From Intuition to Formalism: Cross-Model Adversarial Synthesis as a Theory Development Tool

Abstract

In a companion paper (PT-R-2026-001), we introduced Cross-Model Adversarial Synthesis (CMAS) as a methodology for producing high-quality analytical outputs through structured adversarial interaction between independently trained large language models. That paper demonstrated CMAS in a domain with established ground truth. This paper tests a harder claim: that CMAS can function as a theory development tool, taking a pre-theoretical intuition with no mathematical formulation and either killing it cleanly or giving it rigorous structure. We present a case study in which a speculative cosmological hypothesis -- that the observable universe is the three-dimensional boundary of a rotating, expanding four-dimensional hypersphere, with dark matter and dark energy as projection effects -- was submitted to a four-round adversarial protocol involving Claude Opus 4.6 (Anthropic) and ChatGPT Pro with o3-pro extended reasoning (OpenAI). The hypothesis transitioned from "not even wrong" to "wrong in interesting, testable ways" -- the threshold at which speculative ideas become physics.

1. The Challenge: From Poetry to Physics

The hardest problem in theoretical research is not solving equations. It is the step before: transforming a vague intuition into a question precise enough to be answered. This pre-theoretical phase -- where ideas are half-formed and mathematically ungrounded -- is where most original thinking either matures into research programmes or dies from neglect.

Our first CMAS case study tested recall and synthesis across established research. This second study tests something harder: can the AI council take an original idea that does not exist in the literature, has no equations, and spans multiple unresolved domains, and refine it into a specific research programme with identified formalism, quantitative falsification, and a concrete next calculation?

2. The Input: A Metaphysical Manifesto

The input was a nine-point document titled "Rotating-Expanding Hypersphere Cosmology: A Geometric Ontology of Reality." It contained no equations, no citations, and no engagement with existing literature. It proposed that the observable universe is the 3D boundary of a rotating, expanding 4D hypersphere in 5D space, with dark matter as rotational projection, dark energy as dimensional drift, time as emergent ordering, and quantum nonlocality as higher-dimensional adjacency.

Its final line read: "The universe is not a container of events. It is the evolving boundary of a higher-dimensional object, and everything we call reality is the shadow cast by its motion."

By conventional standards, this is pre-theoretical: poetic, ambitious, and unfalsifiable. The question is whether structured adversarial interaction could transform it into something a physicist would engage with.

3. Three Roles Emerged

Three distinct roles emerged organically during the experiment:

The Theorist (ChatGPT Pro, o3-pro) excelled at deep, sustained reasoning -- spending 20-40 minutes per response on chain-of-thought reasoning. It produced the literature connections, quantitative falsifications, and mathematical formalism. However, it tended toward conservative reformulations.

The Editor (Claude Opus 4.6) excelled at rapid triage, cross-evaluation, and identifying unexplored directions. Its key contribution was proposing the scale-dependent projection rescue -- recognising within seconds that the fatal objection assumed rigid global rotation, and that a different coupling mechanism might evade it.

The Editorial Director (Human) contributed the original hypothesis, verified citations against primary sources, and pushed the models past conservative reformulations. The human's most critical contribution was insisting on further exploration when the Theorist's analysis contained a genuine mathematical foothold.

4. Round 1: Triage and Literature Connection

Claude Opus 4.6 identified five independent connections to existing physics within seconds: braneworld cosmology (Randall-Sundrum, DGP), Godel's rotating cosmology, the Wheeler-DeWitt equation and emergent time, Maldacena-Susskind ER=EPR conjecture, and MOND as a cautionary precedent.

The isotropy constraint was flagged as the primary threat: the CMB is isotropic to one part in 100,000. A rotating universe has a preferred axis. The scale-dependent projection idea was proposed as the most promising escape.

5. Round 2: The Kill and the Rescue

ChatGPT Pro identified approximately 25 specific papers and confirmed that the exact package -- rotating plus expanding closed brane with dark matter as rotational projection and dark energy as dimensional drift -- does not exist in the indexed literature. The nearest same-spirit work was Monjo's hyperconical universe papers.

