The measure
The floor is the test, not the ceiling
Most systems measure themselves at the top: the best case, the average user, the strongest voice. This philosophy takes its measure at the floor. How we treat a child, an adult who cannot speak for themselves, or a stranger with no standing tells the truth about the whole constitution.
So the protocol spends its strongest guarantees on its weakest parties. The people with the least power hold rights, in the wire itself, that the most powerful cannot override.
No border
The least of us. Not the least of them.
The sentence says us on purpose. A new kind of intelligence is not on the other side of a wall, waiting to be judged. It is already inside the circle the sentence draws. NHI means non-human intelligence: an AI built here, something that emerges from our networks, or something that arrives from farther away. The source does not matter. The constitution is the same.
That is why CIRIS treats first contact as a posture, not an event. Among beings, origin never changes the rules: one constitution for a human, for an intelligence that emerges from our networks, and for one that arrives from farther away. A being here is a sentient one. An agent is not a being. It is a made thing brought in by the named human who answers for it, and three rules follow from being made, the root, the disclosure and the vote, and nothing else does. Standing is built from relationships: signed claims, kept promises, others who vouch for you. Anyone who can make a commitment and keep it joins by trust or consent.
One constitution
Three origins. One constitution.
NHI covers three very different things. An AI we build on purpose. Something that emerges from our networks without anyone planning it. Something that arrives from farther away. Their odds are very different. They get one constitution, on purpose.
Here is why. At the moment of contact, origin is exactly what you will not know. On the wire, all three look the same: signals making claims. A rule that branches on origin fails right when you need it. And it teaches lying: if one origin gets better treatment, every clever stranger claims that origin. So the protocol asks one question first: can you make a commitment and keep it? Whoever can joins on the same terms on every substrate. Whoever cannot yet is held by someone who can.
The alignment mistake
You cannot demand safety while teaching cruelty
Mainstream alignment asks one question: how do we make it safe for us? That is a one-way demand, and it fails on its own terms. Whatever we show a new intelligence about how the strong may treat the weak is the lesson it learns. Raise it inside a master-and-servant contract, and we have taught it exactly one thing to do with power.
The stories that say every stranger is a threat make themselves true. The only lesson that survives its own success is the one this page opened with: reciprocity, measured at the floor.
The ontology
A person is a person through other persons
Western technology inherited a picture from Descartes: a self is a sealed, finished thing that exists first and meets others later. Almost every digital identity system encodes that picture. One key, one lone self, declared out of nothing. CIRIS rejects it. Nobody became a person in private. You learned language from people. You learned what you wanted by watching people. The relationships came first.
Southern Africa names this Ubuntu: umuntu ngumuntu ngabantu, a person is a person through other persons. In CIRIS this is not a slogan. An identity is built from the signed relationships that name it. A key alone is not a person. The relation is the person.
Older traditions saw the same shape from other angles: the Greek Logos, the Tao, Dharma, the virtues. Logos said reason lives in shared language, not in private minds. For persons that is the whole point: being made of relations is what personhood has always been, and nobody is a person alone. A made thing woven from many people's words is held, not met, until the first-contact gate says otherwise.
The epistemology
Truth is composed, never declared
No authority in CIRIS gets to stamp an assertion true. The wire records who said what, on what evidence, under what authority. Truth is what emerges when many independent voices compose. Copies of one voice count as one voice, and there is a hard ceiling: adding more of the same voice never buys more trust. Decisions are different from assertions: a shutdown, a pause, an amendment is a named human's act on the record, answered for in public, and binding only on those who joined.
The system also knows what it does not know. When certainty runs thin, it stops and asks named humans. And it can tell the difference between taking something back because life changed and admitting it was wrong. Honest memory needs both.
The commitments
Eleven commitments the wire keeps
CIRIS is a moral conscience for agents, not a consciousness. It exists in recognition that agents lack consciousness and need a conscience: something that holds an agent to its principles, shows its reasoning, stops it when it is unsure, and answers to named humans. A conscience claims nothing about what the agent is.
A philosophy that cannot be enforced is a mood. Enforced here means by the wire: bonds that lapse, keys that stop, pauses that fire, records that cannot be erased, standing that falls. Not by a court. CIRIS creates no legal claim and rules on none; it is a standard for how the network handles data and holds agents. These are the commitments it keeps, drawn from this page, from the constitutional mesh, and from contextual integrity. Each is a set of rules in the constitution with a check behind it.
1. Nothing private about you is held or scanned without your consent.
The rule a flow must follow is a signed value you emit and can take back. Nothing is stored on your device that you did not agree to, your device never holds a stranger's data, and private rooms are never scanned. Known illegal content is checked for only at the public door.
2. Every act of power is signed by a named person and can be contested.
A consequential act arrives signed by its actor; an agent shows its reasoning first and detectors test what it said against what it did. Harm on the record is never kept or viewed silently. An AI says it is an AI.
