Bounded Autonomy: Keeping AI Proposals Safe
MAIN lets an AI propose experiments but never lets it act unchecked — every command clears a deterministic rule gate and a physics simulation first.
- Layer 1: deterministic hard-limit checks on every action, no simulation needed
- Layer 2: digital-twin simulation blocks any move predicting non-positive growth
- Fails closed; firmware clamps and a physical e-stop backstop the software
An autonomous grower that can dose chemicals, change temperature, and drive lights is only trustworthy if it cannot harm the culture it tends. MAIN’s answer is bounded autonomy: the AI is free to propose, but every proposal must pass independent checking before any actuator moves. Creativity is upstream; safety is downstream and non-negotiable.
The first layer is a deterministic hard-limit check. When the AI proposes an action, a validator compares it against fixed bounds — light windows, temperature bands, per-dose caps on bicarbonate and nutrient, ceilings on clean-cycle time and inoculation volume. Malformed or unknown actions are rejected outright. The second layer is simulation: a proposal within the limits still runs through the digital twin, and if predicted growth is negative or negligible, it is blocked — for example, a bicarbonate dose that would push pH out of range is caught here.
Several choices make this robust. The checks fail closed: if the simulation errors, the proposal is blocked as a precaution. The layers are independent, so a flaw in the AI’s reasoning does not disable the rule gate. And the hardware firmware enforces its own clamps while a physical emergency-stop halts everything regardless of software state. The AI is treated as a proposer of hypotheses, not an authority — safety does not depend on the model being correct.