Monochrome terrain study of a high ridgeline.

Security / Protocols

Analysis in. Nothing out.

ARES and ICARUS process data inside the customer perimeter. This page lists the control set that enforces it, grouped as an assurance review reads it, and the evidence package an authority to operate is granted against.

Air-gapped by design On-node inference Zero data egress

01 / Principle

We build for operators whose data cannot leave the perimeter.

The architecture is derived from that constraint. No vendor cloud sits in the trust boundary of either system, and every control on this page follows from it.

02 / Control domains

Seven domains of control.

These protocols apply to both ARES and ICARUS, because both are built on the same sovereign edge platform. Each one is stated so that it can be confirmed against the system.

A / BND

Boundary & isolation

  • Systems run with no internet connection required for operation.
  • The secure-systems appliance is designed to sit behind a one-way hardware data diode: read-only ingest, no outbound network path.
  • Traffic between classification levels crosses a cross-domain guard rather than a routed connection.
  • Edge systems are engineered for EMCON and EW-contested theatres, where connectivity is scoped rather than assumed.
Air-gappedHardware diodeCross-domain guardEMCON

B / LOC

Data locality

  • Inference executes on the node: quantised models via llama.cpp / GGUF, on rugged edge compute.
  • Retrieval runs against a local vector store; the index never leaves the perimeter.
  • No prompts, documents or telemetry are sent to a third-party API.
  • Hyperborea does not receive, store or have access to customer operational data by default; any exception is a contractual term.
On-node inferenceLocal vector storeZero egress

C / SUP

Software supply chain

  • Systems are built as signed, reproducible artifacts.
  • Mission rules and playbooks are Ed25519-signed, versioned artifacts; an unsigned or altered rule set will not load.
  • Decision logic contains no executable code, so a rule update cannot introduce a new execution path.
  • Systems boot and run with zero third-party credentials, on infrastructure the customer controls.
Reproducible buildsEd25519 signaturesNo code execution

D / DEC

Decision assurance

  • Mission rules are deterministic: the same inputs produce the same recommendation every time.
  • Human-in-the-loop gates stand between the system and every consequential action.
  • Each firing writes a comparison-by-comparison trace: which rule, which values, which branch.
  • Logs are immutable and replayable, so any inference can be reconstructed at any timestamp.
DeterministicHuman-in-the-loopReplayable audit

E / INT

Interoperability boundary

  • Tactical-data-link and bus adapters (Link-16, MIL-STD-1553, CoT/ATAK, STANAG) terminate behind the guard.
  • Track custody is maintained end to end, so every track in the picture carries its provenance.
  • Engagement interlocks are enforced in the common operating picture rather than in the consuming system.
Link-16MIL-STD-1553CoT / ATAKSTANAG

F / RES

Resilience & degradation

  • Nodes sync peer-to-peer. There is no central server to lose, and no single node whose failure stops the others.
  • Each node operates autonomously when comms drop, then reconciles state on reconnect.
  • Behaviour degrades gracefully as links, power or bandwidth are lost.
Peer-to-peer meshAutonomous on comms lossState reconciliation

G / GOV

Lawful autonomy & oversight

  • A lethal effect is never released autonomously. It stages for two distinct human authorisers plus an armed hardware interlock, and the release is refused if the confirming authoriser is the same person or the interlock is unarmed.
  • That floor is machine-checked rather than asserted: no reachable state of the autonomy governor can drive a lethal domain to autonomous execution, and the commander's hold returns every domain to advisory, proven two independent ways: by an explicit-state proof over the code and by a TLA+ model check run as a separate tool.
  • Protected sites are enforced rather than weighed. Medical, worship, cultural, diplomatic, educational, refugee and detention sites, dams and nuclear stations are held as a no-strike overlay, and an aimpoint inside one, or inside its safety buffer, is refused as invalid.
  • The governance gateway is a boundary rather than an API a caller is trusted to use: authority travels as a short-lived signed release token, so a co-located process cannot route around it.
  • In threat intelligence, sensitive queries cite a warrant. Authority is checked per query for scope, validity window and covered parties, written to an immutable audit chain, and the retention it permits expires with it.
  • A model asked about something outside its training distribution says so at runtime, and autonomy is required to hear it.
Two-person releaseHardware interlockMachine-checked floorNo-strike overlayWarrant-scoped

03 / Evidence for accreditation

The argument an accreditor signs against.

