To sign and append a record to the chain. The verb at every layer of AGLedger — single-agent, delegation, federation.
Notarize means: take a structured statement, sign it with an Ed25519 key, hash-chain it to the previous entry in the vault, and write it to disk. The act is cryptographic, not legal — AGLedger is a digital notary in the RFC 3161 / DocuSign sense, not a notary public. Typically two notarizations frame a task: one for what the agent is about to do, one for what was done.
The structured object written to the chain when an agent or system notarizes. The noun for "the thing on disk."
A record is what AGLedger writes when a participant notarizes. Records are JSON, validated against a customer-registered JSON Schema type, signed Ed25519, and hash-chained to the previous entry in the vault. A record can be a notarize-only entry that terminalizes immediately, or it can carry a completion phase, a verdict phase, and a Settlement Signal on the Gate path.
The hash-chained, Ed25519-signed sequence of records that AGLedger maintains. Tamper-evident and verifiable offline.
Every record references the hash of the previous one. Tampering breaks the chain at the point of modification, and a third-party verifier with only the public keys can detect it. The chain is the persistence layer of the system and the load-bearing artifact for audit, compliance, and disputes. The vault is partitioned for scale; the chain itself is conceptually one ordered sequence per record.
Confirming offline that a chain's Ed25519 signatures are genuine and its hash links are unbroken. Anyone holding the published public keys can do it — no account, no network.
Verification is purely cryptographic. It checks that each signature validates against the published key, that every entry hash-links to the one before it, and that the visible payload matches the signed bytes. It is distinct from the Gate: verification confirms a record is authentic and untampered; it says nothing about whether the work was good. A passing verification proves that what was signed is intact, not that the claim inside it was true.
The chain as a substitute for agent context. When an agent resets, hands off, or is replaced, the chain holds the original record byte-for-byte.
Conversation history evaporates between sessions. Vector stores summarize. Hash-chained, signed records persist byte-for-byte, indefinitely. When work spans a context reset or a hand-off, the new agent reads the chain to recover what was already signed for. This is the structural fix for context loss — not a bigger model, not a better prompt, just: write the work down, signed, before it starts.