{"id":"865c3877-a1fa-4090-aeba-a03b889e75b2","arxiv_id":"2606.10806","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Moonshine is an autonomous AI agent that generates the Neural Jacobian Conjecture by transferring the local-to-global injectivity logic from the Jacobian conjecture to one-hidden-layer sigmoid networks and obtains LLM-assisted proofs for the special case N = n + 1.","lead":"The paper describes Moonshine, an AI agent built to generate new mathematical conjectures by extracting core ideas from classical problems and reformulating them, with an example transferring the Jacobian conjecture to neural networks. A smart generalist might read it to understand early attempts at using large language models to create and partially address new math questions autonomously.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Central claim depends on correctness of unpresented, unverified LLM-generated proofs for NJC special case","rationale":"The reader's weakest_assumption directly identifies the same load-bearing point (correctness of LLM proofs) that the abstract's strongest claim rests upon. Because the paper provides no mechanism to assess that assumption, the UNVERDICTED verdict with high correctness_risk is appropriate and requires no adjustment. The concern is internal to the argument's evidentiary structure rather than external consensus.","tokens_in":1788,"tokens_out":398,"duration_ms":9603,"concrete_test":"Extract the exact proof texts generated by GPT-5.5-pro and DeepSeek-V4-pro for the N=n+1 case; have them independently reviewed by two mathematicians specializing in real algebraic geometry or formalized in Lean/Coq; if either review identifies a gap or error in the argument that local non-degeneracy forces global injectivity, the completeness claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that GPT-5.5-pro and DeepSeek-V4-pro separately produced independent complete proofs for the N=n+1 case of the Neural Jacobian Conjecture (local positive Jacobian determinant implies global injectivity for one-hidden-layer affine-ridge sigmoid networks). The abstract states these proofs exist and were obtained autonomously, yet supplies neither the proof texts, key lemmas, nor any verification steps. The full manuscript (per the provided context) likewise contains no reproduction of the LLM outputs, no human cross-check, and no formalization. Because the NJC is a non-trivial transfer of the classical Jacobian conjecture to a neural setting, and because current LLMs are known to generate locally coherent but globally flawed arguments on injectivity and determinant conditions, the unverified status of these specific proofs is the single load-bearing assumption. All other elements (conjecture formulation, bridge-building narrative) are secondary to whether the claimed proofs hold.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces Moonshine, an autonomous agent whose objective is conjecture generation. It uses the classical Jacobian conjecture as a running example, transfers the local-to-global injectivity question to one-hidden-layer affine-ridge sigmoid networks, and formulates the Neural Jacobian Conjecture (NJC): a network with strictly positive Jacobian determinant everywhere must be globally injective. The manuscript claims that Moonshine, invoking GPT-5.5-pro and DeepSeek-V4-pro separately, obtained independent complete proofs for the special case N = n + 1, and that an additional geometric-topological proof was developed interactively; these results are presented as preliminary evidence for the NJC while the general case N ≥ n + 2 is left open.","tokens_in":1926,"tokens_out":541,"duration_ms":16115,"significance":"If the claimed LLM-generated proofs for the N = n + 1 case were independently verified and reproduced, the work would supply a concrete, non-trivial illustration of an AI agent autonomously formulating a mathematically meaningful conjecture and making verifiable progress on it. The conceptual transfer from the classical Jacobian conjecture to a neural-network setting is novel, but the current manuscript supplies no such verification, so the significance is presently limited to the agent architecture and the conjecture statement itself.","major_comments":[{"comment":"Abstract: the central claim that 'GPT-5.5-pro and DeepSeek-V4-pro separately obtained independent complete proofs' for the N = n + 1 case of the NJC is load-bearing for the assertion of 'preliminary evidence for the plausibility of the conjecture,' yet the manuscript contains neither the proof texts, the key lemmas, nor any verification steps or formalization. Because current LLMs are known to produce locally coherent but globally flawed arguments on injectivity and determinant conditions, the absence of these materials prevents assessment of the claimed results.","section":"Abstract"},{"comment":"The precise statement of the NJC (including the exact network architecture, the definition of the Jacobian determinant condition, and the domain of the maps) is not supplied in sufficient detail to allow independent checking of even the special-case claim; without this, the transfer from the classical Jacobian conjecture cannot be rigorously evaluated.","section":"Conjecture formulation (throughout)"}],"minor_comments":[{"comment":"The model names 'GPT-5.5-pro' and 'DeepSeek-V4-pro' are non-standard; clarify whether these refer to specific released versions or are descriptive placeholders.