alnn_rewrite/tasks/templates/model_report.typ

// Per-model evaluation report.
//
// Data is injected as a JSON string via `sys.inputs.data` by
// tasks/analysis.py (model_report). The bundled Typst compiler has no
// `json.decode`, so we parse with `json(bytes(...))`.

#let data = json(bytes(sys.inputs.at("data")))

#set document(title: data.title)
#set page(
  paper: "a4",
  margin: (x: 2cm, top: 2cm, bottom: 1.6cm),
  numbering: "1 / 1",
)
#set text(size: 10pt)
#set par(justify: true)

#let fmt-pct(x) = [#calc.round(x * 100, digits: 1)%]
#let fmt-num(x) = [#calc.round(x, digits: 3)]

// ---- Header ----
#align(center)[
  #text(18pt, weight: "bold")[#data.title] \
  #v(-3pt)
  #text(9pt, fill: luma(110))[
    Generated #data.generated · schema v#data.schema_version ·
    seed #data.seed · commit #raw(data.git_commit)
  ]
]
#v(2pt)
#text(9pt)[*Working directory:* #raw(data.work_dir)]
#line(length: 100%, stroke: 0.5pt + luma(200))
#v(4pt)

// ---- One section per model family ----
#let family-block(fam) = {
  heading(level: 2, fam.name)
  text(9pt, fill: luma(90))[
    Source: #raw(fam.source + ".nc") · noise σ = #fmt-num(fam.noise_sigma) ·
    #fam.n_models model(s) · #fam.n_samples test samples · #fam.n_mc MC pass(es)
  ]
  v(4pt)
  block(fill: luma(245), inset: 8pt, radius: 4pt, width: 100%)[
    *Ensemble accuracy:* #fmt-pct(fam.accuracy_mean) ± #fmt-pct(fam.accuracy_std)
    #h(1fr) *Best member:* #fmt-pct(fam.accuracy_best)
  ]
  v(6pt)
  table(
    columns: (auto, 1fr, 1fr, 1fr),
    align: (col, row) => if col == 0 { left } else { right },
    stroke: none,
    inset: (x: 8pt, y: 5pt),
    fill: (col, row) => if row == 0 { luma(228) } else if calc.odd(row) { luma(248) },
    table.header(
      [*Model*], [*Accuracy*], [*Mean pred. entropy*], [*Mean mutual info.*],
    ),
    ..fam.models
      .map(m => ([#m.index], fmt-pct(m.accuracy), fmt-num(m.mean_entropy), fmt-num(m.mean_mi)))
      .flatten()
  )
  v(10pt)
}

#for fam in data.families {
  family-block(fam)
}

#v(1fr)
#align(center)[#text(8pt, fill: luma(150))[ALNN evaluation harness — model report]]