EVALUATION_PROTOCOL.md

# SimplexTasks-12 Evaluation Protocol

## Fixed-Predictor Principle

Every benchmark slice fixes the predictor first and then varies only the conformal wrapper. This keeps the evaluation target on uncertainty allocation rather than on predictor training.

## Default Scores

- Real and synthetic composition tasks use Aitchison distance unless the target lies on simplex vertices.
- CIFAR-10 uses total variation / L1 because Aitchison distance is ill-defined at the one-hot boundary.

## Stratification Rules

- Default strata are entropy bins, boundary bins, or task-specific dominant-group partitions.
- Sensitivity sweeps use fixed alternative stratifications defined from cached prediction vectors only.
- For repeated-split experiments, one response-free stratification map is computed on the frozen prediction matrix `U` before splitting; calibration and test strata are obtained by indexing that same label vector.
- For synthetic repetitions, the same rule is applied once per generated draw before the calibration/test split.
- Stratification maps are not tuned per wrapper and do not depend on calibration/test responses. Separate calibration/test re-binning is diagnostic-only and is not the paper protocol.

## Main Metrics

- Marginal coverage.
- Max disparity across prediction-space strata.
- Worst-stratum coverage.
- Coverage variance.
- SSCV and mean radius; low-dimensional synthetic tasks also report a simplex-volume ratio.

## Wrapper Families

- Global split conformal.
- Group-wise / Mondrian conformal.
- Two-stage normalization.
- Exact or leave-one-out references where affordable.
- Diagnostic variants such as OneShot, TrainRes, and the current weighted implementation.

## Output Interpretation

The benchmark is designed to compare wrapper families under visible allocation-efficiency-compute tradeoffs. It should not be reduced to a single leaderboard or read as a conditional-coverage certification protocol.