15.169 14.6.28 Modos de fallo del clasificador
- Closed-set forcing: every episode receives a known label even when unknown or insufficient.
- Single-label forcing: compatible local, global and failure labels are treated as mutually exclusive.
- Boundary leakage: the model sees future outcome in an online product or uses ground-truth boundaries at test time.
- Athlete leakage: identity-specific style appears in train and test.
- Route leakage: the model memorizes hold layout or color rather than mechanics.
- Scene shortcut: gym, wall texture or camera predicts the class.
- Taxonomy collapse: parent and subtype are merged without preserving semantic granularity.
- Intent hallucination: controlled, tactical or accidental is asserted from success alone.
- Force hallucination: pose or muscle appearance substitutes for measured/inferred load contract.
- Contact hallucination: 2D overlap substitutes for functional contact.
- Over-smoothing: short catches, regrips or true boundaries disappear.
- Over-segmentation: one movement becomes many repeated labels.
- Class imbalance: frequent static actions dominate rare competition movements.
- Missingness shortcut: a modality missing in one class becomes a label cue.
- Mirroring error: left/right labels are not transformed with augmentation.
- Calibration drift: probabilities remain numerically precise after domain change.
- Unknown contamination: unknown test classes or near-duplicates enter training.
- Composite confusion: a local action is mistaken for the complete transition.
- Censored outcome: failure/completion is inferred after the clip ends.
- Review overwrite: model output replaces human adjudication without revision history.
15.169.1 Sesgo de escena
SOAR muestra que modelos de acción pueden depender del fondo y que este sesgo perjudica open-set recognition [P46-R13]. En escalada, color de ruta, gimnasio, ángulo de cámara y atleta son shortcuts especialmente peligrosos. Deben medirse mediante splits y perturbaciones, no sólo enmascararse durante una demo.