Leandro Z. Agudelo

Judge scores increasingly control downstream actions in AI systems, yet a threshold calibrated at one operatingpoint can retain authority after the judge or workload changes. The threshold remains fixed while its relationship tocorrectness, cost, and accepted-cohort performance can drift. Here, we introduce the stale gate, a versioned operationalcontract that links score-based routing to an explicit decision cost and two monitored promises: accepted outcomesshould remain consistent with the judge’s stated probabilities, and the accepted cohort should retain a declared referenceperformance. For calibrated scores, we derive the cost-minimizing threshold and use predeclared restart banks ofanytime-valid e-processes to test whether either contract has expired. In controlled experiments, we show that additiveoverconfidence of 0.15 increased accepted failure from 0.241 to 0.320 and expected cost by 0.327 per task. We findthat the restart bank alarmed in 16 of 5,000 no-drift deployments, tracked first-order delay within 1–4% over fiveshifts, and reduced mean epoch-8 detection delay from 585 to 156 audited labels. Across 80 prospective API-serveddeployments, we observe that the same judge edit produced a 0.424 claim gap with onset-epoch alarms in 5/5 GSM8Kdeployments but a 0.068 gap with no transition in 10/10 MBPP deployments; an honestly reported harder arithmeticworkload changed routing without triggering a claim-contract alarm. These findings show that score authority shouldbe treated as versioned evidence: drift can be priced, failed contracts distinguished, and gates recalibrated when theevidence supporting their decisions no longer applies. Together, these results provide a basis for managing the validityof judge-gated decisions over time, while studies across broader models and long-horizon settings can determine howthese controls generalize to evolving AI systems.

Agent evaluations can compress stochastic generation, execution, and model-based judging into a single score,obscuring whether a measured difference reflects a change in the system or variability in the evaluation itself. Here,we introduce the determinism ladder, a measurement framework that compares fresh generation, temperature-zerogeneration, and content-addressed replay through executable verification, together with a frozen-candidate experimentthat distinguishes variability in judge verdicts from disagreement with executed outcomes. Across HumanEval+ andMBPP+ in two hosted model-and-verifier stacks, we found that replay nearly eliminated the observed within-taskvariation. In the primary stack, live variances of 0.0301 and 0.0142 fell to 0 and 0.0010 under replay, while a secondstack reproduced the same ordering. Temperature zero also reduced generation variance in both stacks, althoughthe remaining variation was stack-specific. On 200 frozen MBPP+ candidates, a stochastic judge introduced verdictvariance of 0.0073. Setting the judge temperature to zero removed the observed verdict variation, yet majority verdictsstill disagreed with executed outcomes on 23.5% of tasks, demonstrating that repeatability and reference agreementare distinct properties of an evaluation. For an illustrative fixed-task difference of 0.02, the measured variancescorresponded to point budgets of 3–8 runs per arm under live evaluation and one under valid replay, subject to the statedplanning assumptions. Together, these results show that evaluation variability can be measured under stage-specificcontrols rather than absorbed into a single score, providing a basis for matching replication and evaluation design tothe claim an agent benchmark is intended to support.

Adaptive AI systems change after evaluation through updates to models, memory, prompts, and tools. Yet, an earlierscore may remain numerically correct while no longer describing the current system. Reuse of evaluation outcomesfor optimization can also compromise the independence of that evidence. Here, we introduce the ContaminationClock, a runtime protocol that links scores to system snapshots and evaluation histories. We isolate measurement fromupdates and record provenance and reuse to assess historical validity, current applicability, and permission for reuseseparately. In three prospective studies, we tested whether memory containing evaluation outcomes preferentiallyimproved performance on reused tasks. In the first study, we found greater gains on reused than on sealed tasks in allten evaluation episodes after adjustment for matched unrelated memory (mean adjusted contrast 0.469; 95% interval[0.377, 0.560]). We reproduced this advantage in another model class, while unrelated-memory shams yielded anear-zero mean contrast. By varying exposure, we further showed that the mean contrast rose from 0.119 to 0.453 asexact traces were supplied for 25% to 100% of reused tasks. We also found that token reordering preserved the meancontrast, whereas removing the verified decision from each trace reduced it to−0.013. By comparison, a leak-freefine-tuning update did not trigger the contamination alert (differential point estimate 0.069; frozen bounds ±0.12). Inthis controlled test, we nevertheless treated the earlier claim as superseded because the model had changed. This workprovides a protocol for making reevaluation an explicit, auditable part of agent development while preserving validhistorical evidence. Future studies with larger models and semantic memory could test whether the protocol keepsperformance claims reliable across successive updates.

