On a phone, the grid is still usable, but it reads best if you start with the guide below and then scroll the matrix horizontally.
Rows are training worlds; columns are evaluation worlds.
Active cells are translated literally, then checked against every computation the grid knows how to make.
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). The same object can be imagined as trainable data, evaluation data, both, reserved for a secure holdout, or unavailable.
Rows show different training scenarios. Columns are evaluation scenarios. One cell means: train on the row world, then evaluate on the column slice.
Use Explore when you want to click around and build a set of evidence cells yourself. Use Compute when you want to ask for one quantity and see the exact cells that define it.
For now: direct cell reading, leave-one-out, evaluation value, group leave-one-out, pair interaction, Shapley value, Banzhaf value, Beta Shapley, row-size scaling, eval-size scaling, diagonal scaling, budgeted subset scans, and a simple unlearning reference comparison.
Several nearby methods still fit the same subset-world picture. Good next candidates include strike curves from the full dataset down to smaller retained worlds, acquisition curves from a seed set upward, regret against the full-data row, composition-stratified scaling, Owen-style group values, sampled or truncated Shapley approximations, and simple datamodel-style response surfaces fit over the grid.
Some of these would only need new formulas over cells we already display. Others would need new controls: group partitions for Owen values, budgets for coreset scans, sampling rules for Monte Carlo Shapley, or a small regression view for datamodels.
Some important methods live nearby but are not fully represented by this fixed subset grid. True influence functions need gradients, Hessians, and an optimizer. Differential privacy is a guarantee over release distributions and all adjacent worlds, not just a few visible cells. Poisoning needs a threat model, feasible perturbations, and an adversarial objective. Dataset distillation and condensation search over synthetic data, not only observed subsets. Active learning and experimental design need a candidate pool, label model, and acquisition policy. Curriculum learning, forgetting events, and data cartography depend on training order and trajectories. Membership inference needs an attacker and observation model.
So the grid can be a diagnostic baseline for those families, but it should not pretend to certify or replace the full method.