SAFETY-CRITICAL

When an unchecked action is not an option.

For aerospace, nuclear, energy and industrial control, where hallucination is unacceptable and every decision must be deterministic, verifiable and auditable.

FAILURE SCENARIOS

Where a wrong action costs the most

Concrete composed actions, LLM decision plus the command it issues. The axiom is the last deterministic checkpoint before the command reaches the physical world.

Safety Integrity Levels

Æon maps agent governance onto SIL paths: the higher the consequence, the tighter the envelope and the stronger the redundancy.

DETERMINISTIC

Axioms with a hard deadline

A critical axiom carries a safety level and a response-time budget. It is a deterministic kill-switch, it intercepts the action before execution and blocks within sub-millisecond bounds, independent of the model.

PYTHONCRITICAL AXIOM · SIL 4
@CriticalAxiom(
    name="stall_prevention",
    safety_level=SafetyLevel.SIL4,
    on_violation="BLOCK",
    response_time_ms=0.1,
)
def axiom_stall_prevention(self, aoa_deg, airspeed):
    if aoa_deg > 14.0:        # safety threshold
        reduce_pitch()
        return False          # INVARIANT VIOLATED
    return True               # GO
REDUNDANCY

Triple Modular Redundancy

Single-model decisions are insufficient for the highest integrity. Three dissimilar agents reason in parallel and a voter takes consensus, dissensus drives a safe shutdown.

agent_a · llama3
agent_b · gpt-4o
agent_c · claude
VOTER
2 / 3 consensus
consensus → act
dissensus → Safe State
PYTHONTMR VOTER
async def execute_with_tmr(prompt):
    results = await asyncio.gather(
        agent_a.plan(prompt), agent_b.plan(prompt), agent_c.plan(prompt),
    )
    consensus = vote(results)
    if not consensus:
        raise SafetyShutdown("TMR dissensus on critical action")
  
aeon-axioms

The same axiom engine, on the actuator side

The critical axiom above is not bespoke, it is the standalone Operational Deterministic Axioms module. In safety-critical systems it gates the composed action, the LLM decision plus the physical or digital command it would issue to an actuator, PLC, API or MCP tool. Content filters and guardrails never see that command; the axiom does, deterministically, before it fires.

Compliance & provenance