Program libraries guide code generation, but a search may return no code or documentation relevant to the requested functionality. In this case, abstaining may exclude solvable tasks, whereas proceeding may waste computation on tasks likely to fail. Here, we introduce the synthesis abstention card, a verification framework for deciding whether missing library information justifies declining a task before generation. The protocol combines confidence intervals for support--failure linkage, retained risk, and calls avoided with checks of transport to the target workload. These checks classify decisions as supported, non-actionable, or uncertified against prespecified requirements. We examined these requirements through controlled studies and external audits. In a small, locally trained model, we found that authoritative implementation notes reduced the probability that all three generation attempts failed from 0.910 to 0.414. In a separate controlled system, abstention avoided 50.6% of generation calls when half the tasks lacked relevant library information. Abstention also reduced failure risk to 0.129 among retained tasks, compared with 0.366 when all tasks were attempted. These benefits persisted when 75% of tasks lacked library information, although retained risk was overestimated beyond the prespecified tolerance. The external audits revealed a further limitation: generic text signals could pass transport checks while providing little failure discrimination. Together, these findings show that library signals can justify abstention only when failure discrimination, retained risk, and savings are validated for the target workload. The card provides a basis for reassessing which tasks an agent is justified in attempting as its knowledge, capabilities, and workload change.

Functional specialization in continuous systems requires balancing adaptation to environmental stress with the preservation of encoded information. Yet, the physical constraints governing how living systems reconcile selective information retention with energetic structural reorganization remain unclear–a trade-off that may offer transferable design rules for adaptive computing. Here we present a unified framework that connects cellular specialization to adaptive machine learning through principles of nonequilibrium optimization. We find that asymmetric fluctuations in key parameters act as tunable drives that, when phase-aligned, are rectified into durable functional specialization. Biologically, we show that in adipose mesenchymal cells, transient nutrient deficits trigger cytosolic pH fluctuations. These oscillations, regulated by the ATP12A proton pump, modulate nuclear compartment dynamics to coordinate mitochondrial specialization and lock in epigenetic memory for phenotype persistence. Notably, modulating ATP12A amplifies these drives, enhancing functional specialization and metabolic homeostasis in disease models. Computationally, we formalize analogous principles using a compositional framework based on operadic algebra, where structural updates act as a topological ratchet converting drive fluctuations into persistent architectures. This reveals a learning regime where gains in predictive information are accompanied by auditable residual costs for non-predictive states and reorganization. As a proof of concept, we deploy a self-organizing machine learning system that discovers latent hierarchy and reconstructs data through adaptive compartment dynamics. We show that internal organization fluctuations increase with residuals, an information-processing ledger guiding adaptive computation. Across both domains, local organization gains are directionally linked to dissipation proxies (pH variance in vivo and residuals in silico), consistent with operation under shared constraints. Overall, this work identifies asymmetric-drive rectification as a fundamental mechanism of adaptive specialization, providing thermodynamically informed design rules for both bioengineering and adaptive AI. One-Sentence Summary: Asymmetric-drive rectification is a fundamental mechanism by which cells and adaptive computational systems convert internal stress into structural memory, encoding functional specialization

