Evaluation

Classify signed scalar-field values as inside/outside.

Returns -1 for inside, +1 for outside, and 0 for exact zeros or non-finite values. Set negative_inside=False for the opposite sign convention.

Return a mask for values within [-threshold, threshold].

Probability that a normal signed distance lies within a surface band.

Return scalar-field gradients for points via surface.gradient.

Return scalar-field residuals for points via surface.value.

Return predictive field variances for points via surface.variance_at.

Merge event-level evidence-margin diagnostics back into score rows.

Classify a non-negative log-evidence margin into qualitative buckets.

Return cluster-bootstrap intervals for paired evidence-margin summaries.

Return one best-vs-runner-up evidence-margin row per group.

Return generic pass/fail gates for grouped claim robustness.

Alias for grouped paired-margin summaries used by downstream reports.

Infer columns that define one paired model-decision unit.

Apply a summary after holding out each value of group_col.

Classify paired model wins using a symmetric log-evidence margin.

Summarize a paired margin-decision table, optionally by groups.

Summarize paired margin decisions over candidate thresholds.

Select the smallest threshold satisfying synthetic specificity gates.

Return rows satisfying all scalar constraints.

Constraint values may be either ("<=", value) tuples or mappings with {"op": "<=", "value": value} keys.

Select candidates under fixed quality constraints.

This is a named wrapper around :func:select_under_constraints for common equal-quality compression/compaction comparisons.

Return a boolean Series marking non-dominated rows.

Return indices of non-dominated rows.

Parameters

table: DataFrame whose rows are candidates. objectives: Objective column names to compare. directions: Either a mapping from objective name to "min"/"max" or a sequence aligned with objectives. feasible_mask: Optional row mask. If supplied, only feasible rows can be on the front. eps: Numerical tolerance for weak/strict comparisons. allow_missing: If true, objective values that are missing in either row are skipped for that pair. A row can dominate another only when at least one comparable objective remains and one comparable objective is strictly better.

Return whether left Pareto-dominates right.

left dominates right if it is at least as good on all comparable objectives and strictly better on at least one comparable objective.

Return feasible rows sorted by one objective plus optional tie-breakers.

Return points as a finite float64 array with shape (N, D).

Parameters

points: Array-like point set. name: Name used in validation error messages. expected_dim: Optional required point dimensionality.

Return directed or symmetric Chamfer distance between two point sets.

symmetric=True returns the sum of the two directed mean nearest-neighbor distances, matching the convention used by common 3D reconstruction diagnostics. squared=True applies the same convention to squared nearest-neighbor distances.

Return a deterministic random subset of points and its indices.

If max_points is None, non-positive, or at least the number of points, the original order is preserved and all indices are returned. Otherwise, max_points unique indices are sampled without replacement and sorted so downstream computations are reproducible and stable under independent chunk sizes.

Return quantiles of directed nearest-neighbor distances.

Distance from each query point to its nearest reference point.

The function uses :class:scipy.spatial.cKDTree when available and falls back to a deterministic chunked dense implementation otherwise.

Return standard directed/symmetric point-set geometry metrics.

The summary keys follow reconstruction terminology:

• accuracy_*: estimate-to-reference nearest-neighbor distances.
• completion_*: reference-to-estimate nearest-neighbor distances.
• chamfer_l1: sum of directed mean distances.
• chamfer_l2: sum of directed mean squared distances.

Return precision/recall/F-score rows for multiple thresholds.

Return precision, recall, and F-score for a distance threshold.

Select a fixed-size subset while preserving proportional tail capacity.

The candidate set is split at tail_quantile of reliability_scores. The tail receives the same retention fraction as the full set, ranked by tail_scores. The complement receives the remaining retained budget, ranked by primary_scores. If the tail is empty the function falls back to ordinary top-fraction selection by primary_scores; if the complement is empty, all candidates are ranked by tail_scores.

Return a boolean mask for the lower or upper reliability-score tail.

Return the threshold separating a reliability-score quantile tail.

Convert a retention fraction to a deterministic retained count.

min_count is applied only when item_count > 0 and retention_fraction > 0. Set min_count=0 for exact zero-retention behavior.

Return a finite one-dimensional float64 score vector.

Non-finite entries are mapped to zero. By default negative values are also clipped to zero, matching the common interpretation of scores as confidence, reliability, probability, or non-negative utility.

Return a bounded tail-rescue quota inside a retained budget.

Top-k selection with a bounded quota rescued from a reliability tail.

Return a deterministic mask selecting retained_count score entries.

Ties are resolved by the optional tie_break_scores and then by increasing original index, making the selection reproducible across NumPy versions.

Return a top-k mask whose size is derived from a retention fraction.