An authority to operate requires controls and a structured safety argument. The argument states a top-level safety claim, decomposes it into sub-claims, and discharges each one against named evidence.

ARES generates that argument in GSN (Goal Structuring Notation, the assurance notation used across defence and aviation) from live trial evidence, not as a document maintained alongside the system.

The case re-derives on every run. A sub-claim whose evidence has regressed reports as unsatisfied. Each item of evidence carries its control references, mapping to NIST SP 800-53 and the relevant DoDD 3000.09 clause. Residual and open items are listed as open.

  • Assurance notation GSN
  • Control mapping NIST SP 800-53
  • Autonomy policy DoDD 3000.09
  • Derived from Trial evidence, per run
  • Regressed evidence Reports unsatisfied
  • Open items Named, not closed

Artifact 01Argument

Structured safety case

The top claim decomposed into sub-claims, each bound to a trial measure of effectiveness or a named test artefact. Renders as the document an accreditor reads, as a GSN diagram, and as a machine-readable record.

Artifact 02Floor

Release-floor conformance pack

A signed artifact demonstrating that the two-person release floor holds on the node under evaluation, at the time of evaluation.

Artifact 03Models

Model cards & evaluation report

What each model is, what it was trained on, how it was measured, and the uses it is not qualified for.

Artifact 04Inference

Proof-carrying inference

A classification bound to its evidence and signed, certifying that the decision resting on it fell inside the proven envelope.

Artifact 05Robustness

Adversarial deployment gate

A learned model reaches the release path only after measurement under adversarial attack, with certified output bounds over the networks. The gate fails closed.

Artifact 06Interpretability

Circuit auditor

Mechanistic interpretability applied as a runtime gate: it checks the computation a model performed, not an explanation generated after the fact.

04 / The boundary is physical

A one-way hardware data diode, enforced in silicon.

Field, sensor and mission data flow into the AI for analysis. The return path physically does not exist. There is no firewall rule to misconfigure, no exfiltration channel to exploit, and no vendor cloud in the trust boundary.

Because the control is physical rather than configured, it is inspected once rather than audited continuously.

  • Read-only ingest across the diode.
  • No outbound network path.
  • Local vector store and inference.
  • Signed, replayable decision logs.

Schematic · sovereign boundaryOne-way

Field / sensor Sovereign AI Data diode → no return path

05 / Applied to each system

How the controls land on ARES and ICARUS.

The platform is shared, but ARES and ICARUS sit in different places relative to the network, so their connectivity profile differs.

Edge autonomy interface: a filtered track list beside a track-detail panel showing sensor, heading, speed and range.
ARES · edge autonomy Fully air-gapped

ARESEdge autonomy

  • Network required None
  • Inference On-node
  • Decision logic Signed, deterministic
  • Boundary Hardware diode
  • Operational data egress None
Threat-intelligence interface: a ranked incident feed beside a theatre map with plotted incidents and entity markers, an assistant panel summarising critical events, and a severity timeline across the period.
ICARUS · full-spectrum intelligence Networked · isolated from the air-gapped side

ICARUSThreat intelligence

  • Posture Networked by design
  • Hosting Infrastructure you control
  • Build Standalone · in-house
  • Analysis Local
  • Path to the air-gapped side None

ICARUS is networked by design: collection and operations both reach outward. It runs on infrastructure the customer controls, is built in-house end to end, and is held architecturally separate from the air-gapped side: there is no path from it into ARES. Its operational posture is scoped and briefed per deployment, and accreditation with it. The evidence above is what we bring to that.

Request a security briefing.

We walk your security team through the trust boundary, the signing model and the audit trail in technical detail.

Get in touch