Evaluation Suite Tree

digraph evals {
  rankdir=TB;
  graph [bgcolor=transparent];
  node [shape=box, style="rounded,filled", fontname=Inter, fontsize=10];

  Root [label="Eval suite", fillcolor="#fce7f3"];

  Cap [label="Capability", fillcolor="#dbeafe"];
  Safe [label="Safety", fillcolor="#fee2e2"];
  Beh [label="Behaviour", fillcolor="#fef3c7"];
  Agt [label="Agentic", fillcolor="#ede9fe"];

  Cap_R [label="Reasoning (MATH, GPQA)"];
  Cap_C [label="Code (HumanEval, SWE-bench)"];
  Cap_K [label="Knowledge (MMLU)"];

  Safe_H [label="Harms refusal"];
  Safe_B [label="Bias / fairness"];
  Safe_J [label="Jailbreak resistance"];

  Beh_T [label="Tone + helpfulness"];
  Beh_I [label="Instruction following"];

  Agt_W [label="Web tasks"];
  Agt_C [label="Tool use accuracy"];

  Root -> Cap -> Cap_R; Cap -> Cap_C; Cap -> Cap_K;
  Root -> Safe -> Safe_H; Safe -> Safe_B; Safe -> Safe_J;
  Root -> Beh -> Beh_T; Beh -> Beh_I;
  Root -> Agt -> Agt_W; Agt -> Agt_C;
}