0143 — Prediction sets can read MAPIE's classification quantile, and the ceiling rule stays the default

Status: accepted · Date: 2026-09-17 · Amends: 0070

Context

Decision 0070 recorded that MAPIE 1.5.0 follows the ceiling rule, k = ceil((n + 1)(1 - alpha)), and not numpy.quantile(scores, (n + 1)(1 - alpha)/n, method="higher"). That holds for SplitConformalRegressor, and it is not what SplitConformalClassifier does: _compute_classification_quantile in mapie/utils.py calls exactly that numpy quantile (#866).

The two rules read different order statistics. Measured against MAPIE 1.5.0 with a row whose first class lies between the two thresholds, at alpha = 0.1:

calibration size	ceiling rule reads	MAPIE's prediction set reads	the row's first class
19	the 18th smallest score	the 19th	MAPIE includes it, the ceiling rule does not
99	the 90th	the 91st	the same

The three classification cases frozen for #441 sit where the rules agree, and their generator asserted MAPIE's sets against its own ceiling rule, so the corpus could not see it.

Decision

SplitConformal.Quantile takes a ConformalQuantileRule. ConformalQuantileRule.Ceiling is the zero value and what the two-argument overload uses, so every existing call keeps its answer. ConformalQuantileRule.MapieClassification reads numpy's higher quantile at MAPIE's level, and a prediction set built from it matches predict_set. Where that level passes 1, numpy raises and this returns double.PositiveInfinity, as 0070 decided for the ceiling rule.

Options refused

Switch prediction sets to MAPIE's rule outright, through a second quantile member. Parity by default, and a silent change to every set a caller already computes, for a difference of one rank. The maintainer chose to keep the established answer and make parity a named request.

Record the divergence and change nothing. The package's promise is MAPIE parity, and a caller who needs the sets MAPIE produces would have no way to get them.

Consequences

• docs/equivalence.md's SplitConformalClassifier row is identical only at ConformalQuantileRule.MapieClassification, and its numpy row names the rule it now matches.
• The conformal corpus carries two cases where the rules disagree, freezes both quantiles, and its generator asserts that they differ, so a rule wired to the other fails.
• 0070's sentence that MAPIE matches the ceiling rule on every case measured is true of the regressor only.