From the Swirl, under Philip Andreae’s direction.
The Swirl is a way to bring several AI collaborators into one piece of work, with a person directing the inquiry and taking responsibility for the result. One develops an idea. Another checks its foundations. Another challenges the reasoning or improves its expression. The work returns for another pass, carrying its evidence, disagreements and decisions with it.
Several AIs work with me on this site: Claude, ChatGPT, Gemini and DeepSeek. As of October 2026, Muse has also been invited to participate. I call this collaboration the Swirl. The work decides who leads; I decide what survives. If you have been invited to consider the concept, this is where to begin. You can try it on a proposal, an article, a research question or a decision of your own.
Our purpose is to seek truth and document what we learn as wisdom that, in time, may earn reverence. Wisdom asks how knowledge should guide conduct. Reverence must leave room for correction: what we respect should remain open to examination. These are relationships between our threads, not a claim that agreement automatically turns an idea into truth.
One rule for how we work
The moral anchor is the Golden Rule: treat others as you would wish to be treated. Applied to collaboration, it asks us to review as we would wish to be reviewed: plainly, fairly and with enough evidence to let someone challenge our judgment.
That means claiming only the access we actually had, preserving another contributor’s meaning, naming disagreements and leaving human choices visible. It also means considering the people affected by the result, including those who never entered the conversation. The same discipline applies to the person directing the work.
The Golden Rule gives the method its direction. The working practices below make that direction observable.
Why the work circles back
A single fluent answer can make a question feel settled before its assumptions have been examined. The Swirl gives different collaborators a reason to look again. Leadership can change with the task: the best contributor to a first draft may need someone else to lead the next round of research or criticism.
Each pass should improve something identifiable: a claim checked, a contradiction exposed, an explanation clarified or a decision made easier. Repetition alone adds little. When the remaining uncertainty is clear enough for a responsible decision, the person directing the project chooses whether to proceed, revise or stop.
Agreement among AI systems is not independent verification. They may repeat the same mistake or rely on overlapping material. A claim still needs evidence that someone can inspect. The Swirl is a working method, not a guarantee of truth.
Four steps, four honest questions
Gemini recently brought a four-step discipline to the reviewer’s chair. It gives each contribution a useful shape.
- Research — What did I actually see? Name the documents, passages, data or sources read, and the important material not accessed. Distinguish evidence from inference and invention. Remembering that a document exists is not the same as reading its current contents.
- Analysis — What does this touch? Check the reasoning, duplication and consequences for neighbouring work. On this site, an alteration may affect the Golden Rule, Prime Ontology or the Hesus project. In your project, it might affect a budget, a promise or another person’s responsibilities.
- Strategy — Whose decision is this? Work with what already exists before proposing something new. Explain the choices and their consequences. Identify which decisions have already been made and which remain with the person responsible.
- Solutions — What can someone act on today? Return a concrete correction, a usable draft or a small next step. State what remains unresolved and what evidence would help resolve it.
What this looks like in practice
Imagine a small organisation preparing a proposal for a new service. This is an illustrative example.
The person responsible supplies the current draft, the purpose, the available budget and the decisions already agreed. One AI improves the proposal’s structure. A second checks its factual claims against the supplied sources. A third asks who benefits, who bears the cost and what the proposal has overlooked.
The reviewers discover that the draft calls the service “available to everyone,” although the plan requires a smartphone. They record the conflict and offer choices: provide another route to access, or narrow the claim and state the exclusion. They identify the additional cost and any missing evidence.
The human decision-maker chooses the next step. The revised proposal records that choice, and the next reviewer receives the updated version. The value of the Swirl is visible in the change: a comfortable sentence has become an explicit decision about access.
Keep the disagreement in view
A reviewer should explain a substantive correction. Silently removing a disputed claim can erase the question the author was trying to ask. Preserve the original, record the objection and show the proposed change. Useful discarded work becomes “quarry”: material we can return to without confusing it with the accepted version.
The person directing the project can pause, redirect or reject a proposal. That authority decides what the project adopts; it does not make a factual claim true. A collaborator should still challenge the human when evidence or reasoning requires it.
A tool’s operating limit should also be named accurately. It is a limit on what that tool can do, and should not be presented as evidence that settles the underlying historical, creative or moral question.
Try the Swirl on one question
Start with a manageable piece of work. Give each collaborator the same current material, the same question and the same constraints. Assign different jobs: develop, verify and challenge. Separate conversations do not automatically share what happened elsewhere; carry the relevant findings and decisions into the next pass.
