14. Señales para visión artificial — 15.166 14.6.25 Pipeline de inferencia

Inicio · Capítulos · 14. Señales para visión artificial

15.166 14.6.25 Pipeline de inferencia

Etapa	ID	Acción
1	`CLS-01`	Load immutable ontology, class manifest, contrast sets and model card.
2	`CLS-02`	Validate synchronization, coordinate frames, modality availability and input quality.
3	`CLS-03`	Generate or ingest episode proposals with uncertain boundaries from 14.4.
4	`CLS-04`	Assemble pose, scene, contact, load, phase and kinematic tensors with masks.
5	`CLS-05`	Encode each modality without replacing missing data by zeros interpreted as observations.
6	`CLS-06`	Fuse temporal and interaction representations in causal or offline mode.
7	`CLS-07`	Predict observable attributes and coarse hierarchy nodes.
8	`CLS-08`	Generate multi-label canonical movement and failure candidates.
9	`CLS-09`	Select relevant confusion groups and run differential heads/rules.
10	`CLS-10`	Apply physical, temporal and ontology constraints with an auditable repair log.
11	`CLS-11`	Estimate open-set novelty, insufficiency and out-of-scope states separately.
12	`CLS-12`	Calibrate probabilities and construct prediction sets on the declared calibration profile.
13	`CLS-13`	Build evidence ledger with supports, contradictions and missing requirements.
14	`CLS-14`	Apply selective publication policy and route ambiguous cases to review.
15	`CLS-15`	Persist predictions, thresholds, revisions and provenance; never overwrite adjudicated truth.
16	`CLS-16`	Monitor drift, confusion-set performance, calibration and unknown rate by domain.

OnlineCausal sólo puede utilizar el pasado y presente; no puede etiquetar un fallo por desenlace futuro todavía no observado.
OfflineBidirectional puede refinar fronteras y usar el episodio completo.
Los logits de ambos modos no son directamente intercambiables ni comparten necesariamente umbrales.
La interfaz debe indicar si una etiqueta online es provisional y cuándo fue revisada offline.