{
  "kind": "tensor-dataflow",
  "title": "Keep hidden coordinates local until the output reduction",
  "description": "Show x[1×2] copied to two workers, first-weight column slices, two local activations, second-weight row slices, and one sum of two [1×2] partial outputs. Caption distinguishes logical tensor size from actual collective traffic.",
  "data": {
    "nodes": [
      {
        "id": "x",
        "label": "Input [1,2]"
      },
      {
        "id": "h0",
        "label": "Worker 0 hidden [1,1]"
      },
      {
        "id": "h1",
        "label": "Worker 1 hidden [0,6]"
      },
      {
        "id": "y0",
        "label": "Partial output [1,1]"
      },
      {
        "id": "y1",
        "label": "Partial output [6,−6]"
      },
      {
        "id": "sum",
        "label": "Sum [7,−5]"
      }
    ],
    "edges": [
      [
        "x",
        "h0",
        "W1 columns 0–1; ReLU"
      ],
      [
        "x",
        "h1",
        "W1 columns 2–3; ReLU"
      ],
      [
        "h0",
        "y0",
        "W2 rows 0–1"
      ],
      [
        "h1",
        "y1",
        "W2 rows 2–3"
      ],
      [
        "y0",
        "sum",
        "reduce"
      ],
      [
        "y1",
        "sum",
        "reduce"
      ]
    ],
    "logical_tensor_bytes": {
      "expanded": 1048576,
      "output": 262144
    }
  },
  "evidence": "Original educational calculation; no measured paper or hardware benchmark."
}