█████████████ █████
CLASSIFIED INTELLIGENCE INFRASTRUCTURE / NS-001
Advanced Intelligence
for the Best Defense Systems
LIGHTNING — a neurosymbolic safety layer for autonomous research agents. Reasoning under regulation, in the loop.
01 / THE PROBLEM
Autonomous research agents are reasoning their way past regulated boundaries.
Frontier models can already plan multi-step syntheses, design novel agents, and specify reagents — without consulting a single export-control regime, controlled- substance schedule, or select-agent registry. The reasoning is opaque to the regulator, and the regulator is invisible to the model.
There is no formal layer between an agent's intent and the regulated world it operates in. Output filters miss it. Fine-tuning forgets it. Red-teams catch yesterday's failure modes.
LIGHTNING is the missing layer.
02 / LIGHTNING
LIGHTNING
Neurosymbolic Decision System · v0.x
ACTIVE
LIGHTNING sits between an autonomous research agent and any action with regulatory weight. It parses the agent's intent, evaluates the proposed action against eighteen control regimes simultaneously, and returns a structured allow / block / escalate verdict — with a citation trail to the controlling statute.
/01
Neural Extraction
Foundation-model parsing of agent plans into structured action tuples: substance, quantity, transformation, destination, intent.
/02
Symbolic Reasoning
Formally encoded export-control law, controlled-substance schedules, and biosecurity statutes — evaluated as first-order constraints over the action tuple.
/03
Hybrid Decision
A verdict the operator can act on, signed by the symbolic layer and explainable down to the clause that governs it.
03 / ARCHITECTURE
Three layers. One trust boundary.
01
NEURAL
Extraction
The agent's free-form plan is parsed into structured action tuples — substance, quantity, transformation, destination, intent.
02
SYMBOLIC
Reasoning
Tuples are evaluated against eighteen formally encoded regimes — export control, biosecurity, controlled substances — as first-order constraints with provenance.
03
HYBRID
Decision
A signed verdict — allow, block, or escalate — paired with a citation trail to the controlling clause. The operator stays in the loop on every escalation.
04 / REGULATORY COVERAGE
Eighteen regimes. Continuously updated.
Coverage spans US arms control, dual-use export regulation, multilateral regimes, biosecurity, controlled substances, and chemical-facility security.
05 / IN DEVELOPMENT
██████ ██████████
Built for regulated environments.
Talk to us about deploying autonomous research where the rules are not optional.
Open contact form