Live demonstration

Every number resolves to a cited source.

Carbon accounting for IT estates under the GHG Protocol Corporate Standard and ICT Sector Guidance. Continuous measurement rather than a baseline resampled at intervals, with row-level lineage from every reported figure back to the record it came from and the emission factor that converted it.

A number you cannot take apart is an assertion, not an account

Every figure carries the source records it consumed, the arithmetic applied, and the emission factor used — by identity and version, never as a literal embedded in code. One request returns the whole chain. This is a real response from the running system.

GET /api/v1/lineage/{result_id}chain complete
source recordmqtt · estate/pdu/winston-salem-nc-data-center/pdu-ws-01
sha-2567d1c9a04f2be6d3e8c5a1b47e09f2d6a…
activity1,259,597.183 kWh · metered
derivation366 smart-PDU readings; closing minus opening cumulative
factor0.270488924 kgCO2e/kWh · EPA eGRID · eGRID2023 Rev 2 · SRVC
citationepa.gov/system/files/…/egrid2023_data_rev2.xlsx
retrieved2026-08-29 · valid from 2023-01-01
result340,707.10 kgCO2e

Measured and modelled are not the same claim

A metered figure is an observation. A modelled one is an estimate from a device inventory. Both belong in an inventory; conflating them does not. Every figure states which it is, on screen and in the report.

metered

Smart-PDU telemetry

On-premise data-centre draw arrives over MQTT and is stored per reading. Where a site is metered, its modelled per-asset draw is excluded from the calculation so the same electricity is never counted twice.

modelled

Device inventory

End-user compute from a registry of devices, power-draw profiles and usage-hour assumptions, with home-office allocation modelled explicitly rather than assumed away.

provider-reported

Cloud

AWS, Azure and GCP carbon exports are stored as immutable source records and passed through against the provider's own published methodology. We do not re-derive their figures.

invoiced

Colocation

Metered facility draw taken from the provider invoice. Contracted PUE is recorded for reference and deliberately not applied — the meter already reads total facility power.

Emission factors are versioned data, not constants

Calculations reference a factor by identity and version. Recomputing against a newer vintage writes a new result and keeps the old one, so a restatement can be explained line by line rather than announced.

  • Transcribed from the publisher's own file. The workbook's SHA-256 is pinned; the extraction tooling refuses to run on a mismatch rather than producing factors from a file nobody verified.
  • Currently EPA eGRID2023 Revision 2 — the latest EPA publishes. Their page still advertises eGRID2024 for January 2026; no such dataset exists yet. When it lands, adopting it is an insert-and-supersede, not a migration.
  • Scope 2 by subregion, not by national average. The same server carries roughly twice the emissions in one US grid subregion as in another. A blended rate would hide the one lever a client can actually pull.
  • Where no citable factor exists, the figure is withheld. It is reported as a declared gap with the reason, never as zero and never as an industry average. The engine cannot substitute one by accident.

Managed on Databricks, or entirely on your own infrastructure

Four append-only stages — raw ingest, normalised, calculated, reported — in Delta Lake format, carrying row-level lineage keys throughout. The same tables run either way. The choice does not change the model, the factors, or the numbers.

4
pipeline stages
append-only
no update or delete path
2 modes
on-premise · Unity Catalog
1 click
self-verifying evidence pack

The evidence pack's manifest carries a SHA-256 of every file it contains, so a recipient can prove it has not been altered without running our software and without taking our word for anything.

Stated honestly, because the alternative is discovered later

We are not SOC 2 certified. The platform is built with the control surface in place so that certification would be an audit exercise rather than a re-architecture. We do not claim a certification we do not hold.
  • Role-based access with a read-only auditor role that holds no write capability at all — an assurer can verify without being able to alter what they are verifying.
  • Immutable user-action log recording who viewed, exported or changed what — including refused attempts. Append-only at the database level, not by convention: it resists rewriting even from a direct SQL session.
  • Multi-tenant isolation, proven rather than asserted: an automated test walks every scoped endpoint as one tenant and confirms another tenant's data is unreachable.
  • Encryption in transit and at rest, nightly encrypted backups with a restore that is rehearsed on every gate run, and per-container egress restricted to the single endpoint that needs it.
  • SSO-ready. OIDC hooks are stubbed behind a single token-issuance function. Stubbed — not implemented, and not claimed to be.