MetaSpace.bio is a rule-based integrity layer: instead of estimating how likely a sensor reading is, it checks a small set of explicit rules (invariants) on residuals and component health and acts on the result. Every decision can be traced to a rule; the simulation contains no machine learning.
This site runs the idea on a simplified digital twin of Landsat-9 (sunlight and eclipse from a computed orbit, battery, solar wings, star trackers). It is a software simulation at TRL 3-4: there is no hardware, no flight data and no certification.
Reproducible evidence for every claim made here is in the evidence bundle.
The runtime checks in the sandbox test this model for internal consistency (bounds, round trips, an independent recomputation of the feasibility). They are not a formal proof.
A Kalman filter keeps an estimate and an uncertainty (a covariance). Its confidence falls when measurements stop arriving or get noisy. A rule-based check asks a different question: is this value physically allowed? If not, it acts at once instead of averaging.
The baseline in this simulation is deliberately simple: a linear constant-velocity filter whose confidence comes only from its covariance, without a residual or innovation check, and which sees GPS data only. It therefore does not notice wrong but regular GPS data, a lost solar wing or a dead battery until the GPS stops. A production navigation filter would usually add residual gating, so the comparison shows what invariant checks add to a covariance-only confidence; it is not a benchmark of real filters.
Both controllers are compared on the main page (detection) and on the navigation plan (task decisions judged against the physics model).