Dual-Track: A Scalable Governance Architecture for Auto Research

The problem is not only how to generate research

Once AI can retrieve sources, propose hypotheses, write proofs, run experiments, and complete papers, the most obvious bottleneck in research seems to be: how do we get it to pose more and better questions?

I ran into a different problem.

AI can easily write an unfinished chain of research as if it were complete. The mathematical language may be polished, the program may run successfully, and the explanations may support one another. But fluency is not proof, running something once is not reproducibility, and an indeterminate intuition does not become knowledge merely because it has been written up as a paper.

So Auto Research needs not only mechanisms for generation, but also mechanisms for governance: on what grounds may a system declare a problem solved?

Dual-track is the basic unit I found for this problem.

Three kinds of problems, two terminal criteria

For practical purposes, I abstract problems in the world into three kinds:

Logical problems                    → formalization
Engineering problems                → reproducibility
Currently indeterminate / metaphysical problems → excluded from consideration

Logical problems ask whether a conclusion follows from stated definitions and assumptions. Their final judgment should be left to a formal system. Informal derivations can aid understanding and serve as proof sketches, but they are not the end state for this kind of problem.

Formalization also does not guarantee that every logical problem receives an immediate answer. A proposition for which neither proof nor disproof has been completed should remain unresolved; “not yet proved” cannot be silently turned into “proved false.”

Engineering problems ask whether a construction can be implemented under stated conditions and whether it produces the expected result when run. Their final judgment should be left to a reproducible process. Theoretical feasibility or code that “looks correct” cannot substitute for actually building and running it.

Currently indeterminate or metaphysical problems are not adjudicated by dual-track. “Metaphysical” here is neither pejorative nor a claim that such questions have no value forever. It means only that, at present, a problem can be stated neither as a formally decidable logical proposition nor as a reproducible engineering proposition; this method therefore cannot declare it solved.

If it later gains clear definitions, observational conditions, or an implementation path, it can be reclassified. Until that transformation is complete, it may inspire research but cannot count as an established result.

The first two kinds constitute the dual-track framework. The third is a boundary, not a third track.

Dual-track is not two validations of the same statement

AI has previously explained my idea as follows: a research conclusion needs two final anchors, one proven by a formal system and the other validated against external reality.

That is not what I mean.

A purely logical problem can end with formalization; it does not need a numerical program to stamp the theorem a second time. A purely engineering problem can end with reproducibility; it does not need the entire system first to become a Lean theorem.

A research project may contain both logical and engineering problems, but that means its propositions must be decomposed, not that its evidence should be mixed.

For example:

• “A family of phase couplings is unitarily equivalent under given assumptions” is a logical problem.
• “After implementing that coupling, does the simulator produce the expected numerical behavior?” is an engineering problem.
• “Can this result improve scans on a real machine?” is an engineering problem at a higher level.
• “Is it worth making this a research direction?” is a judgment about value and resources, not a criterion of truth within dual-track.

These questions are related, but none can impersonate another’s answer.

Why it is a basic governance unit

If dual-track applied only to an isolated proposition, it would be merely a validation principle. The design claim that makes it a scalable governance architecture is this: complex problems should continue to be decomposed, and the same rules can be reused at every level. This is an operational requirement of the architecture; whether it can be executed fully and durably across domains remains to be tested in practice.

Complex problem
→ decompose it into separately decidable subproblems
→ determine the type of each subproblem
→ logical subproblems enter formalization; engineering subproblems enter reproducibility
→ treat the composition relation among subproblems as a new problem
→ classify, decompose, and adjudicate again
→ only after the higher-level proposition itself, its necessary subproblems, and their composition relations have each reached their proper end states may a higher-level conclusion be made within its stated boundary

The “fractal” here is structural self-similarity, not a strict geometric fractal. Whether the object is a minimal proposition, a module, a system, or a complete research project, the governance structure stays the same: classification, adjudication, composition, and then renewed classification and adjudication of the composition.

Local validity does not automatically entail global validity.

Ten components passing their tests separately does not mean the full system already meets its latency target; several mathematical lemmas holding separately does not mean they correctly describe MRI; a simulator reproducing results reliably does not mean those results have transferred to a real machine or to clinical practice. Every transition from a part to a whole creates a new composition problem. That problem must re-enter the dual-track framework.

Therefore, a large problem is not “validated” all at once. It is more like a tree of problems, or a problem graph with shared dependencies: the leaves must stand, the relations that connect them must stand, and intermediate nodes and the root must each stand as well. As long as a necessary branch remains unresolved, a higher-level conclusion cannot bypass it and declare itself complete.

