Horizon-conditioned Event Predictive Architecture. Components: encoder causal Transformer over context, returns h_t targe

Group P timesteps across all C channels into one token.

Random-cut (context, target, Delta_t) sampler over a corpus of sequences. For each sequence we sample n_cuts pairs (t, Delta_t) with Delta_t draw

Reversible per-context, per-channel z-score normalization (Kim et al. 2022).

Pre-norm Transformer encoder block.

Causal Transformer encoder over patched, RevIN-normalized inputs. Returns the last valid token h_t as the context summary.

Sliding-window dataset emitting (context, time_to_event, t_index).

LayerNorm + linear to a single scalar logit per horizon.

MLP that maps (context embedding, horizon) to a predicted future embedding.

Bidirectional Transformer over the target interval, then attention pooling.