Follow-on from the append-only invocation log.
An append-only table with hash chaining is tamper-evident to someone who trusts whoever operates the store. It is not evidence to a third party who does not. For that we need entries signed by the kernel identity, and an export format something else can verify.
Proposed:
- Hash-chain entries, and sign either each entry or periodic checkpoints over the chain.
- Provide an export that a third party can verify offline given only the kernel's public key.
- Document precisely what the signature does and does not attest — in particular that it proves the kernel recorded the invocation, not that the invocation was legitimate.
Blocked on the key-custody issue. Signing with a key stored in plaintext next to the data it attests would not meaningfully improve on the unsigned version.
Related: #777 — this should not ship before the crypto review.
Follow-on from the append-only invocation log.
An append-only table with hash chaining is tamper-evident to someone who trusts whoever operates the store. It is not evidence to a third party who does not. For that we need entries signed by the kernel identity, and an export format something else can verify.
Proposed:
Blocked on the key-custody issue. Signing with a key stored in plaintext next to the data it attests would not meaningfully improve on the unsigned version.
Related: #777 — this should not ship before the crypto review.