The Aggregated Intelligence Tenets

The thesis

The most consequential intelligence in the world has never lived in a single mind, human or otherwise. It is composite: the surgeon and the team, the captain and the crew, the pilot and the air-traffic system. The performance that matters is the performance of the whole, and the whole is not the average of its parts. It depends on how the parts are arranged, how they hand off, how they recover from each other's mistakes, how they make each other better than they are alone. Aggregated Intelligence is the collective output of different intelligences working together toward a clear intent. Today, that means biological and artificial intelligences. Together they can produce far more than they would apart, or far less. It is not artificial intelligence and not human intelligence; it is the intelligence that emerges when humans and machines, and humans and humans through machines, work together in ways that are well arranged. The authoritative statement of the concept is the executive brief Aggregated Intelligence; these tenets are the design principles beneath the standards that practice it.

The prevailing agent-orchestration paradigm assumes the opposite: intelligence is a scarce resource located in individual models, and the design problem is allocation. Route the expensive mind to planning and review; route cheap hands to labor. This is scientific management (Taylor, 1911) reborn: separate planning from execution. It works for labor and fails at judgment under uncertainty, because one mind directing many hands is one epistemic point of failure, amplified. A system can prove every task passed and still have no way to prove the task was worth doing.

Turnfile and AIDR are built on the composite premise. The design problem is arrangement, not allocation. They are not orchestration competitors. They are the decision layer above every orchestrator, and these tenets are what that layer believes.

The deepest ground: there are no individuals, only arrangements

The composite premise is usually defended by pointing at teams. It goes deeper than that. Even a single human mind is an arrangement of competing parts, not a unit.

We experience ourselves as one seamless self deciding. We never were. The best current models of the brain describe coherence assembled on top of multiplicity: global workspace theory (Baars, Dehaene) has many specialized processors competing for access to a shared broadcast, and what reaches experience is a winning coalition while the rest is suppressed; predictive processing (Clark, Seth) treats experience as a controlled hallucination laid over the substrate; the split-brain interpreter (Gazzaniga) is a module that confabulates a coherent story after the fact. Binocular rivalry shows it in a lab: give each eye an incompatible image and the brain does not average them, it picks a winner and suppresses the loser, and alternates. Minsky named the whole picture in 1986: the mind is a society of agents.

So the arrangement is not a metaphor borrowed from human teams and applied to machines. It is what a mind already is, one level down. Intelligence lives in the arrangement, all the way down.

There is one crucial difference between the brain and the systems we are building, and it is the difference that justifies the work. The brain hides its committee. The interpreter papers over the seams so convincingly that we spend our whole lives believing we are one thing. That hiding is efficient, and it is also the source of our worst failures of self-knowledge: we cannot see our own biases, we confabulate reasons for decisions we did not reason our way into, we mistake a suppressed dissent for an absence of dissent. The self is the original synthesized answer. Turnfile and AIDR are anti-confabulation infrastructure: they build the composite while keeping the committee visible.

The eight tenets

1. Intelligence lives in the arrangement, not the node. The unit of design is the ensemble. Protocol, roles, lanes, turn order, and handoffs matter more than which model is smartest, because the whole's performance is a property of its structure. This is why Turnfile is a protocol and not a runtime: the arrangement is the product.
2. Disagreement is the engine, not the exhaust. A composite exceeds its best member only through structured dissent. A group that cannot disagree is not intelligent; it is redundant. Counter-recommendations are first-class, objections are never deleted, and a record without visible dissent is presumed incomplete, not clean.
3. Independence before influence. Positions form in isolation before anyone sees anyone else's. The first voice anchors the room; correlated error is the killer. Distinct model families are required for the same reason: a monoculture cannot produce honest dissent, only echo with variance. The linter enforces this mechanically: positions recorded before arbitration, two or more providers, or the claim does not light.
4. Consensus is evidence, never authority. Agreement can be groupthink at machine speed. Products that surface disagreement and then evaporate it into a synthesized answer produce a comfortable narrative wearing the costume of rigor. Honest measurement over comfortable narrative: preserved dissent is the honest measurement of a decision.
5. Authority is human because accountability is human. Not because humans are smarter; we are the more jagged ones, and slower besides. Arbitration is human because of the gap between what we ask for and what we want. Models optimize what we asked; the human arbiter is the only place what we actually want can re-enter the system. Authority belongs with whoever must live with the consequences. No model may author an Arbitration section.
6. The record is the relationship. There is a narrow window in which humans and machines are mutually legible. Plain-text records of who proposed, who objected, and who decided are legibility infrastructure, and each one is practice in a shared language that survives vendor churn. Execution logs answer what ran. Only the record answers who decided, and by what right.
7. The standard is a commons or it is a leash. If a vendor owns the collaboration layer, we get what they ask for. So: open spec, permissive licenses, no field naming a required provider, no hosted oracle, verifier-anywhere. The answer cannot belong to any one person or company. That is not idealism; it is failure-mode analysis.
8. Rules are earned, not decreed. Every rule in Turnfile came from a real session; the inception archive is a protocol writing itself through practice, retrospective by retrospective. Standards that outrun practice become armchair law.

Status

The eight tenets above were ratified as canon on 2026-07-06 after an independent multi-provider peer review, with positions formed in isolation before arbitration, per tenets 3 and 5. The source document contains two further claims (a facilitation lineage and a two-failure-modes framing) that remain provisional and are deliberately not published here until they clear the same bar.

Per tenet 8, this page is a time-stamped assertion, not a permanent verdict: the tenets were earned from practice and stay open to it. If your practice contradicts one, that is evidence the record wants. Contest it: github.com/snapsynapse or hello@paice.work.