Rendered scene · illustrative · nothing in this backdrop is a measurement

Signal quality
is not provenance

A trust layer between a GNSS receiver and an autonomy stack. It detects spoofed positioning, scores confidence continuously, and degrades the vehicle's authority in stages — full autonomy, coast, finish the leg, hold — instead of one binary trust decision.

Explore use cases ↓
--.---.----
--:--:-- ZULU

Four missions. Four attacks from the same threat model. One arbitras deciding, every epoch, how much authority the vehicle keeps.

Research

ARBITRAS is an assembly of published work — its geometry score, its credential layer and its scoring discipline each come from somewhere, and the relationships below are the ones stated in docs/design.md §3. Where a work carries no link the citation is the reference.

Our papers

  • Authenticating Satellite-Based Augmentation Systems (SBAS) Eric Hyunseung Cho · advisor Prof. Todd Walter · Stanford GPS Laboratory · 2026 A standards-compliant prototype of SBAS message authentication: 250 bits per second, no signature and nothing confirming the sender, and ICAO's answer to it. The upstream problem our credential layer assumes solved — that work authenticates the message, and this one asks what authority the message grants.
  • Designing Cryptography Systems for GNSS Data and Ranging Authentication Jason Anderson · PhD dissertation, Stanford University, Aeronautics and Astronautics · December 2024 The design space our credential layer sits in: data and ranging authentication under GNSS's bandwidth constraint, and the receiver-side processing that turns a verified message into a statement about trust. Its argument that the product can now dictate the cryptography, rather than the reverse, is why a mission authorisation here is short-lived and continuously renewed instead of signed once.
  • Coverage Optimization for Camera View Selection (CONVERGE) Timothy Chen*, Adam Dai*, Maximilian Adang, Grace Gao, Mac Schwager · Stanford University Fisher Information Gain reduced to a tractable coverage surrogate for active view selection. Different domain, same mathematics: the source of the geometry score — an information ratio over the line-of-sight matrix, normalised to (det ratio)^(1/(3+k)) — and of the next-best-observation ranking the console pursues in DEGRADED. The mathematics is what carries over; ARBITRAS holds no imagery and no camera anywhere.

Works cited

Authentication and credentials

  • RFC 4082 — TESLA Delayed disclosure of symmetric keys buys asymmetry without a signature per message. The hash chain and disclosure schedule here are ours; the primitives under them never are.
  • ICAO Doc 10169 — SBAS authentication TESLA, Ed25519, KMAC, compressed certificates. The protocol our credential layer reimplements; HMAC substitutes for KMAC under the library-crypto-only rule.
  • Galileo OSNMA Operational 24 July 2025. TESLA delayed disclosure with live PKI, Merkle root and OTAR. Upstream of us: it authenticates the message and says nothing about authority.
  • RFC 8032 — Ed25519 · FIPS 180-4 — SHA-256 · RFC 2104 — HMAC The anchor signature, the chain and the tag over each authorisation. All three come from cryptography; no primitive in this project is written by us.

Scoring, geometry and detection

  • How Stable Is a PNT Resilience Score? — arXiv 2607.05415 Ally and challenge. Single-number resilience ratings are weighting-unstable: re-weighting flips the winner in up to 22% of draws under nominal conditions. It is why the false surrender rate is reported as a distribution over sampled weight vectors, and why the composite never appears without its four sub-scores and the geometry block.
  • PNTTING — PNT Trust Inference Engine Molina-Markham et al., DHS S&T / MITRE. The closest published work: probabilistic trust inference for PNT user equipment, meant to assist automated decisions. More principled than ours; no credential binding, and not an authority arbiter with a published false-alarm rate.
  • DHS Resilient PNT Conformance Framework v2.0 (2022) · Reference Architecture v1.0 (2022) · IEEE P1952 Four resilience levels, source-agnostic and outcome-based; Zero Trust applied to PNT — "function under the assumption of compromise." Our state machine instantiates that behavioural intent, and is a candidate conformance test for the standard in development.
  • RAIM / ARAIM An independent-fault model rather than a coordinated adversary, answering "can I trust this fix" and not "how much may this platform do." Its subset-geometry reasoning is the closest published relative of the geometry score.
  • Rothmaier, Chen, Lo, Walter — ION GNSS+ 2021 · Psiaki & Humphreys, Proc. IEEE 2016 Spoofing detection by metric combination, and the canonical survey. Our detector is that class of method, kept in the measurement domain and made to carry a confidence rather than a verdict.
  • TEXBAT (UT Austin) · OAKBAT (ORNL) The standard spoofing datasets, both raw IQ. We work one layer up, on RINEX observables, which is why the attacks here are injected rather than replayed.