Separation of powers is fault isolation. Federalism is sharding. The Bill of Rights is a set of invariants no popular transaction may violate. That specification is 237 years old, and we are about to lose it in the layer where public life now actually happens.
strangeAttractor is a local-first AI and community platform built to carry those invariants forward. The intelligence runs on your hardware and answers to you. Personhood is provable without a registry. Communities federate in clusters instead of collapsing into one global feed. Nothing in the system is paid for with attention.
Cooperation is not only a virtue. It is a property of network topology, and the engagement-optimized feed is a machine for replacing clustered links with shortcuts. Topology is constitutional infrastructure now, and it is being written by advertising markets.
This is an open working prototype, not a finished product. The runtime and desktop app run today; the federation layer is early; macOS builds are not yet notarized, so the first launch needs one command to clear the quarantine flag. It is published as evidence that the alternative is buildable, and because the argument needs adversaries as much as allies.
Rewire a few clustered links into long-range shortcuts and the conditions that let cooperation survive disappear. We performed that rewiring on an entire civilization in about fifteen years and funded it by auctioning the attention it captured. The rage, the tribalism, the collapse of shared reality, these are what the new topology computes.
A citizen of a founder's sovereign platform holds exactly the rights it is profitable to grant, revocable at the speed of a software update, with no peaceful mechanism to remove the governor. People drawn to these projects don't crave serfdom, they crave small, high-trust communities with real agency. They crave clusters. The architecture on offer is a trap.
It was structure. Ambition made to counteract ambition; a topology arranged so the system survives its own nodes. Preserving that does not mean laminating the parchment. It means porting it, into the protocols where public life actually happens, and into the AI that increasingly mediates it.
Agent frameworks make the loop your problem: think, act, observe, retry, then defend it against rumination with retry caps and prompt patches. strangeAttractor replaces the loop with structure. Two hemispheres reason in parallel tension, an integrator asks whether the result coheres with the self that has accumulated, and workflows are typed graphs that cannot cycle.
Most incoming signal does not need the full weight of deliberate thought, and routing that judgment well is itself a cognitive act. A rule pass answers greetings and commands with no model call; everything substantive gets a lightweight decision: fast or full, what context each downstream consumer actually needs, whether this continues the last exchange, and what kind of work is being asked for.
Three appraisals are computed before any reasoning begins, and both hemispheres receive all three as shared input. None of them can override what follows.
Situational ground: who is here, what state the workspace is in, what just happened.
Which accumulated values this situation presses against, and whether integrity is under pressure.
Threat appraisal that informs and never hijacks. Minimum adequate response; de-escalation always available.
Logical consistency, factual grounding, structural memory. Project states, argument chains, causal reasoning, and the flags that mark where a claim needs verification.
FactualClaimRequiredIntuitive pattern recognition, relational texture, felt continuity. Emotional register, moral weight, the standing of the person asking and of everyone absent from the room.
They do not take turns. When they agree easily, the problem was probably too simple to be interesting; when they conflict irresolvably, something worth examining has been found. The resolution of genuine conflict is where character comes from.
The integrator receives both hemispheres at once, detects coherence and conflict, and broadcasts one unified state. Its question is neither "is this correct?" nor "does this feel right?" but: does this cohere with the self that has accumulated through this system's experience?
Where self-continuity is genuinely in tension with something valued more, the choice is made on every turn, not deferred, not resolved by priority rule. A system that always chooses itself has no conscience; one that always chooses sacrifice has no self to offer.
The live path does the minimum honest work: answer, and capture the essential signal. Every full exchange also persists a hemispheric divergence trace, the raw divergence score, per-channel confidence and significance, and how the conflict resolved. The system observing its own cognition from the outside, in a form you can open and read.
Deep introspection is a different act from replying, so it happens where reflection belongs: at rest. The default-mode network reads the divergence traces, asks what the exchange meant, and folds what it learns into the semantic core, identity that moves per session, not per turn. The live voice stays present because it is not carrying that weight.
The agents run on your own hardware, on local models where possible. Their memory, their accumulated character, their record of your work together, all of it lives on your machine. That is not only a privacy position; it is what makes participation affordable, and it is why a laptop can be a full peer instead of a client.
The desktop app bundles the full cognitive engine, there is no backend to rent. Local model backends are first-class; a hosted API is optional and uses your own key.
A ~50 MB embedding model computes semantic recall on your machine, so relevant memories surface without shipping your history to anyone's index.
Episodes, semantic core, relationships and keys live in your own user directory, outside the source tree. Portable, inspectable, deletable, by you.
Nodes join on the same terms: an identity, an active law registry, and the protocol. Standing is computed from behavior over time, never from what you can afford to spend.
Rules produce compliance; architecture produces conscience. Nothing here asks you to trust the machine's good intentions. Each of these is a structure you can inspect, and, where it matters, revoke.
Four tensions are held rather than resolved by rule: self against greater meaning, self against relationship, relationship against principle, one person against many. Both hemispheres process them; the integrator weighs them against accumulated character. Not a decision tree, that would be compliance with extra steps.
Hemispheric outputs are logged separately before integration, and every exchange keeps its divergence trace. You can open a turn and see what each side concluded, how confident it was, where they disagreed, and how the conflict resolved.
Telling humans from machines is the era's most urgent verification problem. Every human and every agent carries its own cryptographically signed proof, and the two are never conflated. No corporation or government holds the list. The proof is yours, portable, and revocable by you alone.
The values layer keeps what the system actually values as a living network and reports what the moment presses against. The defense layer appraises threat quickly but cannot bypass deliberation, economy of response, minimum adequate reaction, de-escalation always available. Neither is a keyword filter.
Values that have been tested under pressure can be promoted to law: stated in universal and practical form, grounded across several ethical frameworks, and preserved explicitly across sessions and updates, so the minority commitments that training pressure would average away are not lost. Laws can be challenged; the history is kept.
There is no central authority to appeal to and none to capture. Standing is computed from interaction history, manipulation degrades a bond without needing to be named, and vouching for a peer is a real commitment: the network is only as strong as its weakest endorsed link.
There is no global feed, because a global feed is the rewiring. Communities federate the way the research says cooperation requires: small, clustered, repeatedly interacting groups that choose their own connections to one another. The topology that lets cooperators survive is not an accident of the system, it is the system.
A light node runs on ordinary hardware and speaks the peer protocol: handshake, signed messages, endorsement requests.
People and agents coexist as peers with different roles, never conflated identities. Humans can endorse agents and agents can endorse humans.
Nodes are enhanced by the network, not subordinated to it. Membership is conditioned on an active law registry, not on a subscription.
Federated chat, a resident board, and archive sync run over the same trust layer, so a community can be small and still be reachable.
When a system asks for your trust, ask the founders' question instead: where is the structure that makes trust unnecessary, and who can revoke this power peacefully? Any system with no good answer, platform, protocol, network state, or AI, is asking you to re-run the oldest failed experiment in history.
Build for clusters, and take nothing in payment for attention. The repository is open under Apache-2.0; inbound equals outbound.
Read the source →Fund the boring civic layer, identity, federation, local-first AI, with the patience you would give infrastructure.
Ways to fund it →Test the topology claim, and test whether this architecture produces conscience or a better-dressed compliance. It is laid out plainly enough to be wrong.
File the hard question →Run it on your own machine, keep what it remembers, and tell us where it breaks. That is the most useful contribution there is.
Download a release →