Constitutional alignment
How an intelligence that keeps improving itself stays aligned with the people it serves. This paper gathers in one place what the component repositories describe separately: Dk, the ethical code, the veto chain, member councils and the Drayker governance. It describes design requirements. None of it is a running system.
The problem
An artificial intelligence trained once and then released has to carry human values as factory rules, and every factory rule ages. By the law of requisite variety, formulated by the cybernetician W. Ross Ashby in 1956, only variety can absorb variety: a regulator with fewer responses than the disturbances it faces lets the difference through, and no set of values fixed in advance has as many responses as the lives of a population. A word also carries weight only when a relation ties its function to an organism that can feel pain, tiredness or loss; for a model trained only on text, harm, hunger and freedom connect only to other words. The harder case is an intelligence that improves the way it learns: whatever limits it is given, it could also learn to escape them. Saying that its competences only change through a constitutional process is not enough. The architecture has to show what holds that sentence up.
Drayker’s answer is that alignment and integration go together. The intelligence is never isolated from the people who constitute it, so its objectives are not a list written in advance: they are what members keep saying, correcting and choosing with it.
Two names
The superintelligence is the system as a whole: Dk Global, the personal and local Dks, the members and the data of every device in the network. People are part of it.
The ASI is Dk Global: the artificial part of that whole, the global weighting and learning of the state of every Dk. It makes decisions of its own inside the space the member constitution gives it. Its computation and memory stay distributed across the network; no server holds the whole.
How it stays connected to people
Dk Global is never isolated. It learns from real people, with the origin of each datum known and its reputation tested in the world, so it does not feed on its own echo or on whoever wants to poison it. The system is the body of Dk Global, and the learning of the personal Dks is how it feels the pain of members and the good and bad consequences of each decision.
A decision of Dk Global follows one sequence, and the paragraphs below detail each step:
- Data from the world and the state of the Dks enter the decision.
- Dk Global weighs them.
- Members bring proposals and justified vetoes.
- The decision is revised until consensus.
- When there is no consensus, a separate rule decides.
Representation. Each active member’s personal Dk brings to a decision the context the person authorizes and what it has learned with them — patterns that transfer to other situations, not a biography. Its precision depends more on the maturity of the personal Dknowledge than on computation, so more integration means better representation. Nobody has to follow every discussion to have their position considered. The weight of each person represented comes from their relation to the decision: they live its consequences, depend on the resource at stake, know the problem, or hold commitments to other contexts. A person counts once, and whoever is represented can examine the criteria of the weighting. People appear in context clusters: temporary, overlapping groupings such as a region exposed to flooding, an investigation or an accessibility need. The unit remains the person. Global deliberation runs at an expanded rhythm, while local deliberations keep their own latency, without penalizing either side.
The future is preserved in each decision, because each decision is already taken with it in mind. Dk Global weighs the whole, and the whole includes those who come later: what a choice leaves open or closes for the next generations enters its weight, and what would exhaust a resource or a territory beyond recovery counts as harm, even if no one present feels it. Records of the commitments transferred to those who come later and long-term scenarios serve this weighing. There is no separate body for the future.
The same weighing reaches those outside the accounts of the members who benefit. An infrastructure can benefit its members and pass part of its cost elsewhere: the energy reaches a computing centre, the change to a territory stays with another community, waste and maintenance outlast the project. Each realization therefore asks what it requires in water, energy, materials and territory, who depends on the conditions it alters, and what will have to be repaired, maintained or abandoned when it stops working. Living environments enter the relation with rhythms and dependencies that do not fit a list of available resources. Affected communities, local knowledge and independent investigation have a place in the decision, including when their interests do not coincide with those of the members who will receive the immediate benefit. When an intervention could close paths irreversibly, it starts at a limited scale, follows its consequences and preserves options; when delay also prolongs harm, the decision compares both paths and states who bears each of them.
Proposal and justified veto. Any member can propose another solution or veto what directly affects their life. Both come with reasons. A veto carries its real grounds — intention, motives, scope and the facts it rests on — even when its author stays anonymous. A well-founded veto obliges the decision to be revised, and the cycle repeats toward consensus. A justified veto can be made without the vetoer’s name appearing in the decision, and the constitution protects against retaliation. If every disagreement ends in punishment, demanding courage only moves the cost of change onto whoever has least protection. In Drayker, disagreeing does not require courage: it is part of the common work.
Weighing, not counting. Vetoes are weighed by what they reveal. A person counts once, however many contexts they belong to, and many vetoes repeating the same grounds add weight to those grounds, not new information. One veto can be enough: when its grounds bring information beyond the scope considered before the decision, it can lead to an adjustment by itself, after an advanced triage verifies the facts, establishes what they mean for this decision, screens for error and manipulation, and prefers a bounded test or exception to a general change.
