Autonomy and safety
Autonomy should earn trust before it earns authority.
BASALT introduces autonomy in stages, with hard constraints, deterministic fallback, operator approval and auditable evidence built into every level.
L0
Observed
BASALT watches the site, builds the twin and measures prediction accuracy. It cannot propose or act.
L1
Advisory
BASALT proposes plans and explains the trade-offs. Operators remain fully manual.
L2
Approved
BASALT prepares the plan and machine-level work, but a named operator must approve execution.
L3
Supervised
Typical mature operating levelBASALT executes inside a pre-agreed envelope while operators can hold, amend or reverse the plan at any time.
L4
Bounded
Reserved for time-critical actions where waiting would create more risk, such as safety stops, containment or ride-through. Scope is explicitly pre-approved.
What keeps autonomy controllable
Deterministic fallback
Every cell retains deterministic control that can keep the process in a safe operating state without AI. Losing BASALT removes optimisation, not the basic control path.
Hard constraint kernel
Hard limits — safety distance, force, temperature, power import and permits — are enforced outside the models. No agent can negotiate around them.
Rehearsal before reality
Plans must pass simulation and safety checks before they reach hardware.
Evidence by default
Inputs, rejected options, binding constraints, approvals and outcomes are recorded for operator review and audit.
Standards as a baseline
The architecture uses IEC 62443 zones and conduits, ISO 10218 and ISO/TS 15066 for collaborative operation, and the NIST AI Risk Management Framework as design baselines.
BASALT
See what BASALT could change at your site.
Start in observation mode. BASALT builds and validates a live twin against real operations before any autonomous action is enabled. Authority expands only where the evidence supports it.