Caloric restriction reduces metabolic disease and associated comorbidities. Yet, the molecular mechanisms encoding cellular memory of these benefits remain unclear. Here, we use a functional genomics approach to integrate evolutionary cues, single-cell sequencing, and metabolomics, identifying convergent signals that encode epigenetic memory of adaptive states in visceral adipose tissue natural killer (NK) cells. Cross-species analysis shows that genomic hubs linked to human accelerated regions (HARs) are conserved and active during dietary restriction, forming transcriptional compartments that regulate oxidative stress response and DNA repair genes. Targeted screens and multimodal profiling across human data and mouse models reveal convergent signals needed for persistent adaptation: cooperative transcriptional regulators (NRF2, CIRBP, NR4A2) at key HAR-linked genomic hubs, innate immune mediators (IL15-IL2RB), and metabolic cofactors from linoleic acid oxidation. Activity of these convergent signals reduces DNA damage and methylation and enhances epigenetic plasticity and cytotoxic functions, thereby decreasing tissue fibrosis and senescence while maintaining both local and systemic metabolic plasticity. Mechanistically, linoleic acid metabolism supplies acetyl-CoA to sustain H3K27ac epigenetic marks that maintain coordinated repair and cytotoxic programs after programming. IL15-IL2RB signaling links tissue metabolic state to NK cell function for coherent adaptive responses. We show that engineering of NK cells with this multimodal programming leads to long-term preservation of therapeutic phenotypes in metabolic aging models, rescuing systemic dysfunction through damaged cells clearance and leptin sensitivity. Overall, these findings establish that convergent signals encode persistent cellular phenotypes linked to metabolic plasticity, suggesting design principles for engineering therapeutic memory of cell function.

Energy conservation is at the basis of evolutionary optimization in living systems. Yet, how cells maintain function during metabolic stress while preserving adaptive capacity remains unclear. Here, we use a functional genomics approach to identify metabolic human accelerated regions (mHARs) genomic hubs (evolutionarily tuned regions) in adipose tissue that coordinate adaptive responses to nutrient deficit. During fasting, cooperative regulators PGC1A, YY1, and CTCF activate these hubs, promoting chromatin contacts that regulate lipid cycling genes such as Dgat1 and Atg4c. This activation maintains lipid and pH homeostasis while limiting oxidative stress, thereby preserving cellular plasticity. When disrupted, nutrient-stress adaptations are impaired leading to metabolic dysfunction and glucose intolerance. This is driven by NCOA7, an oxidative stress mediator that impairs lysosome function and transcriptional plasticity, triggering an inflammatory response. In cellular models, this mechanism blocks pH-mediated PGC1A droplet plasticity, further inhibiting mHAR-genomic hub activity and metabolic homeostasis. Human adipocytes show enhanced transcriptional activity of these genomic hubs compared to murine cells, with multi-enhancer mHAR regulation suppressing disease-associated variants linked to ANGPTL3 and PCSK9. Loss of this cooperative plasticity shifts cells to maladaptive responses, activating STAT1-driven disease gene expression. Our findings reveal that investigating the contextual activity of evolutionary adaptations can uncover functional hubs critical for stress resilience. This also shows that cells exploit nutrient fluctuations to drive tunable transcriptional organizations that lead to durable adaptations, identifying actionable targets for metabolic disease.

Objective To delineate organ-specific and systemic drivers of metabolic dysfunction-associated steatotic liver disease (MASLD), we applied integrative causal inference across clinical, imaging, and proteomic domains in individuals with and without type 2 diabetes (T2D). Research Design and Methods We used Bayesian network analyses to quantify causal pathways linking adipose distribution, glycemia, and insulin dynamics with fatty liver using data from the IMI-DIRECT prospective cohort study. Measurements were made of glucose and insulin dynamics (using frequently-sampled metabolic challenge tests), MRI-derived abdominal and liver fat content, serological biomarkers, and Olink plasma proteomics from 331 adults with new-onset T2D and 964 adults free from diabetes at enrolment. The common protocols used in these two cohorts provided the opportunity for replication analyses to be performed. When the direction of the effect could not be determined with high probability through Bayesian networks, complementary two-sample Mendelian randomization (MR) was employed. Results High basal insulin secretion rate (BasalISR) was identified as the primary causal driver of liver fat accumulation in both diabetes and non-diabetes. Excess visceral adipose tissue (VAT) was bidirectionally associated with liver fat, indicating a self-reinforcing metabolic loop. Basal insulin clearance (Clinsb) worsened as a consequence of liver fat accumulation to a greater degree before the onset of T2D. Out of 446 analysed proteins, 34 mapped to these metabolic networks and 27 were identified in the non-diabetes network, 18 in the diabetes network, and 11 were common between the two networks. Key proteins directly associated with liver fat included GUSB, ALDH1A1, LPL, IGFBP1/2, CTSD, HMOX1, FGF21, AGRP, and ACE2. Sex-stratified analyses revealed distinct proteomic drivers: GUSB and LEP were most predictive of liver fat in females and males, respectively. Conclusions Basal insulin hypersecretion is a modifiable, causal driver of MASLD, particularly prior to glycaemic decompensation. Our findings highlight a multifactorial, sex-and disease-stage–specific proteo-metabolic architecture of hepatic steatosis. Proteins such as GUSB, ALDH1A1, LPL, and IGFBPs warrant further investigation as potential biomarkers or therapeutic targets for MASLD prevention and treatment.

