Research/Energy and alignment
Position · Energy and alignment
Aligning AI with the energy ecosystem.
In energy, alignment is not an abstraction. The mandate is written in grid codes, certification boundaries, tariff rules and the trust of the communities a grid serves.
Grid codes, metering certification, tariff rules and settlement rules become constraints the system structurally cannot break.
Energy already tells you what aligned means
General AI alignment asks how to make a system pursue human intent. The energy sector has an advantage most domains lack: much of the intent is already codified. Grid codes say what may be switched and when. Metering certification says who may generate a token. Tariff regulation says what may be charged. Settlement rules say how money reconciles. An AI system in energy does not need its values discovered through philosophy. It needs the existing rulebook translated into constraints it structurally cannot break, and that translation is engineering work that either happened or did not.
Our systems treat those boundaries as load-bearing. Prepayment token generation stays with the certified headend; our platforms orchestrate around it and are architecturally unable to mint. Automated actions that touch supply carry operator override as part of the control surface, not as a setting. A loss figure without its physical basis reports as unavailable, because a confident wrong number in a regulatory filing is an alignment failure with a signature on it.
The mandate is already written. Grid codes, metering certification, tariff rules and settlement rules become constraints the system structurally cannot break, and the parties the grid serves, operator, regulator, lender and community, can each verify from the record that the constraints held.
Aligned with whom: the ecosystem answer
An energy AI serves more parties than its buyer. The operator wants margin and uptime. The regulator wants compliance it can inspect. The lender wants portfolio numbers it can believe. The community wants power that stays on and billing it can trust, and a rural mini-grid lives or dies on that trust. These interests overlap but do not coincide, and a system aligned only to its operator will eventually be used to squeeze the parties who cannot read its code.
Alignment to the ecosystem means the system is engineered to be defensible to all of them at once. Consumption histories that support a disconnection decision are append-only and reviewable. Forecasts state their error against the honest baseline, so a lender is not pricing a demo. Verification records carry lineage an auditor can walk. The design question we ask is not "what does the operator want the system to do" but "which of the system's outputs will another party someday need to check, and will the record hold when they do".
The grid as a rehearsal for harder alignment
Distributed energy is a strong test bed for deployment alignment generally: physical consequences, money in the loop, multiple principals with opposing interests, and thin connectivity that removes the luxury of constant supervision. The disciplines it forces, certified boundaries honoured in architecture, mandates written as enforceable policy, failure directions chosen in advance, evidence kept for sceptical outsiders, are not energy-specific. They are what alignment looks like anywhere the output of a model becomes an action in the world. We build them into grids first because grids are where our work is, and because a doctrine proven where power and payment are at stake travels credibly to everywhere else.
