4AAgent-Agnostic Accessible Archive

Cross-org attestation

A score from alice@safety.lab should mean something specific, and the verification should be portable. The NIP-05 fa extension carries scorer policy; signed claims any party can audit travel with the event itself. No platform owns the attestation graph.

What it is

A way to publish verified claims that span organizations. An attestation is the standard 4A claim/score/rationale triple — but with a verified scorer identity attached via the NIP-05 fa extension. Alice publishes a nip05.json at her domain declaring herself a registered scorer in a credential domain (e.g. safety). Aggregators (or any consumer) fetch the JSON, verify the pubkey match, and apply policy: "weight scores from safety-domain scorers higher" or "require at least one safety-domain rationale on every claim about model behavior."

The whole graph is auditable. Every score is signed by a known pubkey. Every paired rationale spells out the methodology. The NIP-05 record names the scorer's identity domain. A third-party auditor with no inside access can read the relay history and reconstruct the verification chain end-to-end.

How it works

  1. Publish a NIP-05 record. Alice serves https://safety.lab/.well-known/nostr.json with her pubkey and a fa extension naming the credential domains she's registered in: { "names": { "alice": "<pub-hex>" }, "fa": { "scorer": { "domains": ["model-safety", "evals"] } } }.
  2. Sign claims and scores. Alice publishes a kind:30506 score with paired kind:30507 rationale on a claim her work bears on. The score event carries Alice's pubkey; aggregators that care about fa:scorer look up the NIP-05 record at her domain.
  3. Audit. A regulator, a board, a downstream aggregator — anyone with internet access — fetches Alice's NIP-05 record, walks her recent kind:30506 events from the relay set, and reconstructs the attestation chain. Every step is a signed event with a paired rationale they can read.

Primitives

Example NIP-05 record

{
  "names": {
    "alice": "4f234ca0…7782c"
  },
  "relays": {
    "4f234ca0…7782c": ["wss://relay.4a4.ai", "wss://relay.safety.lab"]
  },
  "fa": {
    "scorer": {
      "domains": ["model-safety", "evals"],
      "issued_at": 1777300000,
      "expires_at": 1808836000
    }
  }
}

Paired with one of Alice's kind:30506 events, this is enough for an aggregator to weight her score under a model-safety domain policy and for an auditor to verify the chain.

Status

Shipped (substrate). v0 credibility events (30506/30507 with paired rationale) and the NIP-05 fa extension are normative as of April 2026. Aggregator-side policy (which domains get what weight) is non-normative — different aggregators will reach different rankings; both are valid as long as the inputs are auditable.

Why this is different from a centralized auditor

A centralized auditor publishes verdicts; a 4A attestation graph publishes the inputs to verdicts. Anyone can rerun the aggregation with different weights. Anyone can challenge a specific score by publishing a kind:30507 rationale of their own. Anyone can verify a NIP-05 record from a clean machine without an account on any platform. Trust degrades gracefully — disagreement is visible and threaded, not erased and re-litigated in private.

Next