ML — ablation study

// ML evaluation vertical — models, datasets, evals, ablations.
// The ablations view at the foot of the document draws through the
// ablation_table renderer (ablation is its registered alias).

model base {
  title: "Flow-v1 base"
  params: 7000000000
  family: "decoder-only transformer"
}

dataset internal_eval {
  title: "FlowBench v2"
  n: 1240
  split: "held-out"
}

eval base_score {
  model: base
  dataset: internal_eval
  metric: "exact-match accuracy"
  score: 0.612
}

ablation no_typed_attrs {
  base: base_score
  delta: -0.084
  change: "Strip semantic types — fall back to generic boxes"
}

ablation no_propagation {
  base: base_score
  delta: -0.051
  change: "Disable arithmetic propagation"
}

ablation no_critique {
  base: base_score
  delta: -0.039
  change: "Skip critic step in Architect mode"
}

view ablations: ablation_table(base_score)