All Intelligence is Collective Intelligence

Culture begins with signals that people notice, interpret and contest. Taste points shared attention toward what might matter, while valence marks the pull, fit, novelty, belonging or refusal a signal carries for a particular community. Trust compresses a history of situated validation so nobody has to verify everything alone. Together they shape which experiences enter collective consideration, who is prepared to act on them, and what context must remain attached for that judgment to remain intelligible later.

Consensus coordinates action, but it does not turn agreement into truth. Every act of normalisation preserves some distinctions and erases others, and network topology influences the result: who observes whom, whose judgment travels, which relationships confer trust, and where dissent remains legible. A coordination system must therefore represent both convergence and the structure around it. Otherwise a dominant signal can appear universal when it reflects only the position from which the group was measured.

We treat Cultural Intelligence as cumulative social learning. The causal link to the larger thesis is evaluation: compositional systems need signals about whether an arrangement served its participants, and open epistemic networks need reasons to propagate one learning event over another. We gather those signals where judgment is already professional, from panels of credentialed practitioners scoring work blind and from the reviewers a client engagement has to satisfy before anything ships. The research challenge is to make taste, trust and disagreement inspectable without flattening them into a single score or treating engagement as a substitute for judgment.

At runtime, one model can call other models, symbolic programs, tools, stored memories and a person. The graph also decides when each component acts, what context it receives and where its output goes. No single part contains the resulting capability, because the useful behaviour depends on an arrangement that crosses their boundaries. This is the software form of collective intelligence: a temporary coalition of specialists whose coordination can be inspected, altered and evaluated as a whole.

The search space is therefore the whole graph: routing, control flow, memory, evaluators and the models themselves. We vary each element, run the resulting assembly, and evaluate its outcome and trace against the evidence a real task produces rather than an intuition about architecture. The assemblies worth comparing are the ones client work generates: a brief that has to ship, composed from local models, rented specialists and human review. Evaluation supplies the causal bridge between experimentation and learning. It makes different arrangements comparable, shows where coordination failed, and gives a meta-orchestrator the material from which to revise future decisions without pretending that one topology will be best for every context.

When an assembly works, it can become a reusable artifact that another machine or community inherits instead of rediscovering. Some of its behaviour may also be fused back into model weights where that is useful. In either form, the arrangement preserves a record of how capability was produced. The open problem is to keep that cycle adaptable without losing provenance or overfitting it to a fixed benchmark.

The person inside a local loop is the oracle for that experience. They judge the result, decide what is worth keeping, and can interpret files, preferences and consequences that a remote system cannot fully see. The architecture is designed to leave those materials on hardware they own rather than requiring collection in a central corpus. Local authority matters causally: it gives a learning event an accountable source, preserves context that would be lost in aggregation, and lets participation occur without making surveillance the price of coordination.

A node can emit a signed learning event while keeping the raw experience that produced it. A valid signature binds the event to a key, making its integrity and key-based origin verifiable; it does not establish human authorship by itself. Provenance carries the claimed source and context, while accrued trust and network position help recipients decide whether the event should travel and how much weight to give it. When shared, the event is a claim others can test, contest or compose with.

What compounds is knowledge that carries its claimed source, context and the judgment under which it was recorded. Plural subjective memories can overlap, disagree and change without being merged into a canonical state. This creates an open-ended ecology in which communities coordinate through provenance, trust and evaluated consequences while retaining distinct views. The test is commercial as much as technical: a company puts its archive into a loop only when it owns the loop, and lets anything leave only as a signed claim. The challenge is to make enough of that learning interoperable for the network to accumulate capability without appointing a central memory, oracle or authority to decide what the whole system knows.