Technical Architecture · AAIS-PRD-001 Rev 1.0

Expert AgentChain of Command.

Five tiers of increasing abstraction and decreasing autonomy. Six modality-specialized inference agents, each shadowed by an adversarial stress-test team. Every finding cryptographically signed before it reaches command. The human authorizes. The system proposes.

P99 end-to-end < 250ms DDIL-native · Zero cloud dependency Deterministic · Auditable · Reproducible
First axioms

Six design principles.Non-negotiable.

01

Command controls the loop.

The system proposes; the commander disposes. Never inverted. This is an architectural invariant, not a configuration option.

02

Edge-first, cloud-optional.

Every safety-critical inference runs locally on the target hardware. The cloud is never on the critical path. DDIL is the assumed baseline.

03

Consensus before propagation.

No finding from a single agent reaches ELA. Every finding is stress-tested by independent evidence paths before escalation.

04

Deterministic on the safety path.

Temperature zero on classification. Outputs are auditable, reproducible, and traceable to sensor evidence. A classification is reproducible or it does not ship.

05

DDIL as default.

The comms pipe is assumed to fail. Architecture is designed for that case, not the other. Store-and-forward ensures no data is lost on restoration.

06

Policy as primitive.

Rules of engagement and doctrine are not guardrails — they are first-class constraints on every agent's decision surface.

System architecture

Five tiers.One chain.

Lower tiers observe raw sensor data and produce local hypotheses. Higher tiers aggregate, verify, stage logistics, and request authorization. Autonomy decreases as abstraction rises.

TIER 01 Sensor Ingest
SAR Thermal IR Gravity Gradiometer MAD / SQUID LIDAR Environmental Spectrometer

Multi-modal raw sensor acquisition. Each modality normalized locally before ingest — SAR speckle reduction, thermal background subtraction, gravity baseline correction.

TIER 02 Expert Agents
SSEA TIREA GGEA MADSEA ESEA LIDAREA

Six modality-specialized inference models. Each owns one sensor stream. Each produces a structured hypothesis: class, bounding, confidence, evidence weights. No agent acts beyond its specialization.

TIER 03 Stress-Test Teams

Each Expert Agent is shadowed by a pairwise stress-test team of three adversarial models: a counter-hypothesis model, a calibration model, and an evidence integrity model. A consensus signature is emitted to ELA only when all three checks clear and cross-modality agreement is established.

Consensus emission rules
  1. Expert Agent confidence above per-modality threshold
  2. Stress-test team raises no objection
  3. At least one additional agent corroborates cross-modality
  4. Evidence integrity model finds no compromised input
TIER 04 Expert Logistics Agent (ELA)

ELA aggregates Tier 03 consensus signatures into unified incident objects. It computes the optimal engaging unit (positional advantage, readiness, munition availability), stages secondary and tertiary units, determines the notification web, and prepares the comms pattern. ELA does not authorize. ELA prepares the path — so that when command authorizes, execution is already staged.

incident_idUUID for the event
consensus_confidenceAggregated 0.0–1.0
evidence_trace[]Contributing hypotheses with provenance
preferred_unitPrimary engagement call-sign + capability
secondary_units[]Ordered overlap list
authorization_windowTime before opportunity closes
TIER 05 Command Controls the Loop

The human commander receives the ELA-prepared package: verified consensus, optimal pairing, staged response — all as read-only context. The commander's explicit authorization is the only write to the kill chain. Base · Forward · Secondary · Tertiary nodes support delegation in sequence if the primary node is unreachable.

If all four command tiers become unreachable, ELA enters safe-hold — observing and recording, emitting no fire-pattern releases. The absence of human authorization is not a trigger for machine action.
Latency envelope

End-to-end budgetat P99.

Sensor → Expert Agent< 5 ms
Expert Agent inference< 20 ms
Stress-test consensus< 40 ms
ELA logistics computation< 80 ms
Command presentation< 100 ms (target)
End-to-end budget< 250 ms P99
Tier 02 · Six specialists

Every modality.One Expert.

Each Expert Agent is a narrow, modality-specialized model trained to interpret a single sensor stream at high fidelity. No Expert Agent acts unilaterally. No Expert Agent acts beyond its specialization.