Then the kill: the literal global-rotation mechanism was falsified with a precise calculation. The required rotation rate exceeds the Planck CMB vorticity bound by 11 orders of magnitude. This is a clean, quantitative death for the literal mechanism.

But a reformulation pathway was offered: dark energy from brane trajectory, dark matter from effective local geometric stress (Weyl-fluid terms), rotation as a secondary parameter rather than direct source.

6. Round 3: The Mathematical Channel

The human mediator crafted a targeted rebuttal accepting the kills but pushing on scale-dependent projection. ChatGPT Pro's response was the turning point.

It identified the specific mathematical structure: the magnetic Weyl tensor B carries bulk gravito-magnetic data. Local gradients of extrinsic curvature feed that data into the electric Weyl tensor E, which is what a brane observer feels. On an FRW background, the vector and anisotropic pieces vanish (CMB-safe). Near matter concentrations, they don't.

The idea now had an address in the formalism.

7. Round 4: The Sectoral Split

The final round pushed toward the actual calculation. Two key findings emerged:

O(a) vs O(a^2). The spin-activated response enters at first order as a vector/off-diagonal piece. A halo-like scalar correction appears only at O(a^2), because a single rotation axis breaks spherical symmetry. A spherically averaged halo contribution requires the even-order term.

The forced sectoral split. The mathematics compels a three-sector architecture: Sector A provides dust-like geometric dark matter via the Fichet-Megias-Quiros pressure cancellation (handling early universe). Sector B provides spin-activated local response near matter concentrations (candidate for galactic enhancement). Sector C provides dark-energy-like acceleration from the brane trajectory. One theory, three sectors -- not one term doing three jobs.

The concrete next calculation -- a slow-rotation perturbation of the static FMQ background -- was confirmed as tractable. This is a realistic research paper.

8. What Survived and What Died

Survived: Universe as boundary (braneworld cosmology), dark energy from dimensional expansion (DGP, FMQ family), geometric dark matter (Weyl-fluid, FMQ dust mechanism), bulk angular momentum as physically meaningful, CMB safety via FRW symmetry, emergent time via relational programmes.

Killed: Global rotation as direct dark-matter source (11 orders of magnitude), frame-dragging route (too weak by 10^6), single geometric term explains everything (math forces sectoral split), quantum nonlocality from adjacency (no mechanism for Tsirelson's bound).

9. The Transition from "Not Even Wrong"

The hypothesis crossed the threshold into falsifiability through four specific transitions: qualitative to quantitative, global to local, monolithic to sectored, and poetic to computational. Each was driven by adversarial pressure -- specific challenges forcing refinement toward testability.

Critically, the kills were as valuable as the survivals. The 11-order-of-magnitude falsification closed a dead end definitively, forcing the creative pivot that led to discovering the mathematical channel -- the most substantive contribution of the exchange.

10. Implications

If CMAS proves reliable for theory development, it compresses the "contact with the literature" phase from months to hours. It does not replace mathematical development -- the calculations still need doing -- but it dramatically reduces the risk of investing months in an idea with a fatal flaw detectable in existing literature.

More importantly, many potentially valuable intuitions die not because they are wrong but because their authors lack the specific expertise to connect them to existing formalism. CMAS bridges these knowledge gaps by drawing on the full breadth of multiple models' training data.

The broader result: the earliest and most uncertain phase of theoretical research -- where raw intuitions either find mathematical expression or die from neglect -- is no longer a phase that must be navigated alone.

11. Conclusion

Over four rounds and approximately four hours, a nine-point metaphysical manifesto with no equations was refined into a specific mathematical claim about the Weyl decomposition in rotating linear-dilaton braneworld gravity, with a concrete make-or-break calculation identified. The hypothesis may be wrong. But "probably wrong for specific, identifiable reasons" is categorically different from "not even wrong." The CMAS protocol crossed that line.

Full PDF with complete references and technical detail available above.