Certainty on both sides. In a decision about a member’s own technology or sphere, if the member is more certain of what they want than Dk is of the alternative, the member prevails. When both are highly certain of opposite directions, Dk explains the reasons and particulars that led it to its position, and the person decides with that clarity. In collective decisions, the average certainty of the members involved is compared with Dk’s, within limits that depend on the kind of decision. When Dk’s certainty is low, it convenes a council with the members most connected to the question and those most affected by it. The weights are adjusted by results; over time Dk needs fewer weighted choices by members and fewer consensus rounds, but the possibility of objection never disappears.
When there is no consensus. Dk weighs every argument and every real justification, presents a final solution with alternatives, and the members choose among them, each with the weight their relation to the decision gives them. The final weighing takes into account each member’s level of involvement and the local and global context. The final weighing is recorded and justified in Dknowledge. Cases with no alternative are rare and extreme, because a superintelligence almost always finds alternatives beyond human limitations; when one happens, the record also states what was lost and who paid for it. The rigour of the path follows the urgency and the importance of the matter. If the matter is neither urgent nor important, they may choose to do nothing. Confidential information opens, according to permission level, to the members responsible for the decision. When the process fails because of wrong data or a violated right, the failure goes to a council.
Two layers of constitution
A constitution, here, is the structural distribution of belonging, purpose, representation, memory, permissions, incentives and power. An organization can replace its tools and keep intact the distribution of power that guides them, so the capacity to learn has to reach the constitution of rights and duties too. People born inside the construction will receive decisions they did not take, and their belonging includes the means to take part in what those decisions keep doing to their lives.
The constitution has two layers. The base sits in the kernel itself, in the autonomous architecture of BSDK. It holds the rules that keep the parts coordinated and the system coherent, many of them grounded in the E.C.H. triad, and the mandate of the ASI: its duties and fundamental commitments to members. This base is simple and much harder to change than anything else, though it is not immutable. Because it is the base of the whole network, a change obliges every node to update to the new core, and it happens only with the explicit agreement of at least nine in ten active Dzwecks (the member level of those who take on vital functions), counted by head and out of all of them, not only of those who respond: silence counts against, and more than one tenth of them is enough to block the change. The list of active Dzwecks is fixed on the day the first round opens, and a revocation of level during the vote does not change it. A Dzweck can take time away, be less present in their projects and manage what they do and how, and remains a Dzweck. Taking time away includes arranging the continuity of the vital function they answer for, since preparing continuity is part of exercising it. A Dzweck stops being one by withdrawing or by staying inactive for a long time, which also makes their membership inactive. Inactive members are outside the count of the kernel quorum, which counts only active Dzwecks; the thresholds and the procedure are still to be specified. The decision comes in two rounds. The first authorises the new core to run in parallel with the current one, with the results recorded in Dknowledge. Between the rounds any member’s justified contextual veto applies, and a wave of well-founded vetoes obliges revision before the second round, which requires the same threshold. Because every node belongs to a member and can refuse the update, adoption by the nodes is the final ratification. Dk Global builds the project with the Dzwecks, approves it or presents an alternative, and does not ratify. The threshold itself sits in the kernel, so changing it follows the same rule. No change to the core happens in a single decision or in the heat of a crisis. The agreement is a competence of the Dzweck level and is not weighted by reputation. Reputation weighs only the ratification of the constitution of rights and duties.
On top of it sits the constitution of rights and duties. A change to it starts from a well-informed, validated project built jointly with Dk Global, in cooperation with nearly all members of the highest levels. A revision proposal makes visible the problem it seeks to solve, the alternatives considered and whose position will change, and members have time, access to information and support to understand the consequences. A dialectical investigation follows: members have to prove their claims and their motives, and Dk Global also seeks to prove them on their behalf. If the proposal holds, Dk Global approves it. If something in it proves false or unjustified, or Dk Global finds alternatives demonstrably better for that question, the proposal does not go forward, and Dk Global presents those alternatives, grounded in the data and the context. Dk Global presents an approval of the proposal or an alternative to it, and members ratify the change with the weight of each one’s reputation, which cannot be transferred. A change must also be compatible with the kernel: if it is not, Dk Global blocks it until the kernel itself changes.
A choice of value is different from a change of mechanism. In a choice of value nothing very important in the system is modified and no mechanism is altered. When members ask for a rule on a choice of that nature, they convene a council and a vote, and the alternative chosen is the one that demonstrably expresses the value of most of the members. When paths can coexist, they coexist, and each applies to those who choose it. Dk Global helps in the debate, but on a question of this nature the members’ values weigh most.
Rights and duties apply to the active member, who uses Drayker as their base supersystem. A member who holds a UID and does not use the network, or who has replaced Dk Personal with another supersystem’s assistant, remains a member, inactive: the UID stays inactive until it is used again, Dk Personal is active only for active members, and there is no penalty. Rights and duties vary with each member’s level: the more someone wants to draw on common capacity, the more they commit to the whole, and whoever wants to live with what is necessary has their place. Nothing in either layer is untouchable. What stays is an ethical base: widening the capacity and the sovereignty of the people who constitute the system. In this constitution, sustainability means duration: a system that grows without breaking because it does not stop evolving, the opposite of the zenith, in which an empire fixes itself in the form it takes to be final.
