A verifiable demand-response rail for AI compute.

When the grid asks an AI datacenter to reduce load, GRACE binds who authorized the exact event, what was allowed, which supplied actuator and meter claims were accepted, and the result computed from them, with the baseline method pinned by hash. The bundle verifies offline. The bundle and its Action State record are signed under one operator-side key; the approver signoffs, the actuator acknowledgment, and the meter statement are each signed under their own pinned keys. It does not prove physical meter truth or decide settlement.

COSA moves the megawatts. EMILIA binds the authorization, supplied observations, and deterministic result.

Flexible compute is growing, but grid programs still need evidence they can evaluate.

Duke University’s Nicholas Institute finds the 22 U.S. balancing areas it analyzed could absorb 76–126 GW of new load if that load accepts curtailment of 0.25% to 1% of its maximum potential annual energy use. ERCOT alone adds about 10 GW at 0.5%. But that headroom is only bankable if the curtailment evidence is verifiable enough for a grid operator to evaluate. Today, operators may have to rely heavily on self-reported logs and supplied measurements. GRACE addresses one part of that gap: authorization and tamper-evident evidence over supplied inputs, with the baseline method pinned by hash but not executed.

Authorize → verify → shed → measure → compute → package.

One evidence bundle, evaluated under each program’s rules.

First mover: an AI/HPC datacenter or neocloud operator with a DR aggregator or utility sponsor: the party that holds the interconnection/payment incentive and a grid counterpart that can evaluate the bundle under its program rules.

Available now: the full reference circuit, visible end to end.

A runnable reference circuit is published with its tests. It issues a grid.curtailment order, simulates a shed, signs separately keyed simulated meter evidence, computes the reference result, and emits the signed Proof-of-Curtailment Bundle. The adversarial paths refuse tampered telemetry, forged orders, and replay. The control-room view replays a recorded run of that circuit so each transition is inspectable without implying a physical deployment or complete source population. The actuator and meter adapters are reference simulations, and no physical measurement has been run. A hardware demonstration with a host-approved compute node and independent meter still requires a facility partner.

GRACE does not invent the baseline. It pins the method’s hash so a method swap is evident.

The baseline methodology belongs to the ISO/program (CAISO ELAP, PJM CBL, ERCOT). GRACE pins its digest and binds the accepted authorization, supplied meter observations, and deterministic computation. The result is tamper-evident given the supplied inputs, including the baseline value carried in the meter statement; GRACE does not execute the pinned method. It does not establish that the readings are physically true or complete, that an event qualifies under a tariff, or what should be paid. Settlement remains a separate program decision.

A deployment may preserve an authenticated OT transport event digest as a distinct evidence leg. That digest does not grant authority or prove the physical outcome.