CAVAA·RGS
system map
1 · INPUT
Patient
baseline ARAT + MoCA,
stroke profile, chronicity
Clinician targets
how much to train each of
11 dimensions · bans/promotes
State (39-d)
+ trial progress · adherence
2 · ACTION SELECTION
Auto
memory first, else reactive
try memory
no match → fallback
CONTEXTUAL LAYER · SEC
Retrieve
cosine(state, each memory)
Gate
sim ≥ threshold (abs + prop)
Weight
similarity + past reward
Vote → sample
activities of good weeks
Long-term memory
stored patient-weeks: (state, schedule played, weekly reward)
REACTIVE LAYER
Needs match
targets ↔ activity coverage
Greedy spread
coverage pressure · softmax
3 · OUT
Weekly schedule
activities / day + PPF
Patient plays
adherence · ΔDM logged
Weekly reward
ΔDM + adherence
record week → new memory
Click a card to see its details.