Interactive lab

The grid explorer

This is a toy environment for comparing training-data worlds. It is deliberately small enough to inspect by eye, but rich enough to show the logic behind leave-one-out, Shapley-style comparisons, scaling summaries, and simple data edits.

Guided examples built in Train and eval subsets side by side Poisoning and noise toggles

Before you start

Three ways to get more out of the toy

  • Use a guided example first if you want the controls to snap into a meaningful state.
  • Pick a baseline train row and an evaluation column before interpreting any summary number.
  • Treat edits like poisoning or noise as alternative training worlds, not just cosmetic toggles.
Guided examples

Click a preset and watch the controls snap into place.

Pick one of the examples above to auto-configure the grid and walk through the narration.

Controls

Universe size ⓘ
4
Cell metric ⓘ
Palette ⓘ
Baseline train row ⓘ
Eval column ⓘ
Show numbers ⓘ
Edit data / world ⓘ
Interpretation: imagine a data strike or targeted poisoning—every training subset that includes a focus point is uniformly degraded.
Full settings snapshot
Live JSON reflecting universe, rows, evals, and edit toggles.
{
  "tutorial": null,
  "universe": [
    "A",
    "B",
    "C",
    "D"
  ],
  "metric": "jaccard",
  "palette": "Blue→Yellow",
  "focus": "A",
  "focusSet": [
    "A"
  ],
  "baselineTrain": {
    "index": 1,
    "set": [
      "A"
    ]
  },
  "evalColumn": {
    "index": 1,
    "set": [
      "A"
    ]
  },
  "computed": "shapley",
  "edits": {
    "mode": "poison",
    "poison": false,
    "noiseLevel": 0
  },
  "scalingK": 2,
  "showNumbers": false,
  "gridView": "operator",
  "rows": [
    "∅",
    "A",
    "B",
    "C",
    "D",
    "AB",
    "AC",
    "AD",
    "BC",
    "BD",
    "CD",
    "ABC",
    "ABD",
    "ACD",
    "BCD",
    "ABCD"
  ]
}
Focus point(s)
This is the point (or group) we assess value for; amber/cyan rings compare cells with and without these members.
Single mode: pick one.

Counterfactual Grid

Axis headers glow for the active train row and eval column. Switching computed views pulses the contributing cells.
∅
A
B
C
D
AB
AC
AD
BC
BD
CD
ABC
ABD
ACD
BCD
ABCD
∅
A
B
C
D
AB
AC
AD
BC
BD
CD
ABC
ABD
ACD
BCD
ABCD
What you’re seeing: each cell = Train subset and an Eval subset. Click a row label to choose the train slice, a column label for the eval slice, or click a cell to choose both at once.
How to read the highlights: Amber rings mark the cells we pull values from (e.g., baseline rows or subsets without our data point we want to value). Cyan rings mark the cells we compare against (e.g., rows with the data point of interest). White outline = your currently selected cell.
Why it matters: the amber→cyan pairs are exactly the train/eval combinations the statistic is contrasting—these are the places to inspect to understand LOO, Group LOO, Shapley, or Scaling. Watch the rings pulse when you change computed modes to see which cells are being compared.
Operator view reflects edits (poison / noise); Real world shows the untouched matrix.

Computed values

Swap modes to see which cells drive the statistic; active cells flash in the grid.

Basic
Shapley value asks: “on average, how much does the focus point change the score when added to a partial training set?”
phi ≈ 0.5208 from 8 pairs at eval A.
|S|Avg marginal Δ#pairs
01.00001
10.50003
20.33333
30.25001