PHYSICAL AI / EXECUTION AUTHORITY

Intelligence
adapts.
Authority holds.

As robots become more adaptive, the boundary between what AI proposes and what a machine is permitted to do becomes fundamental.

NEDELV is developing AEGIS—an independent execution-authority architecture for embodied intelligence.

Explore the approach Our research
RESEARCH-STAGE SYSTEM · NOT A CERTIFIED SAFETY CONTROLLER
01 / THE PREMISESCROLL TO EXPLORE ↓
THE MISSION / 001

Adaptive intelligence needs
independent boundaries.

Robotic intelligence is becoming more capable, more general, and less predictable. The authority to act in the physical world should not depend on the planner's confidence—or its own account of what happened.

We're researching a distinct execution boundary that evaluates machine-generated proposals against independently administered constraints, and separates authorization from evidence of execution.

THE ARCHITECTURE / 002

Cognition proposes.
Authority disposes.

AEGIS separates the generation of an action from the decision to permit it—and from the record that a permitted effect occurred.

AEGIS / LOGICAL CONTRACTILLUSTRATIVE SYSTEM MODEL, NOT LIVE TELEMETRY
01 / PROPOSALAdaptive planner

Produces candidate actions. Does not grant permission.

UNTRUSTED COGNITION
REQUEST
02 / DECISIONAEGIS authority

Independently evaluates the proposed action and context.

CONTROL BOUNDARY
ADMITTED
03 / EFFECTExecution interface

Separates a permitted command from a claimed applied effect.

ACTUATOR GATE / RESEARCH
EVIDENCE
04 / OBSERVATIONEvidence & record

Associates simulated outcomes with matching transaction records.

TRACEABLE OUTCOMES
PROPOSAL ≠ PERMISSIONAUTHORIZATION ≠ EXECUTIONEVIDENCE ≠ CERTIFICATION
01

Independent by design.

The planner is not trusted to authorize its own actions. A separate decision path sets the terms of permissible execution.

02

Built for change.

As intelligence changes, the authorization contract is intended to remain separately governed and reviewable.

03

Evidence over assertion.

A decision and an actual effect are different events. The research prototype preserves that distinction in simulated execution records.

THE BOUNDARY / 003

The right to act
is not a property
of intelligence.

It is an independent systems decision. That's the research direction behind AEGIS.

AUTHORITY RESPONSETRANSACTION EXAMPLE / SIMULATED
01×
Proposal refusedOutside permitted action context
DENY
02↻
Planner revisesNew proposal / new evaluation
REPLAN
03✓
Authorized actionExecution evidence required
ALLOW
Illustration of the observed simulated-workflow pattern. Not a physical safety demonstration.
ACTIVE RESEARCH / 004

From observable behavior
to stronger assurance.

AEGIS is an implemented research-stage system, not a certified hardware safety product. Our work is focused on closing the gap between a demonstrated software boundary and dependable physical enforcement.

R / 01EXECUTION

Physical command mediation

Investigating how an independently administered authorization decision can be made binding at a robot's actual command interface.

Research objective ↗
R / 02ASSURANCE

State, timing & uncertainty

Testing request validity and enforcement behavior when sensing is delayed, proposals are stale, and workloads compete for compute resources.

Research objective ↗
R / 03INTERFACES

Embodiment contracts

Defining repeatable authorization interfaces without assuming that different robots share the same physical constraints or control dynamics.

Research objective ↗
EVIDENCE DISCIPLINE

Demonstrated in software simulation: separate authorization, denial, replanning, and receipt-gated presentation. Under investigation: physical mediation, repeatable timing under real workloads, and hardware-specific assurance.

RESEARCH / NOT CERTIFICATION
WHAT COMES NEXT / 005

Build intelligence.
Keep authority.

We welcome conversations with robotics researchers, hardware integrators, and technical partners working on dependable adaptive systems.

Research inquiries