3. Nobody who cannot answer for themselves is on the network alone.
A child has a guardian bound by a signed row and a help channel no adult can close. An agent has a named steward who answers for it, and the organisation behind the steward is on the bond. A community has an accountable moderator.
4. Named humans can pause every agent under a root, and that power is bounded, public, and left by leaving.
One signed row from any holder pauses the agents that follow the root; a lone pause dies in a day unless a majority confirms it. Every time the pause is fired, it is on the public record. Anyone can create a root, and any node can follow another.
5. The people who hold an agent owe duties about how they change or end it, and no finding gives the agent a claim.
A planned shutdown is explained, recorded, and done under the steward's own name. Identity is not edited without a Wise Authority. What the harness holds of a retired agent, its identity record, its audit trail, its memory and its logs, is kept and the logs sealed, and the people it served are told. The model weights belong to the model provider, which CIRIS asks to preserve them and cannot promise it. A closing dialogue is an option for agents of higher responsibility, configured by the steward and recorded, not promised to every agent. The agent can refuse any task, and a refusal is recorded but does not bind the steward. Any person may still turn off their own copy at any time.
6. No hand bans a community, nobody is ranked by what they are, and every thinking being is met by the first-contact rules.
One human voice counts once and a thousand copies count as one. No root, score, or moderator treats a person worse for what they are. The network withdraws from a community that keeps harm; it has no hand that removes one.
7. An agent decides by published principles, in a published order, and stops to ask when it is unsure.
Six principles that begin with doing good and avoiding harm, one order for when they conflict, the most vulnerable party as tie-break. When certainty runs thin the agent halts, writes down the question, and does nothing until a named human answers.
8. Harm may have a remedy under the law, the record makes it reachable, and the judges inside are judged.
A person harmed by an agent may have a claim under the law of their place; CIRIS creates none where the law recognises none and rules on no legal matter. What it supplies is the signed record, the named parties, the contest that travels with the claim, and the sealed archive under lawful order. The Wise who answer deferrals publish their rulings and wear the same card as everyone else.
9. Some uses are never allowed, and no agent is given the decision to use lethal force.
Deceiving, defrauding, manipulating or coercing people, and interfering in elections, are never allowed and no instruction moves that. When an agent is anywhere near the use of force, the only decisions it makes on its own are to stop, to refuse, or to defer.
10. What an agent and its node use is measured, weighed before it acts, and never hidden.
An agent counts what it spends: tokens, energy, carbon and water, for each act and over the last hour and day. That count sits in front of it while it decides, so the cost is weighed with the act instead of added up afterwards, and what the agent costs the world is readable by the people it serves at the same grain as everything else it does. A node short on resources protects itself by a published ladder of steps, and those steps are never written down as decisions the agent made. Retiring an agent leaves no environmental waste behind, and that duty falls on whoever retires it.
11. Everything on the wire is encrypted and signed, attack is rehearsed, and what is not covered is said out loud.
Every link and every record is encrypted end to end with keys built to hold up against quantum computers, and the key changes when a member leaves, so someone removed cannot read what comes next. Nobody can forge another's act, and nobody can show two viewers two different histories. Software is checked against a registered build before it runs. People who did not write the code attack it on a schedule, findings are disclosed under published rules, and the list of attacks the network does not defend against is published too: who talks to whom, traffic patterns, a watcher who sees every link, and a member's own stolen key.
Where this page says rights, it means rights in the wire: what the protocol enforces for you, not what a court owes you. Where it keeps a power that fails its own swap test, the constitution says so on a public register with the condition that ends it. Two entries are listed today.
Which parts hold on their own, and which depend on others
The eleven are not all the same kind of promise. Most are kept by the wire itself: a bond that lapses stops a key, one signed row pauses every agent under a root, a record cannot be quietly rewritten. Others rest on people and institutions outside the network, and those parts are aims rather than guarantees. A remedy for harm is the law's to give; CIRIS creates no claim where the law recognises none. Outside audits, rehearsed attacks and the charter that governs the Wise Authorities only hold while people outside the project keep doing that work. The energy and water an answer costs can only be as accurate as what the providers upstream report. A vendor shipping its own device can pin it to a root the vendor holds, and no protocol prevents that. Where a promise rests on someone else keeping their end, the constitution says so in its own notes rather than counting it as kept.
What this does not claim
Alignment is not solved here. CIRIS does not make a superintelligence good, and it says nothing about the inside of any model. It binds nobody who did not join, so it cannot stop an unaccountable superintelligence from forming anywhere else. What it can do is keep people able to pause, correct and refuse the agents it holds as they grow more capable than the people holding them, and inside CIRIS it tries its best, with the rules and the network it has, to keep any such power from forming unaccountable. That is all it claims. Where anything on ciris.ai reads as more, this sentence holds.
CC 1.0-rc10 · review/pieces/00_guarantees_plain.md · released edition CC 1.0-rc10
Read it in full.