Every source authenticates. Readings are rate limited, validated, de-duplicated and put into one common format, so nothing downstream depends on whose system it came from.
A rolling baseline for every store, signal and hour of the week. Each reading is scored against its own prior history before it is folded in.
Nobody writes a threshold, and nothing is tuned per customer.
A fixed set of critical conditions the platform commits to never missing.
It depends on nothing above it, so the guarantee survives an outage anywhere else in the system.
One agent per store, always on. It is given the live state of the store and reasons over three things it can point to. What comes out is a specific action with the numbers that justify it, not an alert.
Merges duplicates, then resolves who should see a recommendation from the roster crossed with who is actually on shift, and escalates when nobody is.
Chooses the channel, and if the first does not confirm delivery it moves to the next. A recommendation that never lands is recorded as such.
Every reading, decision, delivery and response written once and never altered — any past decision reconstructable exactly.
Outcomes are where the insights come from, and any change to how the system reasons is replayed against this history and compared before it ships.
Gold outline — software LumicIQ ships; everything else is what the store already runs. · Every request resolves to a customer scope before it reaches application code, and fails closed if it cannot.