09 — State-space models and Mamba, intuitively

flowchart LR
    X["each time step"] --> GATE{"input-dependent gate"}
    GATE -->|sharp QRS| WRITE["write into memory"]
    GATE -->|flat baseline| DECAY["decay, save compute"]
    WRITE --> STATE["running summary\n(state)"]
    DECAY --> STATE
    STATE --> POOL["pool over time\nmean + max"]
    POOL --> VOTE["5-logit vote"]

An SSM carries a compact running summary instead of rereading the whole window per step. Mamba's trick is selectivity: the gate opens for spikes, closes for silence — keep the QRS, forget the baseline.

Model	Cost vs length	Memory	Fit for 360 Hz ECG
LSTM	linear, but slow steps	notebook, fussy	crawls on long windows
Transformer	length-squared	full attention	overkill, expensive
Mamba SSM	linear and fast	selective summary	keeps spikes, skips silence

No math needed to proceed; ../research/literature.md holds the equations for later.

Next: 10-fusion-loss-thresholds.md