How scoring and pruning fit together

Finding a minimal sufficient subset (see What "explaining a prediction" means here) happens in two genuinely separate stages, and understanding the division of labor between them explains why the library has both a scorer and a pruning strategy as independent choices.

Stage 1: scoring – a rough, fast importance estimate

A scoring strategy (e.g. ShapleyExplainer) repeatedly makes a random guess at which parts of the sample to keep, evaluates the model on that guess, and tracks whether keeping vs. dropping each item tended to correlate with the model staying confident. After many random trials, this produces a per-item importance score – a Monte Carlo estimate of Shapley value, if you're familiar with that concept: roughly, "how much did this item tend to matter, averaged over many different contexts of what else was present."

This is fast (hundreds of model evaluations, not an exhaustive search) but approximate, and it never actually finds a valid explanation by itself – it only ranks items by how promising they seem.

Stage 2: pruning – a search that's actually verified to work

Pruning takes that ranking and searches for a small subset, checking the model's real output at each candidate subset it tries. Unlike scoring, pruning's result is never taken on faith: the library's own internal contract requires that whatever pruning returns actually satisfies the tolerance you asked for (see Confidence gap and tolerance) – if it can't, that's treated as an error, not a silently-wrong result.

The scoring stage's ranking is only ever a search order for pruning – it tells pruning which items to try adding first, so a good search converges quickly. If you don't trust the scoring ranking for some reason, explain's order=GreedyForward() option ignores it entirely and searches by trying every candidate at each step instead (see Choose a pruning strategy) – slower, but not dependent on the scoring estimate being any good.

Why this split exists

Keeping scoring and pruning as separate, swappable stages means neither one has to be perfect on its own. Scoring can be a cheap approximation because pruning double-checks its suggestions against the real model. Pruning can search efficiently because scoring hands it a reasonable starting order instead of trying candidates in arbitrary sequence. Neither stage assumes anything about why an item matters – both operate purely by asking "what does the model actually output if I mask this" – which is also why the same two-stage design works unchanged for classification, regression, or any other notion of "confidence" you can express as a function of the model's output (see Explain a binary or non-softmax model).