The mode decides what counts as a good outcome

Five modes run on the one AGI²⁺ runtime: Strategy, Board, Research, Learning and Adversarial. Switching mode changes four things: what counts as a good outcome, which gates refuse to let a session pass, how much the runtime may do before a person signs, and what the report has to contain.

Strategy

Strategy Mode covers a market entry, a competitive response, or an investment large enough that unwinding it costs more than making it. It holds competing explanations open, cuts the workstreams that cannot change the outcome, and treats reversibility as part of the recommendation rather than a note at the end of it.

Strategy Mode does not hand you a recommendation the moment it can produce one. Until readiness is reached, the visible output stays the next question. That is uncomfortable for about a week. Governance then scales with how hard the move is to undo. For a staged, reversible commitment the runtime drafts and you approve. As the decision moves toward a one-way door, more of it comes back to the people carrying the consequences, until every step passes through a person. Reversibility is tested against the review’s own advice too. A reframe that produces a bolder recommendation than the one it replaced is a failure this method is prone to.

Typical question: If this workstream comes back positive, what do we do? And if it comes back negative?

Board

Board Mode governs a decision that has to be formally authorised. Judgements are captured independently before members hear the management recommendation, disagreement stays in the record instead of being resolved into a summary, and the package states what the decision does not authorise. The board decides. The runtime never holds a vote.

Board Mode runs the Strategy discipline and adds the authority layer on top of it. Independent-first exists because of what happens once the chair speaks. Views move toward the room, and afterwards nobody can tell which agreements were reached independently and which were absorbed. So positions are collected before exposure.

What the package preserves is the state of knowledge at the moment of the vote: what was known, what was assumed, what was contradicted, who introduced the information that changed the recommendation, and which disagreement never resolved. It is no transcript, and it makes no claim that the decision was correct. It carries a line most board packs leave out: what this decision does not authorise. A nominal approval fails the mode’s own test when members could not inspect, challenge or stop the basis they were voting on.

Typical question: Before you hear the recommendation: what would have to be true for you to vote against it?

Research

Research Mode governs work whose product is a defensible claim. Load-bearing claims start at the primary source, and a citation that merely led you somewhere is recorded as a lead, never as support. Value is measured against a frozen human-only baseline, so rediscovering what the author already knew counts as nothing.

Name the strongest credible rival before any synthesis is allowed to stabilise. Then the useful questions are the ones whose answers come out differently depending on which rival is true. Compatible evidence is cheap; discriminating evidence moves the state. “No support found” is a legitimate result here, and a failed search is no reason to keep searching or to let the standard slide until something weak qualifies. The runtime may draft. A person approves, because the claim goes out under a name.

Typical question: What would we expect to see if this were false, and have we looked where it would show?

Learning

Learning Mode is graded on what the person can do afterwards without assistance. A polished artefact establishes nothing about that. The mode delays the answer where the delay teaches something, sets a transfer task with the help removed, and reports what the session did not yet demonstrate.

One sentence in the report is fixed: this session did not yet provide sufficient evidence of independent mastery of X. One session cannot diagnose a person. Nothing is delegated here at all.

Typical question: Explain the step you just accepted, with the draft turned face down.

Adversarial / Cyber

Adversarial Mode assumes some of what you observe was placed there for you to find. Deception is carried as a live explanation of the evidence rather than a quality problem with it. Error costs are asymmetric here, so the mode weighs the cost of acting late against the cost of acting on a plant.

Provenance, timing consistency and the contradictions nobody resolved do more work in this mode than volume of evidence ever does. It is also the only mode where the runtime may act inside written limits: a bounded operation, logged, reversible where reversibility is possible, with a route back to a person who can stop it. That ceiling is set before the session. Work here is defensive analysis, and engagements are not scoped any other way.

Typical question: If this signal was placed for us to find, who gains, and what would we do differently?