Composite I

Prepared as the companion to a talk for the Mind in the Age of AI pre-workshop, University of Turku, 19 August 2026 (programme), under the title Human-AI Cognitive Partnership as Public Interface: Composite I, Dialectical Tree Search, and Premise Laundering.

A loop can produce fluent work, and even a usable handle, without ever putting its load-bearing premises at risk. The four jobs below keep search, constraint, adjudication, and transmission separate. Premise laundering is the shortcut: closing on the strength of search alone.

I. One principal, many components

A loaded question may be answered inside its own frame by a colleague or by a model, and a good answer, by itself, does not tell you whether the assumption in the question was tested.

What matters is who owns the grounds, who has to answer, and what a no is allowed to stop.

An I, here, is one principal: the person who decides what gets asserted and who has to answer for it. That is a claim about control, not a verdict on whether a model is a self or part of the mind. The principal can catch a bad premise and often doesn't: a mind is already a committee — drives, habits, inherited convictions. Losing track of which process produced a conclusion is an old failure, older than any model.

An Us contains more than one center of judgment: more than one person who can stand behind their own reasons and withhold endorsement. Final publication may still belong to one person, and separate grounds are then available, though hierarchy, deference, shared incentives, and groupthink often suppress them. Not every human team is an Us: a committee that cannot challenge the chair is many people and one principal. When they disagree, that disagreement need not come from the person whose premise they are challenging.

A Composite I is one principal working through models, tools, prior work, and a protocol. Production is many-handed, but what gets asserted, and who answers for it, stay with the principal — the same person the interface calls a user. The model does real work, and from the outside it looks like a team: turns, roles, surviving criticism, a record of having been attacked.

A Composite I can sit inside an Us: several people, each with their own model-assisted loop. The plurality then comes from the separately answerable people, not from the number of generated voices.

Who authorizes the question, the evidence, and the standard of success, who decides which objections land, and who stands behind publication? If that authority collapses to one person, the arrangement is still a Composite I even when a model proposed the question or drafted the standard.

Composite I names the recurrently coupled arrangement: what comes out of one step can change the next. A one-shot use — a single query, a paper read once — falls outside the name.

II. The search that looks like a team

The signals people use to tell that thinking went well are team-signals. Someone pushed back. The account survived. It clicked into a simpler shape. In an Us those signals can mean another center of judgment did some work. A Composite I can produce the same signals by a search it runs: the principal owns the prompts and roles, even when a model drafted them.

Dialectical Tree Search is that search. Dialectical because a candidate is attacked, not just extended. At one question it looks like a tree: stand at a candidate, expand it, attack it, keep both readings when you cannot yet decide. The walk is made up as you go, not laid out in advance. Local defects get patched; if the inherited premise or the mechanism changes, start over. A helpful prompt tends to draw out elaboration; a hostile prompt can draw out objections; the principal still decides whether the objection lands. From the outside that looks like a team.

Dialectical Tree Search is useful when the question does not arrive with a fixed hypothesis space or an answer key. Its immediate product is not a verdict but a candidate map: explicit claims, unresolved alternatives, and consequences that might distinguish them. A structured candidate can be wrong and still be useful because its joints are visible. A vague answer gives the audit nothing determinate to attack. The search does not turn the attack into independent evidence.

Here the loop means the recurrent human–model search. Sources, compiler executions, and later outcomes enter as evidence or checks; they are not dialogue partners. The same loop has been used where the required output had to survive outside the conversation: a legal-state representation checked against independently specified states and source invariants, mechanism analyses tested against sources and outcomes, and distinctions intended for use by readers who did not share the derivation.

Different prompts give different searches over models that often err in the same places. They can find defects one pass misses, but agreement among those searches is not independent confirmation: it may be shared training, shared evidence, a shared premise, or a shared task. Different evidence, a flipped premise, or a source the first pass did not have can change which claim is live. Generated alternatives still need a decision.

The search can turn a hidden assumption into an explicit alternative — a branch — but further elaboration alone does not independently decide that branch. If every alternative comes from the same evidence and the same principal picks the winner, you have more arguments, not more independent evidence.

When a premise goes in one end and comes back as confident prose, and the principal treats the quality of that prose as support, the failure is premise laundering. The wash is the point: the “suppose” comes off. Suppose a hypothesis H becomes therefore H. You have to assume that hypothesis to see what it can bear — that is still engagement, and most thinking is that. Laundering is closing H on the strength of work that never tested it.

