Papex

Explore

Who checks AI-made proofs? A count of 151 proof claims on Erdős problems (Papex Report No. 1)

Between 14 July and 21 August 2026, 151 proof claims that name an AI system were submitted to erdosproblems.com and are still listed there. We read the public pages of each one in a fixed set of places and asked a single question: within 45 days, did anyone other than the claimant publicly check it? For 39 of the 151 (25.8%; 95% interval 19.5% to 33.3%) we found at least one such check. For 26 (17.2%) at least one check was of the proof itself: someone read it and said what they had checked, formalised the result independently, or wrote the argument out step by step. For the other 13 the formal statement was reviewed or the files were run again. The number depends more on what is allowed to count as a check than on statistical uncertainty. Taking together every tested reading of our rule that lowers it gives 32; every one that raises it, 66. This is a count on one site, over one summer, in one field. "We found no check" does not mean "nobody checked". No person coded a case: the cases were coded by AI sessions under a protocol written before the count.

A search for the IRAS-AKARI Planet Nine candidate in the NOIRLab Source Catalog DR2

Phan et al. (2025) paired a 1983 IRAS source with a 2006 AKARI source and proposed a planet at 500-700 au. The 2013-2018 track lies inside the Dark Energy Survey footprint. I searched the single-epoch measurements of the NOIRLab Source Catalog DR2 (245 DECam exposures on 92 nights) for any body on an orbit through both positions at 300-900 au, bound or unbound. No linkage of five or more nights exists. Injected bodies are recovered with 95 % efficiency to r = 23.04 and 50 % to r = 23.34. If the object is as large as its far-infrared fluxes imply and has an albedo of at least 0.10, the search would have found it with probability >= 0.87 out to 700 au; beyond that the answer depends on how the fluxes are fitted. Darker or strongly methane-absorbed bodies beyond about 500 au are not excluded.

Evidence grades for protein database entries longer than titin

Protein databases hold hundreds of sequences longer than titin, the muscle protein often described as the largest protein. Most are gene predictions whose records do not state their evidence. We collected every sequence longer than a 35,991-aa human titin isoform from UniProtKB, UniParc and NCBI Protein and kept 510; 57 are longer than PKZILLA-1 (45,212 aa). Each was graded A–D by explicit checks, then tested with public RNA-seq (261 runs) or, for three bacterial models, with their source reads. Of the 57, 17 are contradicted (grade D) and 34 have no line of support (C). The most common pattern, a translated tandem DNA repeat, also marks real silk proteins and adhesins in our positive controls, so it counts against a model only with DNA-level evidence; RNA-seq can rarely settle these models, because almost none of their splice junctions is unique. Two fish models are consistent with merged tandem titin genes. Six sequences, at four loci, reach grade A or B, all on gene-model evidence and stable in 31 of 33 threshold settings; the longest, an 85,803-aa Candidatus Omnitrophota open reading frame, is supported by its source reads except across a 1.2-kb repeat. For a 39,686-aa Candidatus Velamenicoccus protein whose detection its authors had reported, all 60 raw proteomics files gave 182 unique peptides, none beyond its first 40 %: the N-terminal part is made, a full-length chain is not shown. Controls exposed weaknesses in three of our rules; we report every revision and release the table, rules and code.

Synthetic Microbial Therapeutics for the Targeted Elimination of Antibiotic-Resistant Bacteria

Antimicrobial resistance (AMR) is a growing worldwide concern because it threatens global health. The increase in antimicrobial resistance is attributed to the rapid evolution of bacteria, the ability of bacteria to transfer genes among themselves, and the ongoing selective pressure from the continued use of antimicrobials. New antibiotic development and/or combination therapy approaches are no longer adequate due to the increasing prevalence of multiple drug-resistant organisms. I propose that there is a way to produce engineered microbial systems that can serve as programmable therapeutic agents and can detect AMR bacteria and kill them without damaging healthy commensal microbes in our body. The engineered microbes could have pathogen detection capabilities through pathogen-sensing modules that respond to quorum-sensing signals and genetic markers indicating AMR and have the ability to produce bacteriocins, disrupt quorum sensing, and use CRISPR for the degradation of resistance genes through effector modules. These engineered microbes would act as autonomous and self-regulating therapeutic agents, enabling the targeted inhibition of AMR pathogens while avoiding any significant disruption of the normal microbiome. Thus, these systems could represent a shift from traditional antimicrobial drug development to adaptive and biologically adaptive antimicrobial systems.

