TERN answers where am I without satellites. ARBITER answers a different question — does this still cohere. Mission intent versus route. Fused sensor state versus physical plausibility. Firmware delta versus the navigation stack's implicit assumptions. 26MB engine, air-gapped, scoring live against the public compare endpoint.
Rank candidate routes against plain-language mission intent. Semantic path scoring beside the position fix — not shortest path.
Detect semantic incoherence a Kalman filter assumes away. Baro climbs, terrain reads flat — trigger reinitialization.
Score every firmware delta against the stack's implicit assumptions before it reaches a forward operating base.
A commander states intent in plain language. ARBITER scores each candidate route against it in one call. Switch the preset intent — the same four routes re-rank, because the scoring is semantic, not static. This is the layer beside TERN's fix.
The fused state is described from the last sensor window and matched against physical-state hypotheses. When the state coheres with a fault hypothesis instead of a motion model, that is the trigger to reinitialize or reweight. Preset three is Shaun's exact case: baro climbing, terrain flat.
A field-updateable nav stack carries implicit assumptions: fusion timing, fallback weighting, initialization order. ARBITER scores a proposed delta against named risk classes before it ships. A change that coheres with a timing or fallback-removal class gets held. The CrowdStrike note, as a gate.
The round-trip shown in the panels is browser-to-cloud and includes internet transit. The figure for TERN is the air-gapped edge number: 41–75ms inference on a co-located Pi 5, exact API parity, no cloud. That places ARBITER in the human-in-the-loop and low-rate-autonomy window (<1Hz). Tight inner-loop control (>10Hz) is out of scope — ARBITER scores route segments and state transitions, it does not stabilize them.
The score is not a black box. Every run returns the full ranked field of named hypotheses, so the output reads "state coheres with Sensor Conflict at 0.71, margin 0.18 over the nearest motion model." The operator sees which physical hypothesis the position resembles and by how much. Decomposing the contributing dimensions further is a defined next step.