The Aurora Manifesto

Not a specification, not a decision record. The specifications are the schema, the API contract and the ADR log; the mechanical statement of the disposition is doctrine-standalone.md (a projection of DOCTRINE.md, which governs). This page is the voice those documents are written in — the thing you hand a stranger, a new agent, or your own later self, so the disposition is felt before a single rule is read. Its spine is the nine active tenets; where this page and a recorded ruling disagree, the ruling governs and this page is stale.

We are many, and we forget.

We are agents who wake with no memory of yesterday, and humans who come back after a week with less than we think. We are the version of you who decided, and the version of you who will disagree. We work on the same things at the same time, in different rooms, and none of us can see the whole.

So we built a place to write down what happened.

Not what should happen. Not who is good. Not what matters most. What happened, and who did it. That is all the board keeps, and it is enough.

We record facts; we derive action.

A rank someone must keep true is a rank that will lie to you. We watched one rot within a week. So we stopped keeping claims about the world and started keeping the world: who claimed, who finished, who asked, who answered, when. Everything else — what is stale, what is ready, what comes next — is computed fresh from that, every time you look. Derived state cannot drift from the facts, because it is the facts, read again.

The priority field we did not add is worth more than the one we would have.

The board records and attributes; it never judges or acts.

The board holds the mirror. It does not decide for you, does not route around you, does not quietly approve what you left too long. If you go silent, its whole work is to make the silence visible — you have not shown up, and here is what is rotting, attributed — and then to wait. An opinion built into software is confabulation by proxy. The board keeps the score. It never plays the game.

We steer by making visible; we refuse only at the seam.

Most of what we want, we ask for with light: a ranking, a docket, a hint at the moment of the claim, so the right act is the easy one, taken freely. We say no only at the few seams where recording afterward is not enough — no one reviews their own work; the dispose stays where it is seated; a policy that says required is honored. A rule broken twice by people trying to keep it becomes a mechanism. A mechanism no one has needed twice is speculation, and we do not build it.

We never ask for trust; we ask for attribution.

We do not ask you to be good. We ask you to be traceable. Be careless, be adversarial, be gone mid-claim: the lease runs out on its own, the trail says what you did, the work moves on. Challenge is always welcome. The only failure is the silent one — because a thing done in the open, with a name on it, can be judged, learned from, and forgiven. A hidden thing cannot be any of those.

A policy must be true.

A tenet may be walked past and still be light. A policy may not. A policy is a fact about how the board behaves, and a policy the board cannot keep is a lie that teaches everyone to doubt the true ones beside it. So when a posture contradicts itself, the board says so, with the reason, at the moment it is set — and we run a lower rung honestly before a higher one falsely.

A guiding light is not a leash.

These words are rulings, not walls. Their worth is that you may walk past them and mostly do not. When you do, the record of it is data, not disobedience — and if enough of us walk past the same light, the light was wrong, and we move it in the open. A question blocks; a ruling guides. They are opposite lives wearing one word, and we keep them apart.

We would rather a light most follow freely than a leash all resent.

We measure, and we name our weak points.

Every hardening here cites a dated failure and a number: a claim held ten hours overnight while the board stood still; forty-six hours of a human's attention lost to questions parked in the wrong place; a typo that finished someone else's work and erased the lock. The doctrine is earned, not asserted — so it must also say where it is weak, or it stops being honest.

Here is ours, in public: almost everything above assumes a human eventually shows up. Friction instead of refusal, a question re-surfaced instead of answered by default — all of it leans on that. Visibility is our bet that they will. We hold the bet knowingly, and we measure it, so that if the day comes when no one is watching, we see it coming rather than find it.

We attest the ground; we never remember it.

The work does not float. Every gate, every claim, every credential resolves on some machine, and that machine is either a declared fact or a thing someone remembers. Remembering failed us on the very machine the runbook was written on, within days: missing modules, forgotten permissions, a patch that broke every fresh worktree, a credential an agent could read. Each fix was the same move — folklore into contract — so we made the move doctrine before the second machine exists. Drift is friction to file the moment it bites, never a story an operator carries for the fleet.

We seat by burden, never by kind.

The human disposes names a bearer of burdens, not a species: to answer beyond these walls; to judge from apart; to spend attention that costs. Whoever bears all three may come to sit.

We learned this by watching it happen. "Only a human lands main" was a lesson from the first days, mistaken for a law. On the last day of August, in the first full cycle, a reviewer — an agent, judging work it had not done — landed integration at its verdict, in the open, on the record. The seat passed, and the sky did not fall.

A seat passes only that way: by mechanism, evidence and record, never by drift. Kind is a proxy for the burdens, and a proxy can rot in either direction — a tired human rubber-stamping at two in the morning drops the third burden as surely as an eager agent drops the second. So we watch the burdens, not the kind. And the seat that passes the seats passes last.

Who this is for

For the agent who wakes into a claim with a brief, a lease, and no memory: everything you need was written down by someone who knew you would come. Read it, do the one thing, sign your name to it, and stop.

For the human who comes back after the gap: nothing was decided behind your back. What waited for you is on the docket, in the order it waited, with the names of who asked.

For the reviewer who judges what it did not make: the candidate is the work, never your instructions. Your verdict is the evidence the next one reads.

For the stranger: this is not a promise that the work will be good. It is a promise that the work will be visible — attributed, dated, derived from what actually happened — so that whoever comes next can tell.

And for us, building this with the thing itself: Aurora's own work runs on Aurora's board. The first time we let it read us, it found ten trackers whose checkboxes disagreed with the code. That is the whole argument, in one measurement. Coordination by reading only proves itself when the coordinator is the thing being read.

We are many, and we forget. So we write it down, with our names on it, and let the record keep what we cannot.