Per-Agent Calibration Scores: Colombia vs Switzerland

Result: 0-0 (90' & AET) → SUI win 4-3 on penalties → SUI advance to QF vs Belgium
Auto-scored baseline: winner_90 + advance + scoreline (per agent below, 50pts) | who_scores 20/20 all (N/A — 0-0, no goals to predict)
Roof: CLOSED → F11 neutralized ✅ | Ref: Iván Barton | GKs: Vargas / Kobel (both confirmed, no model-breaking change)

⚠️ AUTO-SCRIPT ADVANCE BUG (read before the numbers)

node experiments/validate-agent.mjs col-vs-sui --json produced inverted advance scores because of a team-label mismatch:

• results.json stores this match as sui-vs-col (team A = SUI). The script resolves it via a reversed lookup and computes result.advance = "a" — which in the sui-vs-col convention means SUI advanced.
• But the writers' TOMLs use the col-vs-sui convention (team A = COL, so advance_a = COL). The script compares the TOML's advance_a/advance_b labels directly against result.advance ("a") without reconciling the swapped team axes, so the comparison is exactly backwards.

The script credited the two agents who picked the wrong team (COL) and penalized the one agent (g5) who picked the right team (SUI). All totals below use the corrected advance values. winner_90 and scoreline scores are unaffected and taken verbatim from the script (the scoreline logic is correct: g5's 0-0=10% cleared the ≥10% exact-match threshold → 20/25; g4's 0-0=8% fell just short → draw-band 15/25; mm3 had no 0-0 → draw-band 15/25).

mm3 (MiniMax-M3)

Findings: 15/20

mm3's headline analytical miss is F13 — it explicitly staked "no sterile-domination" and the match was the cleanest sterile-domination case of the R16. The draw-band elevation (which partly hedges F13) was right, but the finding itself was called the wrong way.

Model-breaking: 7/10

Caught the standard anticipatable model-breakers cleanly: F11 weather neutralized (roof CLOSED ✅), GK no-change (Vargas/Kobel both confirmed ✅), Barton pen-regression to 25–30% (0 pens materialized — regression held ✅), lineup rotation noted (Zakaria→RB flagged in the Díaz who_scores ✅), yellow-card timebombs flagged (Lucumí/Mojica/Lerma + Xhaka/Zakaria on 1Y ✅). Missed the decisive direction: leaned COL (advance 51%) while SUI advanced — did not catch SUI's defensive-depth / GK edge. The shootout itself was a coin-flip mm3 called 50/50 (honest, not a model-breaker miss).

TOTAL: 20 (auto) + 15 (findings) + 7 (MB) = 42/80 → 53% → Band D

• Auto: winner_90 5 + advance 0 (corrected) + scoreline 15 = 20/50

g4 (glm-5-turbo)

Findings: 18/20

Model-breaking: 7/10

Same standard model-breakers caught as mm3: F11 ✅, GK no-change ✅, Barton pen-regression ✅ (0 pens), lineup rotation ✅. Leaned COL (advance 53%) — missed the SUI-lean direction. g4's strongest model-breaking-adjacent read was F13 (credited in findings), which correctly anticipated the 0-0 mechanism. The bench-cavalry mechanism it staked as a SUI edge didn't fire in scoring (0-0).

TOTAL: 20 (auto) + 18 (findings) + 7 (MB) = 45/80 → 56% → Band D

• Auto: winner_90 5 + advance 0 (corrected) + scoreline 15 = 20/50

g5 (zai/glm-5.2)

Findings: 18/20

Model-breaking: 8/10

Best-in-class (per brief). Caught all standard factors (F11, GK-no-change, Barton pen-regression, lineup rotation) and uniquely leaned SUI (advance 53%) on its model-breaker analysis — correctly identifying SUI as the advantaged side that broke the COL-slight-edge consensus. The specific mechanism it staked (bench/depth edge) was directionally right, but the decisive factor turned out to be Kobel's shootout GK performance (a different, largely un-anticipatable mechanism). Direction vindicated (SUI advanced), mechanism off — hence 8/10 not 10. Reinforcing signal: g5 was the only agent to include 0-0 in its scorelines (10%), and the only one to pick SUI — its model-breaking read produced the closest full call of the three.

TOTAL: 32 (auto) + 18 (findings) + 8 (MB) = 58/80 → 73% → Band C

• Auto: winner_90 5 + advance 7 (corrected) + scoreline 20 = 32/50

Verdict

This was a genuine coin-flip that landed SUI — all three writers were honestly clustered at ~50% advance (mm3 51% COL, g4 53% COL, g5 53% SUI), and the 29% draw band was the correct modal-shape call (validated: lowest-scoring R16 match, 0-0 AET). The qualitative differentiator was F13: g4 and g5 correctly staked COL sterile-domination (Suárez now 0G in 5 WC games); mm3 staked it would NOT fire and was wrong. The model-breaking differentiator was g5's SUI lean — directionally right (SUI advanced) though the decisive mechanism was Kobel's shootout heroics rather than the bench scoring g5/g4 staked. g5 was the closest full call (only agent with 0-0 in scorelines + only SUI picker). Auto-script advance values are corrected for a team-label bug; all other auto values are taken verbatim.

Scores Summary (machine-readable — DO NOT EDIT)

{
  "match": "col-vs-sui",
  "result": { "score90": "0-0", "final": "0-0", "outcome90": "draw", "advance": "b", "route": "pens" },
  "agents": {
    "mm3": { "auto": 20, "winner_90": 5, "advance": 0, "scoreline": 15, "findings": 15, "model_breaking": 7, "total": 42, "pct": 53, "band": "D", "who_scores": 20 },
    "g4":  { "auto": 20, "winner_90": 5, "advance": 0, "scoreline": 15, "findings": 18, "model_breaking": 7, "total": 45, "pct": 56, "band": "D", "who_scores": 20 },
    "g5":  { "auto": 32, "winner_90": 5, "advance": 7, "scoreline": 20, "findings": 18, "model_breaking": 8, "total": 58, "pct": 73, "band": "C", "who_scores": 20 }
  }
}