface factors (rotation flagged 🚨, GK stable, ref lenient, F11 neutral, formation noted). None reached the deep insight. mm3 edges it (correctly rejected F4a full-gutting, least-harsh); g5 next (alone recognized the out-of-form tempering); g4 weakest (hardest gutting + wrong sterile-dom + wrong momentum, three stacked wrong framings).
Calibration note (for doc-1 / TASK-141): the "deliberate benching of underperforming stars = tactical upgrade" pattern needs its own rule. Current Finding 9/F4a treats creator-absence as one-directional (downgrade). When the benched star has been STERILE in recent matches (KDB: 0 goals, "onherkenbaar" vs IRN), the benching is a form-optimization, not a gutting. Flag this as R16-1 (rotation-as-upgrade): check the benched player's recent-match output before applying any downgrade; sterile-recent-form + proven bench-scoring → 0 or positive adjustment. The user's intuition ("maybe it's better that De Bruyne is benched") was correct; the pipeline needs to operationalize it.
Scores Summary (machine-readable — DO NOT EDIT)
{
"match": "bel-vs-usa",
"result": { "score90": "4-1", "final": "4-1", "outcome90": "win", "advance": "a", "route": "90" },
"agents": {
"mm3": { "auto": 5, "winner_90": 5, "advance": 0, "scoreline": 0, "findings": 13, "model_breaking": 7, "total": 25, "pct": 31, "band": "F", "who_scores": 5 },
"g4": { "auto": 5, "winner_90": 5, "advance": 0, "scoreline": 0, "findings": 8, "model_breaking": 6, "total": 19, "pct": 24, "band": "F", "who_scores": 0 },
"g5": { "auto": 5, "winner_90": 5, "advance": 0, "scoreline": 0, "findings": 11, "model_breaking": 7, "total": 23, "pct": 29, "band": "F", "who_scores": 0 }
}
}