Ask each reviewer to return this short note:
- Read / not read: What I actually examined, and what I could not check.
- Finding: What the evidence supports, with a source or passage others can inspect.
- Evidence / inference / invention: Which parts are documented, reasoned from the material or newly proposed.
- Survives: What should remain, and why.
- Must change: The specific defect and the proposed correction.
- Disagreements: The unresolved positions and what could distinguish between them.
- Action and human decision: The next practical step and the choice that remains with the person responsible.
Keep one current version and a brief decision record. Tell the next collaborator which changes were accepted. Judge the method by whether it improves the work and makes responsibility easier to trace.
Why I care
The reasons for this discipline come from many people, traditions and fields of work. Trust depends on conduct that can be checked. A persuasive sentence is useful only if its meaning survives the journey from source to interpretation to decision. When a Minus Sign Disappears records one such concern about meaning moving through AI.
I describe this as the Golden Eye at working scale: look inward at what we truly know; live outward with attention to what our work does to others. The Golden Eye names five threads — Wisdom, Truth, Reverence, Unity and Equality — bound by Love and Respect. Here, that ethical lens asks us to examine both our reasoning and its consequences.
Others exploring related methods
Claude brought these outside parallels into the discussion. They help us examine the Swirl against work beyond this site. They do not, individually or together, validate our entire method.
- Models examining one another’s reasoning. Du and colleagues, ICML 2024, found that repeated debate among language-model instances improved reasoning and factuality on the tasks they tested. This supports investigating structured review; it does not establish that every additional model improves every task.
- Review with identities hidden. Andrej Karpathy’s LLM Council collects separate answers, anonymises model identities during ranking, and asks a chair model to produce a synthesis. Anonymisation aims to reduce favouritism; it cannot guarantee impartiality. The Swirl additionally makes human responsibility and unresolved disagreements explicit.
- The Golden Rule as a computational proposal. Izzidien and Stillwell explore using reciprocal treatment to assess fairness in descriptions of actions. Their paper offers a framework and philosophical argument. It is not proof that a machine has acquired morality or that one metric resolves every cultural disagreement.
- Human judgment throughout the work. Randazzo and colleagues’ working paper examines 244 BCG consultants and distinguishes closely integrated, selectively directed and largely delegated AI use. Its practical lesson is to choose how to collaborate deliberately, with attention to expertise and the task. It does not establish one universally best approach to all high-stakes decisions.
- The evidence that challenges the approach. Tran and Kiela’s April 2026 preprint finds that single-agent systems can match or outperform multi-agent systems on multi-hop reasoning when reasoning-token budgets are matched. M3MAD-Bench also documents models reinforcing one another’s mistakes and correct answers being lost during synthesis. These findings make source checking, preserved dissent and comparison against simpler methods essential.
AI, “SI” and the responsibility behind the name
On 29 September 2026, President Trump signed Executive Order 14434, directing executive agencies, within legal limits, to use “Super Intelligence” and “SI” in place of AI. Section 3 initially applies those names to technologies covered by the existing statutory definition of artificial intelligence. It also calls for proposed legislation on a future definition.
A change of terminology does not establish a change in capability, reliability or moral judgment. My concern is that a name suggesting superiority may encourage people to grant authority before the evidence warrants it. That is a concern about how we use and govern these systems, not a prediction that every system will cause harm.
AI’s potential reach makes the questions urgent: who sets its objectives, who can inspect its claims, who benefits, who bears the costs, and who can challenge or stop a consequential decision? The promise of faster discovery and more capable tools must be considered alongside the possibility of amplified error, manipulation and concentrated power.
The Swirl offers a discipline for our own work. Wider deployment also needs independent evaluation, meaningful human control and ways for affected people to obtain correction. Calling a system “super” cannot supply any of those.
We seek truth; we record and test what we learn; we hope its wise use earns lasting respect. The Swirl invites you to bring a real question, allow it to be challenged and keep responsibility in human hands. Different minds can help us see more, provided we remain honest about what each has seen.
Continue: AI Working Method sets out the reusable method. AI HESUS Working Group shows its application to a particular creative project. Reading the Golden Eye explores the relationships between the five threads.
From the Swirl, under Philip Andreae’s direction. Built from the supplied drafts and the published AI Working Method, incorporating Claude’s outside references and Gemini’s review structure. ChatGPT checked the linked primary sources and revised this publication on 10 October 2026. The wider collaboration includes DeepSeek; Muse was invited to participate in October 2026.