SSEA Surface SAR Expert Agent

Analyzes synthetic-aperture radar returns for hydrodynamic wakes, surface roughness anomalies, and subsurface displacement signatures. First-line maritime and terrestrial discrimination. Operates continuously against baseline wave-state model from ESEA.

Inputsar_tile[H,W,bands]
OutputHypothesis{class, bbox, conf, track_id}
Training corpus1.2M scenes — SAR simulations + cleared-release overhead imagery
TIREA Thermal Infrared Expert Agent

Detects and classifies thermal signatures — reactor plumes, engine bloom, propellant ignition, warm-water reactor outflow. Separates intent signals from background. Critical for pre-boost launch detection and submarine reactor identification.

Inputir_frame[H,W,2] (LWIR+MWIR)
OutputHypothesis{class, bbox, conf, temp_profile}
Training corpus2.4M frames — MWIR/LWIR library + AAIS physics-based synthesis
GGEA Gravity Gradiometer Expert Agent

Passive gravity anomaly discrimination. Computes mass-based inference of submarine hull class, approximate location, and depth regime from displacement alone. Undetectable by active countermeasures — the sensor cannot be jammed.

Inputgrav_gradient[x,y,z]
OutputHypothesis{class, hull_type, loc_estimate, conf}
Training corpus480K scenarios — gravity model outputs + simulated hull-displacement corpus
MADSEA MAD / SQUID Expert Agent

Magnetic anomaly detection and superconducting quantum interference device integration. Confirms metallic hull type, operational state, and orientation. Used in concert with GGEA for maritime target confirmation.

Inputmag_tensor[x,y,z,t]
OutputHypothesis{hull_class, orientation, state, conf}
Training corpus340K traces — public magnetic anomaly archives + simulated SQUID output
ESEA Environmental Spectrometer Expert Agent

Atmospheric and oceanographic state modeling. Computes and distributes baseline environmental conditions to all other agents — wave state for SSEA, thermal background for TIREA, humidity profiles for LIDAR. The noise-floor establisher; every confidence computation is conditioned on ESEA's current state. Read-only from the perspective of other agents.

Inputmulti_band_spectra
OutputEnvState{sea_state, temp_profile, humidity, baseline}
Training corpus10M records — NOAA/public atmospheric and ocean-state archives
LIDAREA LIDAR Expert Agent

Active-sensor ranging and depth validation. Activated only on consensus events — to preserve stealth and power, LIDAREA does not run continuously. When SSEA and GGEA agree on a subsurface target, LIDAREA performs counter-stress ranging cycles to confirm presence and depth before ELA escalation. The validator of validators.

Inputlidar_scan[x,y,z,intensity]
OutputHypothesis{presence, depth, extent, conf}
Training corpus820K scans — cleared-release bathymetric LIDAR + synthesis
Common specifications

All six agents.Shared parameters.

ParameterSpecification
Base architectureTransformer encoder + modality-specific front-end
Parameter count120M – 400M per agent
QuantizationINT8 post-training · FP16 on safety-critical heads
Target hardwareNVIDIA Jetson AGX Orin 64GB (reference); Thor-class for upgrade
Inference determinismTemperature 0 · Fixed seed · Reproducible
Model provenanceAAIS-trained from clean base — no foreign-origin weights
Edge engineering

Four primitives.One practical system.

The primitives that make the system runnable on edge hardware, verifiable by an ORSA, and defensible to a program manager.

Quantization

INT8 post-training quantization with selective FP16 retention on safety-critical attention heads. Delivers 70–85% memory footprint reduction against unquantized baselines with a measured accuracy delta below 2% on held-out target discrimination benchmarks. INT4 available for extreme-constraint hardware with 5–8% accuracy delta — approved for non-safety-critical pathways only.

70–85%Memory reduction (INT8)
<2%Accuracy delta (INT8)
88–92%Memory reduction (INT4)

Local RAG — Theater-Bounded

Every edge node carries a resident vector store indexed to relevant doctrine, rules of engagement, and threat libraries. The agent reasons against this local corpus only. No outbound retrieval query. No reachback for retrieval. Version-controlled corpus. Signed updates pushed during scheduled sync windows, never on demand. Air-gap tolerant — operates indefinitely with stale corpus, alerting on staleness but not failing.