Regular physical exercise has long been recognized to reverse the effects of diet-induced obesity, but the molecular mechanisms mediating these multi-tissue beneficial effects remain uncharacterized. Here, we address this challenge by studying the opposing effects of exercise training and high-fat diet at single-cell, deconvolution and tissue-level resolutions across 3 metabolic tissues. We profile scRNA-seq in 204,883 cells, grouped into 53 distinct cell subtypes/states in 22 major cell types, from subcuta-neous and visceral white adipose tissue (WAT), and skeletal muscle (SkM) in mice with diet and exercise training interventions. With a great number of mesenchymal stem cells (MSCs) profiled, we compared depot-specific adipose stem cell (ASC) states, and defined 7 distinct fibro-adipogenic progenitor (FAP) states in SkM including discovering and validating a novel CD140+/CD34+/SCA1-FAP population. Exercise- and obesity-regulated proportion, transcriptional and cell-cell interaction changes were most strongly pronounced in and centered around ASCs, FAPs, macrophages and T-cells. These changes reflected thermogenesis-vs-lipogenesis and hyperplasia-vs-hypertrophy shifts, clustered in pathways including extracellular matrix remodeling and circadian rhythm, and implicated complex single- and multi-tissue communication including training-associated shift of a cytokine from binding to its decoy receptor on ASCs to true receptor on M2 macrophages in vWAT. Overall, our work provides new insights on the metabolic protective effects of exercise training, uncovers a previously-underappreciated role of MSCs in mediating tissue-specific and multi-tissue effects, and serves as a model for multitissue single-cell analyses in physiologically complex and multifactorial traits exemplified by obesity and exercise training.

Metabolism plays a central role in evolution, as resource conservation is a selective pressure for fitness and survival. Resource-driven adaptations offer a good model to study evolutionary innovation more broadly. It remains unknown how resource-driven optimization of genome function integrates chromatin architecture with transcriptional phase transitions. Here we show that tuning of genome architecture and heterotypic transcriptional condensates mediate resilience to nutrient limitation. Network genomic integration of phenotypic, structural, and functional relationships reveals that fat tissue promotes organismal adaptations through metabolic acceleration chromatin domains and heterotypic PGC1A condensates. We find evolutionary adaptations in several dimensions; low conservation of amino acid residues within protein disorder regions, nonrandom chromatin location of metabolic acceleration domains, condensate-chromatin stability through cis-regulatory anchoring and encoding of genome plasticity in radial chromatin organization. We show that environmental tuning of these adaptations leads to fasting endurance, through efficient nuclear compartmentalization of lipid metabolic regions, and, locally, human-specific burst kinetics of lipid cycling genes. This process reduces oxidative stress, and fatty-acid mediated cellular acidification, enabling endurance of condensate chromatin conformations. Comparative genomics of genetic and diet perturbations reveal mammalian convergence of phenotype and structural relationships, along with loss of transcriptional control by diet-induced obesity. Further, we find that radial transcriptional organization is encoded in functional divergence of metabolic disease variant-hubs, heterotypic condensate composition, and protein residues sensing metabolic variation. During fuel restriction, these features license the formation of large heterotypic condensates that buffer proton excess, and shift viscoelasticity for condensate endurance. This mechanism maintains physiological pH, reduces pH-resilient inflammatory gene programs, and enables genome plasticity through transcriptionally driven cell-specific chromatin contacts. In vivo manipulation of this circuit promotes fasting-like adaptations with heterotypic nuclear compartments, metabolic and cell-specific homeostasis. In sum, we uncover here a general principle by which transcription uses environmental fluctuations for genome function, and demonstrate how resource conservation optimizes transcriptional self-organization through robust feedback integrators, highlighting obesity as an inhibitor of genome plasticity relevant for many diseases.