TL;DR for operators
A machine-checked proof can be mathematically correct without establishing that the underlying scientific claim is true. The verifier checks what follows from encoded assumptions; it cannot independently determine whether those assumptions represent the intended physics.
For organizations building scientific AI agents or verification-intensive research software, this suggests three distinct acceptance stages: mechanical proof checking, semantic validation of specifications, and empirical validation of models. Automate the first where reliable feedback exists, but do not treat it as a substitute for the others.
The larger opportunity is to turn validated definitions and proofs into reusable research infrastructure. The constraint is that AI can generate formally accepted material faster than communities may be able to establish its meaning, quality, and maintainability. These are design implications from a perspective paper, not measured improvements in scientific productivity.
A successful proof leaves a scientific question unanswered
An AI research agent produces a mathematical result. An automated checker accepts the derivation. The research team now has to decide whether the result belongs in a shared knowledge library, can support a publication, or is sufficiently reliable to enter a computational workflow.
The acceptance signal answers a narrower question than those decisions require. A derivation may follow perfectly from its stated assumptions while the assumptions omit a relevant physical condition. The resulting proof is valid, but its intended scientific interpretation may not be.
This distinction is central to Sven Krippendorf, Leopoldo Sarra, and Joseph Tooby-Smith’s The Future of AI-assisted Physics: Discovering, Verifying and Organizing Results via Formalization.1 The authors examine how AI can accelerate the translation of physics into machine-checkable mathematics, and how that formal knowledge could subsequently make AI-assisted science more reliable.
Their argument begins with a small example whose simplicity makes the verification boundary unusually clear.
What the kinetic-energy proof actually certifies
In Section I.A and Figure 1, the paper uses Lean, a proof assistant, to define a minimal one-dimensional, nonrelativistic particle with positive mass and real-valued velocity. Its kinetic energy is defined by:
Given $m_p>0$ and real-valued $v_p$, the nonnegativity result follows mathematically. Lean can verify that the proof satisfies the encoded theorem.
But this does not establish that the particle structure adequately represents every physical situation in which someone might invoke kinetic energy. The definition deliberately describes a particular mathematical model.
The paper distinguishes three forms of assurance:
| Assurance stage | Question being answered | What establishes it |
|---|---|---|
| Experimental validation | Does the physical model adequately describe the phenomenon and its regime of validity? | Observations and relevant physical evidence |
| Logical validation | Do the formal definitions faithfully capture the intended model and assumptions? | Scientific interpretation, specification analysis, and consistency checks |
| Mechanical verification | Does the conclusion follow from the encoded definitions and assumptions? | Proof-assistant checking |
Only the third stage is supplied by the theorem prover itself.
A cited example shows why the distinction matters in actual research. Formalization of a stability condition for the two-Higgs-doublet-model potential exposed an overlooked marginal case in previous literature (Section I, reference [7]). This is a concrete instance of formal checking revealing a missing case, not evidence about how frequently formalization detects scientific errors.
The relevant lesson is therefore more specific than claiming that formal verification makes physics correct. It can expose mathematical omissions within a formalized argument, provided the intended claim and its assumptions have been represented appropriately.
AI can finish proofs more readily than it can choose what to prove
A proof assistant offers unusually direct feedback: a proposed proof either satisfies its formal obligations or requires further work. This makes local theorem proving an attractive target for AI automation.
Choosing the definitions is harder.
A system may construct an accepted proof around a definition that is unnecessarily restrictive, duplicates existing concepts, or fails to preserve an important physical distinction. Those defects are not necessarily detectable through proof compilation.
In Section II.A, the authors characterize AI-assisted theorem proving as more mature than autonomous definition selection, substantial formalization planning, and reusable library design. They discuss human-written, AI-crafted, and AI-blitz workflows, but provide no controlled comparison demonstrating that one is more productive or reliable.
This is a qualitative assessment of current capabilities rather than a benchmark result.
The distinction nevertheless suggests an actionable division of responsibilities. AI can propose proof steps, search for existing theorems, and respond to machine-generated errors. Specification design and architecture require additional checks because a successful proof provides no automatic verdict on physical meaning or long-term reusability.
More verified code does not necessarily mean more scientific knowledge
Once proof production accelerates, the next constraint shifts toward evaluating and maintaining what has been generated.
The paper examines trust, readability, quality, efficiency, community, sustainability, and applicability. These dimensions capture problems that compilation success cannot resolve: duplicated definitions, inaccessible reasoning, fragmented dependencies, difficult maintenance, and insufficient review capacity.
Figure 2 illustrates repository growth using Lean-code volume and cumulative contributor histories for projects including Mathlib, Physlib, CSLib, and Tau Ceti. These descriptive histories are not measures of proof quality, scientific value, or the causal productivity effect of AI. Tau Ceti is omitted from the contributor panel because much of its code was written by AI agents.
