A hardware-native enforcement layer for AI inference carbon governance. Every request intercepted, evaluated, and cryptographically sealed. Suitable for ESRS E1 limited assurance.
The February 2026 Omnibus I Directive has narrowed mandatory CSRD scope to approximately 5,000 large enterprises facing mandatory limited assurance audits with legal consequences for non-compliance. The primary audit failure mode identified in CSRD literature is material misstatement — unverifiable emissions estimates that cannot withstand independent assurance scrutiny. AI inference emissions (GHG Protocol Scope 3 Category 11) are the fastest-growing and least-governed component of corporate carbon liability.
ACERM eliminates material misstatement risk for AI emissions by providing per-transaction, cryptographically verified audit evidence. Production V1 has been validated in a live operational environment, governing 1,000+ inference requests across ten AI agents with verified carbon avoidance of 73.6kg CO₂e.
We define the Computational-Grid Synchronisation Gap as the structural disconnect between millisecond-scale AI inference execution and monthly-scale carbon accounting. This gap has three dimensions:
AI inference decisions are made in milliseconds. Grid carbon intensity is measured at 30-minute intervals. No existing system bridges these timescales at the enforcement layer.
AI agents operate entirely above the application layer. The electrical grid operates at the physical layer. No hardware-level interface exists between them in standard deployments.
ESRS E1 and CSRD mandate emissions reporting but specify no enforcement mechanism. Organisations can report high-carbon AI usage with no obligation to prevent it.
ACERM addresses all three dimensions simultaneously by operating as a mandatory network proxy — a hardware-level chokepoint between AI agents and upstream inference providers — that queries live grid telemetry at the moment of each inference request.
Several tools exist for AI carbon measurement and reporting. None address real-time enforcement at the inference request level.
Implemented as a binary process on an edge-computing device — the mandatory and exclusive network chokepoint between AI agents and upstream providers. Exposes an OpenAI-compatible endpoint. Any AI agent can route through ACERM by changing a single environment variable.
Maintains continuous connection to the UK National Grid ESO Carbon Intensity API, polling at 30-second intervals. Returns current intensity in gCO₂/kWh, forecast, generation mix, and grid state. Implements Carbon Integrity Safety Protocol (CISP) — fail-closed on telemetry loss.
Applies configurable governance policy to each intercepted inference request. Evaluates agent authentication, kill switch state, grid telemetry, operating mode, carbon intensity, and agent priority. Produces PERMIT, DELAY, DOWNGRADE, or DENY outcome.
Maintains authenticated AI agents, priority classifications, and team assignments. Stored in Supabase (EU West London) with local caching. Authentication uses SHA-256 hashing of the X-ACERM-KEY header value. Refreshed every 60 seconds.
/v1/chat/completionsOpenAI-compatible inference proxy (X-ACERM-KEY required)/api/heartbeatSystem health check and live grid state/api/auditAudit chain summary and ESRS E1 export/api/leaderboardAgent performance rankings/api/savingsCarbon avoidance metrics/api/toggle-modeShadow / Active mode switch/api/killEmergency kill switch — deny all non-HIGH requests/ws/oracleReal-time grid state WebSocket broadcastACERM uses the UK National Grid ESO Carbon Intensity API as its primary telemetry source, maintained under CC BY 4.0 licence by National Grid ESO.
Every enforcement decision is recorded as an immutable ledger entry in a SHA-256 sequential hash chain. The chain provides tamper-evident, independently verifiable evidence suitable for ESRS E1 limited assurance engagement.
Any modification to any entry breaks the hash chain from that point forward. Detectable by any independent verifier.
The genesis hash and chain structure are publicly documented. Any third party can verify chain integrity without access to ACERM systems.
Chain entries include all fields required for ESRS E1 Scope 3 Category 11 disclosure — per-transaction, with timestamps, agent identity, and carbon calculation.
The following results were verified in a live operational environment on Apple M-series edge hardware — May 2026.
ACERM audit chain provides per-transaction Scope 3 Category 11 data with timestamps, agent identity, carbon calculation, and chain hash. Suitable for limited assurance engagement under ISAE 3000.
ACERM satisfies the evidentiary requirements for CSRD mandatory limited assurance by providing independently verifiable, tamper-evident carbon accounting at the individual inference request level.
Carbon calculations align with GHG Protocol Scope 3 Category 11 methodology. Energy per token estimate is conservative and benchmarked against published Google data.
ACERM is content-blind by design. Only request metadata is inspected. No prompt content, no personal data, no user identifiers are processed or stored. DPA template available.
ACERM provides the Category 11 AI inference emissions data absent from all reviewed UK government supplier Carbon Reduction Plans — closing the structural gap created by PPN 06/21's mandatory category subset.
If you are conducting a limited assurance engagement under CSRD and require further technical detail on ACERM's methodology, audit chain structure, or evidentiary standards, we welcome professional correspondence.