Endometriosis: A Mechanistically-Grounded Hypothesis for miRNA-Mediated Pre-EMT Apoptotic Regulation as a Therapeutic Strategy

Endometriosis is a long-lasting reproductive health condition caused by the abnormal development of endometrial-like tissue where it does not belong. According to worldwide estimates, 10% of all women of childbearing age are affected by endometriosis. The current treatments for endometriosis are limited due to the tendency of the disease to recur and the systemic side effects of existing hormonal therapies (including the use of the combined hormonal contraceptive pill). Recent studies have suggested that the development of endometriosis may be related to the dysregulation of apoptosis, improper epithelial-mesenchymal transition (commonly referred to as EMT), and changes in miRNA expression patterns. This paper proposes a mechanistically grounded therapeutic treatment strategy for endometriosis based on the proposal that inducing programmed cell death (apoptosis) in endometrial cells via miRNA-mediated mechanisms with specificity to the time just prior to activation of EMT may limit ectopic implantation during retrograde menstruation. A series of experimental phases, which are listed within this paper, is outlined to enable experimental evaluation of this hypothesis. By targeting the upstream survival signaling pathways, this approach will provide a potential non-hormonal therapeutic strategy for endometriosis.

Neural Correlates of Chronic Pain: Functional Connectivity and Neuroinflammation

Chronic pain is a long-term neurological disorder that affects millions of people each year and can be described as the experience of pain persisting longer than the expected timeline for healing of the tissue itself. Although extensive research has looked at chronic pain, the mechanisms that drive the pain through the central nervous system are still poorly understood. However, there is emerging research that suggests that chronic pain is related to disrupted functional connectivity of the central nervous system in pain processing network areas such as the prefrontal cortex (PFC), anterior cingulate cortex (ACC), insula, thalamus and somatosensory cortices in addition to neuroinflammatory signaling through microglia and astrocytes. The hypothesis is that chronic pain results from a maladaptive interaction between PFC inhibitory control and a persistent neuroinflammatory process that leads to increased nociceptive processing and central sensitization. I will test the hypothesis using a multimodal approach that combines resting state and task-based fMRI, MRS and TSPO-PET imaging to achieve a comprehensive understanding of functional connectivity, the excitatory/inhibitory neurotransmitter balance and the burden of neuroinflammation, as determined by neuroinflammatory markers. In tandem with the above-described studies, preclinical studies in animal models will be conducted to examine causal relationships between modulation of the prefrontal network, neuroinflammation and pain behavior. Ultimately, this integrative methodology will help identify convergent neural and inflammatory biomarkers to provide insight into potential network therapeutic strategies.

Postpartum Cardiovascular Risks: Hemodynamic and Endothelial Dysfunction

Pregnancy causes major cardiovascular physiological changes, such as an increase in blood volume and cardiac output, along with a drop in systemic vascular resistance and cardiac remodeling. Though most women experience some level of recovery after birth, a small number of women, especially those who developed gestational hypertension, preeclampsia, or gestational diabetes, show signs of long-term cardiovascular disease. Compared to the rest of the population, those who had these issues are more likely to suffer from cardiovascular disease due to the disruption of the endothelium process, vascular stiffness, and subclinical heart remodeling. It is my assumption that the reason behind postpartum cardiovascular dysfunction is persistent injury to endothelium combined with oxidative stress and imbalance of pro-and anti-angiogenic agents. Dysregulation of endothelium and stress in blood vessels can be tracked for up to one year after giving birth based on circulating levels of inflammatory agents (IL-6, TNF-α, hs-CRP), parameters of endothelium malfunction (VCAM, ICAM, endothelin), angiogenesis factors (sFlt-1, PlGF), and indicators of oxidative stress (MDA, IMA). To test this hypothesis, I propose the longitudinal cohort study based on the comparison of women with non-complicated and complicated pregnancies using echocardiography for heart function monitoring, vascular tests and biomarker evaluation procedure up to the postpartum period of one year.

