Seal · Green AI

The greenest AI call is the one that never runs. And it is the one you can evidence.

When a policy refuses a call, the provider is never reached. No tokens, no inference, no energy. That refusal is signed, so the fact of it survives a challenge. What it avoided is a different kind of number, and we keep the two apart on purpose.

In the signed record

The refusal happened

Outcome BLOCKED, the policy in force, the time, and the hash of the request. Signed with Ed25519, checkable offline against the published key. A refused call is a fact, not a forecast.

Fact. Verifiable by a stranger.

Outside it, and never inside

What that refusal avoided

Energy per token modelled on the ML CO2 Impact method, static regional grid averages, per-model coefficients that are order-of-magnitude and not vendor-confirmed. Not metered at the device. No live grid feed.

Estimate. Read the method, versioned and dated, and hold the figure to it.

What we will not do

No carbon number is signed. The receipt records that a call was refused. It never records a carbon figure, because a number nobody metered does not belong beside a signature.
No claim of neutrality. We do not say carbon-neutral, net zero, or offset. We say a call was avoided, and here is the modelled estimate of what that would otherwise have cost.
An estimate, stated as one. Framed for CSRD and ISSB reporting as an indicative figure with a documented, versioned basis, which is what those regimes ask an estimate to carry, not a measurement dressed up as a fact.

This is the whole of the green claim, and its restraint is the point. A sustainability figure with no verifiable event underneath it is what this market is already full of. Ours has the event in the record and the estimate beside it, and it tells you which is which.