SetRaceAgent outperforms five of seven tested LLMs

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• TrackerAgent outperforms six of seven tested LLMsfinding0.873
  In the 98-game slice, TrackerAgent had a higher win rate or TrueSkill than all LLMs except Gemini 3 Flash.
• Two heuristic code agents (TrackerAgent and SetRaceAgent) outperform most tested LLMs.claim0.838
  Calibration that conditional logic can beat cost-efficient LLMs in this setting.
• SetRaceAgentmethod0.775
  deterministic code agent that greedily pursues quartet completion, bidding aggressively on near-complete sets
• Card-counting heuristics suffice to outperform most LLMs tested.claim0.760
  TrackerAgent's second-place ranking calibrates the benchmark and highlights LLM shortcomings.
• TrackerAgent and SetRaceAgent have TC tightness τ ≈ 0.2–0.25, looser countersfinding0.745
  Code agents trade bargaining precision for acquisition pressure.
• Two heuristic code agents outperform most tested LLMs, and behavioural traces surface recurring LLM failure modes including overbidding, self-bidding, bankrupt TC initiation, and weak opponent-state adaptation.quote0.741
  Abstract sentence summarising performance and failures.
• Hardest composition for LLMs: two TrackerAgents (C2, C7), only G3-F still wins majorityfinding0.733
  card-counting pressure compounds with multiple TrackerAgents
• Five independent LLM scorers from four labs produce identical rankings (Spearman ρ > 0.8).finding0.711
  Scorer bias validation: Claude Haiku, Gemini Flash, GPT-5.4, Grok 4, Kimi K2.5 all converge on same model ordering.