Why it suits Auto Research

People can review a limited number of results; they cannot easily review, one by one, the large volume of results continuously produced by a high-speed automated research system. The more capable the models become, the less scalable it is to rely only on “having an expert check it at the end.”

Dual-track turns the acceptance conditions of research into a protocol that can be executed recursively.

AI can investigate different subproblems in parallel; formal checkers can adjudicate encoded logical propositions; engineering environments can build and reproduce implementation results. When one branch fails, the system can stop that branch, narrow its higher-level conclusion, and preserve where the failure occurred, without having to retell the entire project.

More importantly, a generator cannot change an end state merely through a more fluent explanation. A logical problem that has not reached a formally checked proof or refutation cannot be talked into being true or false; an engineering problem without a reproducible result cannot be rescued by theoretical possibility.

This does not mean people disappear from research. People still choose directions, define risk boundaries, handle anomalies, and decide on resources and value. But credibility no longer requires any one person to understand and approve every intermediate step produced by the system.

In this sense, dual-track is not an ordinary step in an Auto Research workflow. It is the smallest governance structure that the entire workflow can call repeatedly.

Beyond the boundary, other methods are needed

Dual-track has clear boundaries.

A formal system can prove only the propositions actually written into it. Definitions and assumptions come from the current knowledge base; they may be incomplete and may be revised in the future. If a domain problem is mistranslated before entering the formal system, the proof may still be completely correct.

Engineering reproducibility is likewise valid only for the stated model, data, device, and conditions. Reproducibility at the simulation layer cannot replace experiments on a real machine, and results on a real machine cannot automatically replace clinical results. Expanding a claim outward by one layer creates a new engineering problem and requires reproducibility conditions to be specified again.

Dual-track also does not decide whether a result is novel, important, or worth consuming resources. Problem discovery, semantic modeling, novelty assessment against the literature, value choices, and risk governance all require other methods.

These are not missing third, fourth, or fifth tracks of dual-track. They are other governance structures that work around the basic unit.

Codex’s earlier misunderstanding of my idea arose precisely because it saw these boundaries and concluded that formalization and engineering reproducibility were merely two incomplete anchors, requiring ever more validation layers to be added to dual-track. A more accurate approach is to preserve dual-track’s minimal responsibility, then establish adjacent mechanisms for problems beyond its boundary.

A basic unit can be called at scale only if it remains simple.

How it emerged

I did not design dual-track after reading a theory of Auto Research.

In the earlier Minimax project, I repeatedly tightened an optimistic result using cross-validation, random baselines, and minimum description length. In retrospect, that can be understood as an experience on the engineering track. But I did not have the name “dual-track” at the time, nor can I prove that it directly led to the later method.

After entering the entirely unfamiliar field of MRI in May 2026, logical propositions and engineering problems appeared in large numbers at the same time. The project record wrote down what was then its most central rule on May 8:

Logical problems are proved through formalization; engineering problems are proved through reality: build it, and obtain the result we expect.

The complete formulation of the three kinds of problems, along with the recognition of the fractal structure, emerged gradually through later, repeated clarification of this method. It may have been influenced by many prior practices, but it was not a framework I built by following any particular researcher or paper.

Independently arriving at an idea cannot prove that no one else has proposed something similar. Formal verification, engineering reproducibility, recursive decomposition, and composition are not new components.

What I believe remains worth testing is their structure: first characterize the problem; assign distinct end states to the two decidable kinds; leave indeterminate content outside the boundary; then let the same governance unit act recursively on subproblems, composition relations, and the complete research effort.

That is what turns dual-track from a validation principle into a scalable governance architecture for Auto Research.

What can be claimed now

MRI is the first testbed where dual-track was explicitly proposed and put into practice, but it cannot yet establish that this architecture is universally effective.

What can currently be confirmed is that the project has left behind logical results that passed formal checks, engineering results that can be run and reproduced, and paths that were overturned, narrowed, or judged trivial. At minimum, the process is capable of preserving failures rather than only success stories.

What has not yet been confirmed is whether complex research can be stably decomposed into complete problem trees over time; whether composition relations will become a new bottleneck; whether different agents will repeat the same errors during classification and modeling; or whether this governance approach truly yields a higher proportion of valid research than other workflows.

These questions do not resolve themselves simply because the name “dual-track” sounds complete.

They should continue to be decomposed and sent, separately, to the places capable of adjudicating them.