Five progressive tests. Expected scores under trust-score-v2. Most AI energy claims stop at Test 1.
Progressive attestation
Five trust levels — computed deterministically from the evidence flags present in each certificate.
| Layer | Points | What it proves | Adversary required to fake |
|---|---|---|---|
| Independent Meter | +0.40 | An independent physical meter (PDU, branch circuit, facility or BMC) corroborates the reported power. The one source the audited operator cannot fake from inside the host. | Control of a meter outside the audited host |
| TEE Quote | +0.40 | Hardware-rooted attestation (Intel TDX and AMD SEV-SNP — live in production; AWS Nitro planned). Proves a genuine TEE committed to the samples. Alone it reaches only cross_checked: it does not identify the code that measured. | CPU vendor silicon break |
| Signed Exporter | +0.30 | The exporter response carried a verified Ed25519 signature. Cryptographic exporter integrity, on the host side. | Theft of the exporter signing key |
| Machine Fingerprint | +0.20 | The exporter's machine identity matches the agent's. Binds the reading to a host. | Control of the host's identity |
| Dual Source | +0.15 | NVML and RAPL agree. Both are counters on the audited host, so this is internal consistency, not independent truth. | Control of the audited host |
| MAXIMUM | 1.00 | Weights sum to 1.45 and the score is clamped at 1.00. Signatures, Merkle proofs and the Polygon anchor protect the record after signing and add no points. | |
No account. No API key. No reliance on our declaration alone.