AiEDs - AI Energy Disclosure Standard

AiEDs is the provider-neutral standard for reporting modeled energy and carbon from AI work. This published document is the citable source of truth: it is rendered at randomknights.xyz/aieds, its exact source bytes are served at randomknights.xyz/source/AIEDS.md, and the sha256 of those bytes is shown in the page header. Cite the methodology version together with that hash. The Required Record section below is the record contract a structured disclosure must meet.

Current Method

Methodology 2.4.0 is current. Versions 1.x are superseded and must not be used for new disclosures. Records stamped 2.0.0, 2.1.0, 2.2.0, or 2.3.0 stay valid, because every 2.x release after 2.0.0 was clarifying or additive and no published number moved. The v2 response path is energy-first:

energyWh = inputTokens / 1000 * modelInputWhPer1k
         + outputTokens / 1000 * modelOutputWhPer1k
energyWh = energyWh * providerPue
gCO2e = energyWh / 1000 * gridIntensityGPerKWh

Every coefficient, PUE, and grid-intensity value must identify its methodology version and evidence source. Do not blend disclosures from different impact model versions in one aggregate.

Metric Order

1. Level 1: energy in kWh or Wh, then carbon in gCO2e.
2. Level 2: operational facts such as input tokens, output tokens, cost, and latency.
3. Level 3: educational comparisons derived from Level 1, never substitutes for it.

Tree-Time uses one Mature Reference Tree, or MRT, at 21,000 g CO2e per year:

treeTimeMinutes = gCO2e / 21000 * 525600

The 21 kg value is canonical in methodology 2.4.0 and matches the shipped constant. Any 22 kg figure is superseded 1.x prose and must not be used.

Required Record

A disclosure must identify:

• stable record id;
• subject kind and name;
• scope: device, usage, inference, or training;
• time window;
• energyKWh and gCO2e;
• grid intensity and region when known;
• confidence;
• methodologyVersion;
• generating source and generatedAt timestamp.

Include compute provenance when available: tokens, GPU seconds, FLOPs, and hardware. Structured records must validate against the schema before release.

Confidence And Display

Use the confidence level supported by evidence. Direct measurement is stronger than a hardware factor, which is stronger than a token or FLOP proxy. Never upgrade confidence because the estimate looks plausible.

Every rendered disclosure states:

Energy and carbon are modeled estimates.
Tree-Time and equivalents are educational comparisons.

Do not describe an equivalency as an offset, credit, certification, verified outcome, or restoration result. Methodology changes require owner ratification, a version bump, and an append-only changelog entry in the canonical methodology.