A binary LSTM can outperform caution/warning thresholds on latency and coverage in the same open loop
Positive detection is an observation that can gate hypothesis / evaluation — not a fault ID
Train / holdout still share the HSS31 generator — in-distribution detection, not a fidelity leap
Soft faults (#11, #26) stay closer to nominal; residual dynamics after recover can look “shaky”
RQ3 compares methods inside the loop — interpretability, evidence needs, failure modes
Thresholds remain auditable baselines; learned detection is one modular swap
Gated KG / time-series after LSTM arm; match Study A schedules
Full noise grid (4×3×32×10): F1 0.92–0.99, detect 31/31 every cell — classifier slides stay in backup