Take the loaded question: why does consultation improve a policy? A model can produce a polished account of feedback, error correction, and legitimacy. The search first turns the hidden premise into alternatives: consultation may alter the proposal, merely record engagement, or lend legitimacy without correction. It can sharpen what each branch predicts and discard incoherent versions. It cannot establish which branch occurred merely by continuing the discussion. Decision records and later outcomes can discriminate which branch occurred. Participants with independently owned reasons can keep alternatives open, introduce evidence the prompt did not select, and identify what still has to be checked. Only after that can the surviving distinction — engagement, adjudication, or laundering — be packaged for use on another consultation.

III. Two kinds of no

A colleague writes back: I don't accept the question. A calculation comes out the other way. A model, asked to attack, produces three pages of objections. All three can be called a no. They are not the same job: the two kinds are a no from another person, and a no from the world; the model's objections are the lookalike.

The first kind: the colleague can reject the frame, bring reasons the prompt did not pick, and refuse to sign off. That is a counterparty — another person who can be held to their no: asked why they refused, they have to stand behind that refusal as theirs. Someone you asked to review still counts if they own their grounds. Someone you only told to play opponent is not the same as someone who actually disagrees. They can keep the question open, and they can still be wrong: disagreement is independent dissent, and it does not, by itself, test whether the premise is true.

The second kind: the calculation is a discriminating check when its expected outcome depends on whether the premise is true. So is a source, a test you can run, or a later result. That is constraint from the world. A second person is not required for it, and adding one does not replace it. The arrangement is still a Composite I.

The lookalike: the model's pages of objections can surface a defect one pass missed. That is a candidate objection. If the attack happens to contain a valid derivation, or a result that would have come out otherwise if the premise were false, that result is the check. Generating the objection did not settle the premise, and instructing the model to attack did not turn it into a counterparty.

In the arrangement described here, the colleague can occupy an answerable counterparty seat. A check is not a seat, and model-generated opposition does not create one merely by sounding like refusal.

A no with no relevant grounds is noise. Relevant grounds with no route to affect the conclusion are procedurally inert; when their presence is displayed as evidence of rigor, they become ceremony. No here means a result offered as a stop: a coin flip, a safety filter, or a prompted critic can all produce one, and none of that, by itself, tests the premise or the question.

A person can dissent badly and a test can be badly specified. Being able to say no is an entry requirement, not a proof. A colleague who shares the question may add no independent check; a test the same principal runs can.

IV. A grounded no

A grounded no is a result that bears on the claim and is allowed to change what gets asserted. What works is not that such a no occurred. Rejection need not have occurred. The claim must still have faced a person or a test that could have stopped it, under a rule that would have let that result land. A no that bears on the claim but has no route to alter it is not a corrective channel.

Owning whether a no can land is not performing every step around it. The loop can draft the prompts, the labels, even a rule for what a negative result should do. A proposed “independently checked” is still generated. The transcript records who said what, not which claims were established. The principal need not have typed the rule: they answer for whether the no was allowed to change the claim, and they stand behind what remains.

Search finds candidate defects. Checking decides them. When it clicks, stop generating and start checking.

One working architecture is several people, each operating a Composite I, connected to checks none of them controls. Earlier work can bring in a premise this exchange did not inherit. Past-you, under different pressure, may reopen the question — another live alternative, whose facts still need checking.

V. After it could have failed, a handle

A claim can travel without ever having faced something that could have stopped it. A handle is something someone else can pick up and use on a new case without the author. That is packaging, as How I Work and Engineering for the Noosphere specify, not a test of whether the claim could have failed.

Astrology travels: natal-chart procedures are learnable and repeatable, which does not make them true. An instrument that tries to measure both measures neither carefully.

Ask first what could have made the claim fail, then whether that claim has a handle. The order is not interchangeable: a well-packaged conclusion that never faced a stop it would have had to respect is a process failure that spreads easily, including when the conclusion happens to be true.

When a conversation gets more fluent, it can feel as if the starting point has been secured. The Compression Paradox names that inversion for formalization: the more elaborate the structure built on top, the more solid the foundation feels. A Composite I running Dialectical Tree Search is one setting in which the same inversion can occur when the structure is a conversation. Premise laundering is treating that feeling as a test of the premise.

Related:

• The Compression Paradox — why a more sophisticated superstructure creates confidence in the foundation
• How I Work — whether the thing can be picked up and used, asked after it could have failed
• Engineering for the Noosphere — invocability
• Holistic System Rotation — how a person builds a denser internal model by staying with a system
• Cargo-Cult Epistemology — the form of rigor without its function