Model card: Position EKF

Module: src/nous/estimators/position.py

Backlog: BL-026

Inputs

• GNSS fix observations from PositionSubsystem.sensor_obs() (lat, lon, alt_m, with per-axis sigma).
• IMU integrals (accel + gyro) once the L2 IMU model lands.

Outputs

Estimate with point = {lat, lon, alt_m, vN, vE, vU} and a 6x6 covariance serialised to JSON.

SLA

• Update latency: under 5 ms per call on the reference profile.
• Covariance bound at the 95th percentile under nominal fix quality: horizontal sigma <= 5 m, vertical sigma <= 8 m.

Known failure modes

• GNSS multipath in urban canyons inflates the actual horizontal error beyond the filter's reported covariance; treat indoor or near-wall fixes as confidence_low.
• IMU-only periods (lost fix) accumulate error linearly with time; past 30 s of dead reckoning the covariance bound is no longer defensible.