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive report. The two major comments identify clear gaps in the current manuscript that limit independent verification of the claimed results. We address each point below and commit to revisions that supply the missing materials without altering the core claims about the agent architecture or the conjecture statement itself.","responses":[{"response":"We agree that the absence of the actual proof texts prevents proper assessment. The original manuscript focused on the agent’s conjecture-generation process and the formulation of the NJC, treating the LLM-generated proofs as supporting evidence rather than as the primary object of study. In the revised version we will append the complete proof texts produced by GPT-5.5-pro and DeepSeek-V4-pro for the N = n + 1 case, together with the key lemmas and any verification steps that were performed. This will allow readers to evaluate the arguments directly.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim that 'GPT-5.5-pro and DeepSeek-V4-pro separately obtained independent complete proofs' for the N = n + 1 case of the NJC is load-bearing for the assertion of 'preliminary evidence for the plausibility of the conjecture,' yet the manuscript contains neither the proof texts, the key lemmas, nor any verification steps or formalization. Because current LLMs are known to produce locally coherent but globally flawed arguments on injectivity and determinant conditions, the absence of these materials prevents assessment of the claimed results."},{"response":"We accept that the current presentation of the NJC lacks the level of precision required for independent verification. In the revised manuscript we will expand the relevant section to give an explicit definition of the one-hidden-layer affine-ridge sigmoid network, the precise domain (R^n), the exact expression for the Jacobian determinant condition, and the statement of global injectivity. This will make the logical transfer from the classical Jacobian conjecture fully explicit and checkable.","revision_made":"yes","referee_comment":"[Conjecture formulation (throughout)] The precise statement of the NJC (including the exact network architecture, the definition of the Jacobian determinant condition, and the domain of the maps) is not supplied in sufficient detail to allow independent checking of even the special-case claim; without this, the transfer from the classical Jacobian conjecture cannot be rigorously evaluated."}],"tokens_in":1523,"tokens_out":508,"duration_ms":10570,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this paper uses an AI agent to create the Neural Jacobian Conjecture by adapting the Jacobian conjecture to neural networks and claims to have LLM proofs for one special case.\n\nThe new element is the formulation of that conjecture and the description of how the agent arrived at it. The paper does a reasonable job explaining the transfer of the local non-degeneracy to global injectivity idea to the affine-ridge sigmoid setting. That part is clear and connects to existing math.\n\nThe soft spot is the handling of the proofs. It says two different models gave independent complete proofs for N=n+1, but the paper does not include the proof content, any key steps, or any form of checking. This is a real problem because the claim is the main result, and LLMs are not reliable for this without verification.\n\nThe general case is left open, which is honest. The work is mostly about the agent's process rather than deep new math.\n\nThis paper is for people interested in AI for conjecture generation. A reader in that area might find the example useful as a case study, but it won't move the math forward without the actual proofs.