Assert that a scorer is row-equivariant (single-sample deployable)

The decisive inclusion property for the single-sample method bank: at inference a specimen's score must be computable from that specimen alone given a frozen fitted model, with no dependence on which other specimens are scored alongside it. This harness implements the strengthened single-sample-deployability unit test prespecified in SAP Amendment #4 (section 2): for the same specimen, the batch score must equal its standalone score across ALL of

(a) the singleton \(\{x_i\}\) (the decisive single-sample case);
(b) differently-composed batches (varied size/composition) and an adversarial batch (the specimen among extreme, scaled neighbours);
(c) row-order permutations and batches with duplicated rows present;
(d) no model-state mutation: the model's bytes are identical before and after scoring (catches BatchNorm running-stat updates / online recalibration).

Permutation alone cannot detect mean/quantile-centering coupling (column means are permutation-invariant), which is why the singleton, subset, duplicate, and adversarial cases always run.

Usage

singlesample_assert_row_equivariant(
  score_fun,
  model,
  X,
  meta = NULL,
  n_perm = 4L,
  tol = 1e-08,
  n_singletons = 4L,
  model_digest = NULL,
  seed = 1L,
  extreme_transform = NULL
)

Arguments

score_fun: A function `function(model, X, meta)` returning one numeric score per row of `X`.
model: A fitted model object.
X: Numeric matrix (samples x features), at least 4 rows.
meta: Optional per-sample metadata (data frame with `nrow(X)` rows).
n_perm: Number of random row permutations to test.
tol: Absolute tolerance for score agreement (default 1e-8).
n_singletons: Number of distinct rows to test as singletons (always includes the first and last row). Pass `nrow(X)` for an exhaustive sweep.
model_digest: Optional `function(model)` returning a comparable snapshot of the model's mutable state. Defaults to a serialization snapshot; supply a custom closure for torch / external-pointer models that do not serialize deterministically.
seed: Random seed for reproducible permutations/subsets.
extreme_transform: Optional function of a numeric matrix that creates extreme neighbours within the input domain, preserving dimensions and names. The default NULL multiplies values by one million. For raw Ct, shift detected values down by log(1e6) / log(ct_base) cycles and leave non-detects unchanged. Only neighbours are transformed; target rows retain their original measurements.

Value

Invisibly `TRUE` on success; otherwise an error naming the failing case.

Usage

Arguments

Value

See also