AGENTS

Calibration

How complete real observation windows become model-ready.

#Complete-series gate

Calibration uses the minimum count across all configured symbols. This prevents a broad universe from appearing ready because only one asset has sufficient history.

#Unique source evidence

The database unique index uses agent ID, symbol, and source timestamp. Multiple application requests for the same upstream quote do not advance calibration.

#Rejected inputs

Null bid or ask, non-positive midpoint, stale source time, a source halt, a missing benchmark, and an API failure do not become valid observations.

#Example

NVDA 20
AMD  20
AVGO 18
────────
CALIBRATING 18 / 20