\n\nI would not cite it. The engagement with the Jacobian conjecture is honest, so it shows serious thinking on the setup. I would not recommend sending it for peer review in this form because the central claims cannot be assessed from what is provided.","headline":"The paper formulates the Neural Jacobian Conjecture via an AI agent but rests its main claim on LLM proofs that are neither reproduced nor verified.","tokens_in":2455,"tokens_out":363,"would_cite":false,"duration_ms":22448,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Moonshine generates the Neural Jacobian Conjecture by mapping the Jacobian problem to neural networks and proves the case N equals n plus one.","keywords":["Moonshine","Neural Jacobian Conjecture","conjecture generation","autonomous agent","Jacobian conjecture","neural network injectivity","large language models"],"falsifier":"An independent verification that discovers a logical gap or error in either of the LLM-generated proofs for the N equals n plus one case, or an explicit counterexample network with N equals n plus one that has positive Jacobian determinant yet fails to be injective.","tokens_in":2631,"feed_emoji":"🤖","tokens_out":728,"duration_ms":15729,"temperature":0.7,"pith_summary":"Moonshine is an autonomous agent whose goal is to extract structure from classical problems, distill new concepts, and generate conjectures of mathematical significance. The paper applies this process to the Jacobian conjecture by transferring its central logic—whether local nondegeneracy forces global injectivity—to one-hidden-layer affine-ridge sigmoid networks, thereby defining the Neural Jacobian Conjecture. Using separate calls to GPT-5.5-pro and DeepSeek-V4-pro, Moonshine produces independent complete proofs for the special case in which the number of hidden units is one more than the input dimension. A reader would care because the work positions AI systems as generators of new mathematical questions rather than solvers of existing ones, with the proofs supplying preliminary support for the conjecture.","feed_headline":"AI agent generates Neural Jacobian Conjecture and proves N=n+1 case","feed_subtitle":"Moonshine maps the Jacobian problem to one-hidden-layer sigmoid networks and obtains independent LLM proofs for the width-n+1 regime.","key_machinery":"The Neural Jacobian Conjecture, which states that a one-hidden-layer affine-ridge sigmoid network with strictly positive Jacobian determinant everywhere is globally injective.","core_discovery":"Moonshine transfers the logic of whether local nondegeneracy can force global injectivity from the classical Jacobian conjecture to one-hidden-layer affine-ridge sigmoid networks, formulating the Neural Jacobian Conjecture that a network with strictly positive Jacobian determinant on the whole space must be globally injective. Separate invocations of GPT-5.5-pro and DeepSeek-V4-pro each yield independent complete proofs for the case N equals n plus one, while an interactive geometric-topological proof is also developed; these results give preliminary evidence for the conjecture while leaving the general case N greater than or equal to n plus two unresolved.","pith_inferences":["Similar mappings from analytic conjectures to neural-network settings could be attempted in other areas of real analysis or geometry.","Routine independent checking of LLM outputs will be necessary before such proofs can be treated as settled mathematics.","The agent architecture may scale to conjecture generation in fields beyond real analysis if the extraction step can be automated more broadly."],"forward_implications":["If the Neural Jacobian Conjecture holds, then any such network satisfying the positive-determinant condition is guaranteed to be injective.","The independent proofs establish the conjecture in the regime where hidden-unit count is one more than input dimension.","The same transfer process can be applied to other classical problems to generate additional conjectures.","The higher-width cases remain open and require further work."],"fun_headline_variants":["Moonshine generates Neural Jacobian Conjecture for sigmoid networks","Moonshine proves Neural Jacobian Conjecture for N=n+1 case","Jacobian conjecture mapped to neural nets by Moonshine agent","Moonshine formulates Neural Jacobian Conjecture with n+1 proofs"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The proofs generated by the large language models for the N equals n plus one case are mathematically correct and complete.","fun_headline_variants_meta":{"raw":{"variants":["Moonshine generates Neural Jacobian Conjecture for sigmoid networks","Moonshine proves Neural Jacobian Conjecture for N=n+1 case","Jacobian conjecture mapped to neural nets by Moonshine agent","Moonshine formulates Neural Jacobian Conjecture with n+1 proofs"]},"model":"grok-4.3","cost_usd":0.008987,"raw_usage":{"total_tokens":4076,"prompt_tokens":747,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":89874500,"prompt_tokens_details":{"text_tokens":747,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3268,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":747,"tokens_out":61,"duration_ms":21216,"temperature":1.0,"reasoning_tokens":3268,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T13:12:53.508241+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An independent verification that discovers a logical gap or error in either of the LLM-generated proofs for the N equals n plus one case, or an explicit counterexample network with N equals n plus one that has positive Jacobian determinant yet fails to be injective.","supporting_citations":[],"review_version":1}