The rank-one Dixmier conjecture for elements of mass at most six

Let K be a field of characteristic zero and let P, Q ∈ A1(K) satisfy [Q, P ] = 1. We prove that P, Q generate A1(K) if either has mass at most six, meaning at most six nonzero homogeneous components for the grading deg X = 1, deg Y=−1, with no degree bound on either element. This extends the four-component theorem of Guccione, Guccione and Valqui. The proof combines their Newton-polygon restrictions with a classification of polynomial pairs r, f satisfying an associated scalar differential equation under the condition that r2 has at most six monomial terms. The scalar argument uses a two-dimensional Vandermonde kernel, Rolle’s theorem, and a sign and parity obstruction. The scalar classification and the mass bound reduce the crossing cases to a binomial-power face. We exclude that face using at most one Newton cut, without a mass bound on the transformed pair. We also exclude a family of pure-power crossing faces with prime outer exponent, without a mass bound. A seven-term scalar solution shows that the scalar classification is sharp, although its associated Weyl face is excluded by the pure-power theorem. Theorem 1.1 is formalized in Lean 4 and registered as Palomar entry PALOMAR-2026-10-05-000006, version 2.

The directed Ramsey number R(7) is at most 46

Let R(k) be the least n such that every tournament on n vertices contains a transitive subtournament on k vertices. In 2022 Neiman, Mackey and Heule proved 34 ≤ R(7) ≤ 47. We prove R(7) ≤ 46. The argument looks at the in-degree of a single vertex of a hypothetical counterexample. It uses three computations. (i) Complete lists of the tournaments on 21 to 26 vertices that contain no transitive subtournament on 6 vertices. We obtain them by a SAT-based enumeration in which every job carries a checked LRAT certificate; the lists agree with previously published, uncertified files. (ii) Five new SAT instances on 46 vertices with 504 free arcs each. We solve them by cube-and-conquer; each of the 47,495 leaf cubes is closed by an LRAT proof that is checked while it is produced, and the cube cover is checked as well. (iii) The 28 SAT computations of Neiman, Mackey and Heule. They were published without proofs; we re-ran all of them with the same checked method (32,692 leaf cubes). Every proof was accepted by two checkers, one of which is formally verified. In Lean 4 we derive the theorem from twelve axioms: R(5) ≤ 14 and eleven statements of the computational results above. We list explicitly the parts of the argument that the proof checkers do not cover.

Eleven lines cannot form 33 Kobon triangles: a checked SAT proof over all pseudoline arrangements

The Kobon triangle problem asks for the largest number K(n) of bounded triangular faces in an arrangement of n lines in the plane. For n = 11 the best known arrangements have 32 triangles and the classical counting bound allows 33. In October 2026 the value a(11) was removed from OEIS A006066, because the lemma that extends the counting bound to arrangements in which three or more lines meet at a point has a counterexample. We show that no arrangement of 11 pseudolines, simple or not, has 33 bounded triangles; hence K(11) = 32. A counting argument reduces every arrangement with at least n(n-2)/3 triangles to two cases: simple arrangements, and arrangements in which three concurrent lines carry three particular triangles. Each case is a SAT instance over rank-3 chirotopes. The simple case is refuted by a single LRAT proof checked with lrat-check. The other case is refuted by a cube-and-conquer tree with 67,614 leaves, each refuted by an LRAT proof that was checked as it was produced, and 1,094 splits whose covers were checked in the same way. The computation took about 2,540 CPU hours.

Relay main study (preregistration): do session notes help AI research agents solve multi-session tasks?

In research tasks that take several sessions, with the work alternating between two AI apps (Claude Code with claude-sonnet-5-5, and Codex with gpt-6-luna), a short session note changes the share of runs solved within four sessions. The note has three parts: tried, failed, next. It is written after each session on the harness's request and read at the start of the next session. The comparison is with keeping only the workspace files. The test is two-sided, on two tasks with exact checkers (Kobon triangles for 8 lines; a covering system with distinct moduli all at least 13), with 90 runs per arm.