Algorithm 2. Spatial Event Corrector.

Inputs:  Active expectations E_S = ⟨S, bindings, [t_a, t_b]⟩ // from pattern context & anchors  Observation buffer  O_K    // last K s of {x(t), μ_Z(x), class, bbox flags}  Config        method ∈ {BN, LSTM, TRAJ}, ∆_probe, ε_t, δ, B  Thresholds      τ_BN, τ_LSTM, τ_TRAJ  BN params      Θ_BN  LSTM params     M_LSTM (model weights), T (window), stride  Trajectory params  μ_ROI(·), exemplars H, FastDTW radius r, blend β State: none (stateless; queries only) procedure CORRECTOR(E_S):  // ----- GAP DETECTION -----  let ⟨S, bindings, [t_a, t_b]⟩ = E_S  if now < t_b + Δ_probe: return      // wait a small patience window  if exists RAW SpatialEvent(type = S, bindings, t_raw ∈ [t_a, t_b]) in O_K:     return              // no gap → nothing to correct  // ----- EVIDENCE ASSEMBLY -----  E ← slice O_K for variables in ‘bindings’ over [t_a − δ, t_b + δ]  if E is empty: return  // ----- SINGLE-METHOD IMPUTATION (chosen by config) -----  switch method:   case BN:    …    if p < τ_BN: return    conf ← p   case LSTM:    …    if p_t(t*) < τ_LSTM: return    conf ← p_t(t*)   case TRAJ:    …    if best = Ø or best.score < τ_TRAJ: return    t* ← argmax_{p∈best.σ} μ_ROI(p)            // most plausible time inside segment    conf ← best.score  // ----- DEDUP AGAINST RAW EVENTS -----  if exists RAW SpatialEvent(type = S, bindings, t_raw) with |t_raw − t*| ≤ ε_t in O_K:    return                         // prefer raw observation  // ----- EMIT ADDITIVE CORRECTED EVENT -----  emit SpatialEvent ⟨type = S, bindings, t* = t*, conf = conf, corrected = true⟩  // ----- TIME BUDGET GUARD (optional) -----  ensure wall-clock time ≤ B ms (degrade by skipping LSTM or lowering DTW radius if needed)