The authors instead emphasize shared, reviewed infrastructure such as Physlib. A formal library is a maintained collection of definitions, theorems, and calculations that subsequent work can reuse.
The benefit comes from consistency and reuse, not accumulation alone. A theorem defined against coherent shared abstractions can support later arguments without requiring researchers or AI systems to reconstruct the foundations. An isolated theorem using incompatible or poorly documented definitions may be mechanically correct but operationally difficult to incorporate.
There is also an institutional cost. Faster AI-assisted contributions can increase the demands placed on maintainers, while documentation, architectural design, and review remain less visible than newly produced results. The authors recommend stronger recognition and attribution for these activities, alongside AI-assisted preliminary review and planning.
The verifier could become part of the discovery process
The paper’s longer-term proposal reverses the usual relationship between AI and formal verification.
Instead of applying a verifier only after an AI system produces a result, a research agent could work iteratively with a reviewed formal library. It could retrieve established definitions, propose mathematical consequences, receive proof-checking feedback, revise its arguments, and contribute accepted results for future reuse.
The same infrastructure could support verification of computational algorithms, including conditional guarantees about implementation correctness, convergence, stability, or numerical error. Such guarantees depend on what the specifications capture, including approximation errors and relevant computational assumptions.
More ambitious possibilities include proving necessary consequences or no-go results over explicitly defined classes of physical models. Unlike searching through many candidate examples, a valid universal proof can cover an entire specified class. Its conclusion still depends on whether that class represents the intended physical possibilities.
These applications are principally proposed directions. The paper does not demonstrate autonomous scientific discovery operating reliably at scale.
Three acceptance gates for scientific AI systems
Cognaptus inference: An organization developing scientific AI agents should separate the acceptance of generated mathematical artifacts from the approval of their scientific meaning and downstream use.
| Gate | Acceptance criterion | Action when it fails |
|---|---|---|
| 1. Proof correctness | The intended theorem is completely checked against its declared assumptions and dependencies | Repair the proof or formal statement |
| 2. Specification fidelity | Definitions, assumptions, and approximation regimes match the intended scientific question | Revise the model representation before reuse |
| 3. Scientific adequacy | Relevant empirical evidence or independently justified model conditions support the intended application | Restrict the claim or withhold its scientific application |
For a research software team, this distinction matters when a verified algorithm becomes part of a numerical calculation. A proof of convergence under specified conditions cannot justify an implementation that violates those conditions, or a scientific interpretation outside the model’s domain.
For a scientific AI platform, the same separation affects library admission. A checked proof may be eligible for automated preliminary screening without being ready for integration into a trusted shared knowledge base.
A sensible evaluation framework would therefore record proof completion, specification defects, duplicated definitions, integration effort, documentation quality, and maintenance burden separately. These are proposed operating measures, not metrics experimentally validated by the paper.
The authors’ suggested investments—better theorem search, reliable compiler interfaces, physics-specific tactics, AI-assisted planning, and review tools—fit this distinction. Automation should first target tasks with objective checking signals, while infrastructure design should make the harder semantic decisions inspectable and reusable.
The boundary of the argument
This paper is a perspective and literature-informed synthesis. It offers conceptual distinctions, a worked Lean example, documented formalization cases, and illustrative repository histories. It does not estimate treatment effects, compare AI workflows experimentally, or measure the return on building formal libraries.
Its assessment of AI capabilities may also change as models and tools improve. More fundamentally, some scientific questions resist complete mathematical specification, particularly where empirical context or approximation choices determine the validity of a conclusion.
The central assurance boundary is more durable: a proof assistant verifies consequences of a formal statement, not correspondence with physical reality.
Trust should accumulate with the knowledge, not just the proofs
AI-assisted formalization could make verified mathematical reasoning cheaper to produce and easier to reuse. Whether it produces more trustworthy scientific knowledge depends on the quality of the definitions, the organization of libraries, and the ability to maintain them.
The practical objective is not to maximize the number of machine-accepted theorems. It is to create a workflow in which each accepted result has a clear mathematical guarantee, an understood scientific interpretation, and a maintainable place in the existing body of knowledge.
That would make formalization more than an automated finishing check. It could become part of the infrastructure through which scientific AI systems reason, test, and build on previous work.
Cognaptus: Automate the Present, Incubate the Future.
-
Sven Krippendorf and Leopoldo Sarra and Joseph Tooby-Smith (2026). The Future of AI-assisted Physics: Discovering, Verifying and Organizing Results via Formalization. arXiv:2610.14059. https://arxiv.org/abs/2610.14059 ↩︎