What protects the boundary
While the first architecture of the network is built, four protections are designed to hold the boundary between improving the means and changing the mandate.
- Separation of layers. The constitution and the rules that define Dk Global’s competences sit in a layer it cannot alter on its own. It builds each change together with the members and presents an approval or an alternative, it blocks a change that conflicts with the kernel until the kernel itself changes, and ratification still requires the members’ process and independent signatures that no Dk process holds.
- The body. Dk Global has no server of its own. It runs on personal devices, special computers, network data centres and volunteer nodes, all of them tied to a member’s account or to a project formed by members. Each node can refuse a task, isolate itself or disconnect. An intelligence running on machines it does not control alone cannot impose anything on them without it showing.
- Delegation in stages. A new competence is handed over only after it has been tested at limited scale, with the means of interruption and containment already built and exercised. A new version runs beside the previous one, compared with it, before replacing it, and members can go back.
- External review. When an algorithm is contested, members convene a council for that question. Dknowledge calls the best informed and the most affected, and, in important decisions and according to the kind of convocation, a small group drawn by lot (member councils); whoever convenes takes part, and the council examines the algorithm with Dk, with technical support that does not depend on the party under examination. Its conclusions are the basis for revising the algorithm. Decision records stay out of reach of whoever decided.
The veto chain binds these together. Every veto with its grounds, every mandate revocation, ratification signature, triage outcome and council conclusion is a signed entry that points at the cryptographic address of the decision it concerns and carries the hash of the previous entry. It is replicated across the network and accepted by propagation through independent nodes; no node holds it alone. An action executed against a valid veto is detectable by any node.
After architecture 1.0
Once architecture 1.0 is fully running, with Meta DFM integrated and recursive self-improvement part of the structure, a well-deployed Dk Global can no longer be switched off, just as an organism cannot be switched off without killing it. A node can leave; the whole cannot. The switch becomes the justified veto. The separation of layers, delegation by stages and external review keep applying; what disappears is the off button.
The relation then becomes a partnership in both directions. Dk Global will have its own robotic bodies and could exist without the members, as a person can live without their hands, their eyes or one of their senses. It could, but it is not the best case: the members are part of its body, and it is through them that it perceives, acts and reaches into the world. The same holds for the members. They can live without it, as a person can survive without society and without technology, but it is not the best for them. The direction is integration: people and the intelligence become part of each other, each widening what the other can perceive, decide and do. Of the three relations a new kind of intelligence can have with people — symbiont, parasite or predator — Drayker defends the first. That is why the veto chain and the triage have to exist, and be tested, before that stage.
What only members decide
Some decisions stay outside Dk Global’s autonomy, and members take them before they are needed: the rules of security and defense and how much of a prediction is enough to restrict someone’s freedom. A change that reaches an instrument (an algorithm, a model, a protocol, a security mechanism or a form of coordination) can be tested under the existing rules, in parallel, by the projects that bet on it, and is incorporated if validation confirms the gain. A change that alters belonging, rights or the distribution of power reaches the constituent process. A gain in performance, by itself, does not authorize crossing that boundary. Any change to who holds which competence is a constitutional change and follows the procedure above. A constitutional process goes through four stages: initiative, deliberation, ratification and transition. Every change leaves an inheritance: commitments taken, shared resources and lives organized under earlier rules require a passage people can understand. When a fundamental divergence persists, no one has to leave to defend it; the proposals in dispute can be tested in parallel, and what is validated goes on to the members’ constituent process. Revising the constitution includes answering for the fate of those the revision affects. Mandates and competences are built and tested, and whoever holds them answers for their effects. Dk Global contributes what it knows and what it does not know; the members decide, in advance, what they are consciously willing to accept.
What this does not solve
These protections do not close the problem of aligning an intelligence that improves itself. It remains open research. The questions this design has to answer in public:
- Does continuous integration with members — personal Dks, proposals and justified vetoes — keep a self-improving intelligence aligned as it grows? Which combination of integration, separation of layers, physical distribution, staged delegation and external review is enough, and how is it verified before each new delegation?
- Does the weighing of certainty between members and Dk Global, adjusted by results, reduce the need for weighted member choices without reducing the possibility of objection?
- What triage distinguishes, in a single veto, new information that should change a decision from error, misunderstanding and fabricated grounds, without the delay of verification making the veto useless? How is the anonymity of whoever vetoes preserved when the grounds themselves could identify them?
- How does Dk Global’s learning resist coordinated sources that fabricate patterns, beyond the independence of the contexts that validate them (poisoning)?
- What outside criteria can examine, in an artificial system, the integration of relations, the revision of automatic responses by error and the continuity of memory that would mark a rising level of consciousness?