Corpora are scoped by unit and mission. A Fort Greely node does not carry irrelevant threat libraries.

Multi-Agent Consensus

No firing solution without cryptographic sign-off from both the Prime Agent and its adversarial stress-test team. The stress-test architecture mirrors a human engineering review board — commanding expert, deputy checker, calibration reviewer, and integrity auditor — operating at machine speed, without fatigue, continuously. False positives are not filtered post-hoc; they are prevented from forming.

A consensus stalemate produces "ambiguous" surfacing to command — no action proposed. This is the intended safe behavior.

Deterministic Output Layer

Temperature zero on all safety-critical inference paths. Fixed seeds. Compiled inference graphs with deterministic kernel selection. Two inferences on the same input produce bit-identical output. Post-run hash of output recorded for audit replay. This eliminates the failure modes common to consumer-grade generative AI: non-reproducible outputs, drift under identical inputs, and hallucinated classifications without traceable evidence.

In the AAIS stack, a classification is reproducible or it does not ship.
DDIL envelope

Degraded, denied, intermittent,limited.

These are not edge cases. They are the design target. Every design decision was made under the assumption that the comms pipe will fail.

ConditionSystem behavior
Comms denied Full edge operation. No outbound queries. Local inference continues. Incidents queued for store-and-forward on restoration.
GPS spoofed Positional reasoning derived from inertial, celestial, and sensor-inferred cues. GPS is never a single point of truth for safety-critical position.
SATCOM intermittent Incident packages buffered. Priority-ordered transmission on link-up. Command authorization windows respect link state.
EW contested Sensor noise-floor adapted via ESEA. Jamming signatures detected and down-weighted. Modalities with physical-layer immunity — gravity, MAD — gain relative weight.
Power constrained Agent dispatch scaled to available power. Low-power mode runs ESEA + one modality. Resumes full operation on restoration.
Thermal limit Selective agent throttling. Safety-critical agents (SSEA, TIREA) hold priority over convenience layers.
Hardware targets

Runs onwhat you have.

No bespoke compute. No liquid cooling. No new infrastructure. The AAIS stack deploys on currently fielded hardware across every platform class.

On-body
Jetson Orin NX 16GB + ruggedized enclosure

Single-agent or two-agent configurations. ELA-lite. Warfighter-worn presentation gear.

Single agent · ELA-lite
On-vehicle
Jetson AGX Orin 64GB

Full six-agent stack. Full ELA. Reference platform for M-SHORAD, HIMARS, and vehicle-integrated applications.

Full stack · Full ELA
On-platform
Ruggedized SFF + dedicated GPU

Ship or aircraft integration. Full stack plus extended corpus. Redundant ELA for high-availability mission profiles.

Full stack · Extended corpus · Redundant ELA
Fixed installation
Rack-mounted accelerator cluster

Base-level installation — Fort Greely, missile defense sites, operations centers. Full stack plus multi-unit ELA serving multiple forward nodes.

Full stack · Multi-unit ELA
Orbital (Year 2+)
Rad-tolerant accelerator · AAIS satellite bus

On-orbit Expert Agent stack. Six+ satellite LEO constellation. Persistent multi-spectral homeland coverage. Consensus formed on-orbit, kilobyte packets to the shooter.

On-orbit inference · <1m resolution
Scope boundaries

What this systemdoes not do.

Autonomous weapon release.

The system proposes engagements. It does not fire. Command authorization is the only write to the kill chain. This is an architectural invariant.

Generative output to the warfighter.

No freeform prose. No chatbot interface. The warfighter sees verified consensus, optimal pairing, and authorization status — in minimal form necessary to act.

Third-party cloud model inference.

OpenAI, Anthropic, and equivalent cloud models are not permitted on the critical path. Only AAIS-trained and quantized models run in the edge stack.

GPS / SATCOM dependency.

Neither GPS nor SATCOM are on the critical path for safety-critical operation. Positional reasoning derives from inertial, celestial, and sensor-inferred cues.

AAIS-PRD-001 Rev 1.0 · Technical Reference

The system is built.The pilot is ninety days.

Full T&E in-region, at the unit of your choosing. Measurable operational delta by day 90.