HouseFriend · Safety by Design

Why AI Must Not Directly Control Critical Home Engineering

AI is useful for context assessment and comfort adaptation. Safety for water, gas, heating, access and power loads must remain in a deterministic and verifiable environment.

Engineering insight · 10 minute read

AI Proposes or Selects Within an Allowed Range

A model must not create commands outside the engineering constraints of a subsystem.

AI may select preferred lighting, suggest a climate setpoint or flag unusual energy use. The command passes through a policy layer that checks property mode, allowed range, manual-control priority and interlocks.

If data is insufficient, the model does not guess a critical action. The system uses a known safe state, requests confirmation or escalates the event to a responsible person.

Interlocks Operate Independently of AI

Unsafe state combinations are blocked by automation that can be verified with repeatable tests.

Water protection defines leak sensors, valve position, timeout and regulated manual reopening. Boilers and climate systems enforce minimum and maximum setpoints, equipment protection and vendor requirements. Locks and access follow separate authorization rules.

AI cannot disable an interlock, alter a protection range or hide an alarm. Such changes require an engineer role, audit log and repeated verification.

Manual Control and Offline Operation Are Mandatory

The owner must retain a clear way to operate the home without an application, AI or external connectivity.

Physical switches and local panels continue to execute essential commands. If AI fails, scenarios fall back to deterministic rules. If internet access is lost, the local controller preserves critical functions and logs undelivered external events.

Maintenance mode allows an engineer to safely remove a subsystem from automatic control, perform work and return it to service using a checklist.

A Decision Must Be Explainable and Reviewable After an Event

The log connects input signals, selected scenario, applied constraints and the actual command.

Each significant action records time, property, zone, data source, configuration version, initiator, policy-check result and actuator state. NeroCore uses this history for diagnostics, change control and HouseFriend service operations.

The log does not replace protective automation, but it helps explain behavior, reproduce conditions and confirm that a constraint worked.

Acceptance Tests Failure, Not Only the Ideal Scenario

The system is ready only after normal operation, failures and recovery have been tested.

The test program covers internet loss, AI-service shutdown, sensor failure, conflicting signals, manual override, power recovery and log verification. Results are recorded in a protocol.

HouseFriend does not replace certified security, fire, gas or medical systems and does not guarantee life safety. Its role is comfort, engineering context, notifications and integration within the approved architecture.

Define Safety Boundaries Before Developing Scenarios

MOSCOS will prepare the HouseFriend architecture, criticality matrix, interlock rules and acceptance-test program.

Discuss Safe Architecture