Graph explorer

The graph explorer is the widest surface on the site. On a smaller screen, it helps to start with the guide below or use the grid first, then come back here when you want the lattice view.

Open the graph Try the grid first Read the memo first
What are we simulating?

We imagine an AI operator is training some machine learning model on different slices of data. There is a small set of data objects to choose between, named A, B, C, D, and so on. The operator will train a model on some slice of training data (e.g., A, B, and C) and then evaluate on a set of data (e.g., just A and B).

How do I read the graph?

Fix one evaluation slice first. Then read each node as a possible training set and each edge as a one-step move in dataset space. Moving downward removes data, moving upward adds data, and longer paths let you trace a strike or coordinated ablation instead of only reading a single local comparison.

How do I use this page?

Choose a graph lens, a score rule, a training world, and an evaluation slice. Then click nodes or use the quick actions to move through nearby training sets while the lens panel stays locked to the same selected node.

What questions can I ask here?

Local ablations, coordinated data strikes, Shapley-style edge sweeps, and scaling layers across the subset lattice.

Graph explorer

Walk the subset lattice directly.

Instead of scanning the full train-by-eval matrix, fix an evaluation slice and move through the graph of possible training sets. Nodes are training worlds; one-step edges become ablations, augmentations, or steps inside a larger strike path.

Universe size
4 datasets
The graph has 16 nodes once every possible subset is enumerated.
Score rule
Normalized overlap: a rough performance proxy when train and eval look similar, but still only set similarity.
Graph lens
Highlight one single-step deletion from the selected training node.
Clicking a node in the lattice updates this too.
Focus contributor
Choose the dataset whose edge additions or deletions you want to inspect.
Subset lattice

Nodes are training sets; colors read off the active eval slice.

Ablation edgeTrain ABCDEval ABCDJaccard
Selected train nodeActive eval sliceCurrent walkUnselected edit edge
Selected targetTrain ABCD | Eval ABCD | Score 1.0000Train ABCD is locked in against eval ABCD. Click a node or use the command buttons to move.
B∅ABCDABACADBCBDCDABCABDACDBCDABCD
Midpoint nodes can get wide as the universe grows. The canvas scrolls horizontally so the full lattice stays inspectable.
Current lens

Follow one ablation edge

Ablation delta: 0.2500

Move from train ABCD to ACD while keeping eval ABCD fixed.

Selected score1.0000
Train nodeABCD
Eval sliceABCD
Node degree4
f(ABCD, ABCD) - f(ACD, ABCD) = 0.2500
One-step edits

Neighbors of ABCD

Eval ABCD
Single deletions
Single additions
  • The full training set has no outgoing augmentation edges.