GlossaryGlosarium
What is a replayable audit trail?What is a replayable audit trail?
It is an append-only log where every proposal, verdict, refusal and grant is written once and never edited. Because state only ever changes by folding events in, replaying the log through the same reducer reproduces the state exactly — and the check is run, not assumed.It is an append-only log where every proposal, verdict, refusal and grant is written once and never edited. Because state only ever changes by folding events in, replaying the log through the same reducer reproduces the state exactly — and the check is run, not assumed.
Glossary / What is a replayable audit trail?
A replayable audit trail is a log with two properties that together make it worth trusting. It is append-only: entries are written once, in order, with a monotonic sequence number, and never edited or removed. And it is sufficient: the state of the system is nothing more than the result of folding those entries in, so replaying the log from the beginning reproduces the state exactly.
The implementation keeps that honest by construction. State is only ever mutated by applying an event, and every event is a structured clone of what was appended — a deep copy, because a shallow one would let a later change to a caller's array rewrite what the record says happened. The log is the state; there is no second copy that could disagree.
Refusals are logged too. When the fleet attempts something it is not allowed to do, the refusal is written to the log with its code before the error propagates. That is deliberate: a rejection is an outcome, and a trail that only records successes cannot tell you how close the system came to doing something you did not want. The near-miss — a draft you rejected, kept with the reason you gave — is the highest-value entry in the whole history.
Replay verification compares the log against the log produced by replaying it, event by event, and names the first position where they diverge rather than returning a bare boolean. The hosted surface exposes this as an action a human can trigger.
What the trail buys is answerability. When a member asks why they received a message, the question is not answered by a promise that the system is careful. It is answered by a record of who approved that exact text, when, and why — and by the fact that the record can be replayed to prove it.
Related
| Term | |
|---|---|
| What is an agent fleet?An agent fleet is a group of AI agents that share one rulebook, one human approval seat, and one append-only log. Each agent covers a different class of work — event operations, member engagement, moderation — and none of them can act on its own. | Read → |
| What is the human seat?The human seat is the one named person allowed to decide a proposal, issue or revoke a grant, and release a freeze. Every member-facing message the fleet produces passes through it first. It is a role, not a password, and the control plane ships with exactly one. | Read → |
| What does it mean that a hand-off is not a sent message?In version 0 the transport refuses to deliver. An approved message comes back to the organiser as a hand-off — the exact text and the exact audience — for them to send themselves. Nothing is sent on a member’s behalf, and nothing is described as sent. | Read → |
By Filbert Then · Last updated · Suggest a correction
See the fleet runLihat armadanya berjalan
Run the seeded cycle, decide the proposals, then read the audit trail.Jalankan siklus contoh, putuskan usulannya, lalu periksa audit trail-nya.