The name is a claim about the system
AGI²⁺ stands for human-sovereign social hybrid intelligence. The name is a claim: that in consequential work, intelligence is a property of a governed arrangement of people, machines and persistent institutional state, and a model on its own is one component inside it. The squared term counts the human side twice, the person and the group, and the plus is the machine added to them.
Where the method comes from
The ordering discipline is old. Decision analysis has sorted policy questions from tactical ones since the 1980s. Assumption-based planning filters for the assumptions a plan actually rests on, and question decomposition has been routine in forecasting work for about as long. All of it worked. All of it needed a skilled facilitator in the room to pull a usable frame out of a messy real situation. That is why the method stayed inside consultancies and rarely reached the people who were making the decisions.
The extraction step is now close to free.
So the honest version of the claim is narrow: the novelty sits in the starting cost. A leadership team can begin from the material it already has (a memo, a board paper, a partner proposal) without booking a facilitator first, and the first thing it sees is a completed strawman to correct rather than an empty page.
What the theory adds on top of the lineage is the runtime: ten states, S0 to S9, that carry a decision through frame, hypotheses, question, evidence, belief, readiness, decision, action, outcome and learning. Each state has a condition it must pass, so an action does not count until it is authorised and an outcome does not count until it has been observed. A late reframe rewinds the states that depended on the old frame instead of quietly leaving them in place. Those ten states are Layer 1 of the current architecture. In v23 they were the whole runtime; today they are one of three layers.
Who
Pertti Jalasvirta wrote the theory. The current architecture is AGI²⁺ v26.0, dated September 2026, and it carries a version lineage in which each release records which failure mode it repairs and which earlier ideas were dropped for want of evidence. The failure mode this release repairs is the one the site used to leave open: what holds a decision together in the months after it is taken.
Andres Pihor is the tech lead and AI expert on the project. He builds the runtime and the tooling the engagements run on.
The architecture
Three layers, running together
Long-horizon intelligence emerges from a governed human-machine system that preserves and revises its institutional state across time.
The layers are not stages. A consequential decision touches all three at once.
Layer 1 — Decision cognition
Layer 2 — Human sovereignty and governance
Layer 3 — Persistent institutional state
Status
v26.0 is built around a hypothesis that can be tested and can fail: that long-horizon intelligence needs more than memory, and requires persistent institutional state, explicit commitments, falsification, and a way to notice when a mechanism has drifted or an actor has adapted. Nothing published yet establishes it.
The method has been demonstrated on a real decision. The system that runs it is still being built, and every engagement is run as a specification exercise for it, under a rule that anything the facilitator does twice gets automated before the third engagement. The engagements are the specification, which is the plain way of saying that the software is not finished and nobody here will tell you otherwise.
Last updated 8 September 2026