nx_race_vs_claude.nx

nx_race_vs_claude.nx -- compare the TEAM's capabilities to CLAUDE's, head to head (operator: "compare their capabilities to yours via the team and you racing off"). For each task type we hold an accuracy for the team and for Claude; the verdict is TEAM_WINS / TIE / CLAUDE_WINS. Crucially, where the team TIES on accuracy it still wins on COST (free) + REPRODUCIBILITY (deterministic -- same input, same output, which a sampled LLM cannot guarantee). So the honest picture: the team has caught Claude on the mechanizable tasks (and beats it on cost/reproducibility there); Claude still wins the semantic / creative rung -- exactly the open LLM-gaps. As PPMI->embeddings->Nishi-LLM land, the team closes those. The accuracies are Claude's honest self-assessment, grounded in the proven team tests. license_tier: ORIGINAL