25 lines
1.5 KiB
Markdown
25 lines
1.5 KiB
Markdown
# AERE Proof of AI
|
|
|
|
An AERE Proof Protocol (AIP-23) envelope that proves the provenance of an AI agent's action: which model, which version, which
|
|
prompt, which tools, which output, which agent and who asked. It is verified by the SAME verifier as every other AERE proof, with
|
|
no new verification code. Notarization and the post-quantum certified point come from AIP-23: the statement hash is notarized on
|
|
AereNotary and covered by a post-quantum anchor.
|
|
|
|
```
|
|
import { buildProofOfAi } from './proof-of-ai.mjs';
|
|
const envelope = buildProofOfAi({ model:{name,version,weightsSha256}, prompt, input, output, tools:[...], agentId, requester, createdAt });
|
|
// -> an AIP-23 envelope; verify it with the AIP-23 reference verifier (verify-proof.mjs, published at aere.network/tools)
|
|
```
|
|
|
|
The envelope holds DIGESTS only: the prompt, the input, the output and the requester are stored as SHA-256, never raw. It lines up
|
|
with the on-chain contract `AereAIProof` (modelId, inputHash, outputHash, requesterHash), but it is PORTABLE: it exists without the
|
|
chain and needs the chain only for the certified point.
|
|
|
|
Test: `node proba-proof-of-ai.mjs` (6/6, with a negative control: an output changed after hashing is INVALID). What it shows: one
|
|
standard, many kinds of proof. The same verifier, a new kind of proof, no new code.
|
|
|
|
What it is not: no third party has reviewed this component, and a proof of provenance says what was recorded, not that the model's
|
|
output is correct.
|
|
|
|
(Until 2026-10-05 this README was written in Romanian.)
|