16работ
6соавторов
  1. Статья

    Accidents Open the Story and Choices End It in 136 Animated Films and Fifty Years of American Surveys

    Drobyshev D.

    In Kung Fu Panda (2008) the old master Oogway tells his student that there are no accidents. We

    asked how stories, people and the world actually treat chance. We coded every plot-driving event

    in all 81 features released by Pixar and DreamWorks Animation from 1995 to 2026 (1,189 events

    in 52,864 words of plot summaries) as caused by chance, by a character's choice, or by destiny,

    with five hypotheses committed before coding began. Chance drove 370 events (31%), choice

    800 (67%) and destiny 19 (2%). Chance hurt the protagonist about twice as often as it helped

    (56% against 29% of chance events), while choices helped about as often as they hurt (50% and

    42%). Chance was concentrated early: it caused 41% of events in the first fifth of a story and

    20% in the last fifth, it set off the inciting incident in 41 films, and it resolved the central conflict

    in none. Thirteen of the 14 films that invoke fate, prophecy or a chosen one resolved their

    conflict by a choice. Against our prediction, the films almost never declare a chance event to

    have been meant to be (4 of 370 chance events). A replication planned after these results and

    before any of it was run, on all 55 single-story features of Walt Disney Animation Studios from

    1937 to 2025 (783 events), reproduced the pattern and dated it. Before The Little Mermaid

    (1989), 6 of 19 Disney features ended by luck, fate or an outside rescue, and chance helped the

    hero about as often as it hurt; since 1989 the same was true of 2 of 36 Disney features and 4 of

    81 Pixar and DreamWorks films, and chance decided none of the 117 endings. In the General

    Social Survey, about one in seven Americans who chose a side said that luck rather than hard

    work matters most for getting ahead, a share that barely moved between 1973 and 2024.

    Holding that view went with a 7.3 point lower probability of being very happy (95% CI 5.7 to 9.0)

    after adjustment for age, sex, race, education, income and marital status, and the gap was of

    similar size in the lowest, middle and highest income thirds. Agreeing that miracles happen, or

    that life is meaningful because God exists, went with more happiness, but the difference shrank

    to statistical noise once religious attendance was held constant. Astrological sign, our test of

    pure chance, behaved like noise: 22 of 432 sign-against-rest tests passed p < 0.05 (5.1%) and

    none survived correction for multiple testing. Five simulations show why chance is easy to

    mistake for design: one-in-a-million events become routine at scale, a deterministic system loses

    a difference in the tenth decimal within 30 steps, the richest of 1,000 agents was also the most

    talented in 1.5% of 200 runs, and in lives built from 50 independent random shocks, removing

    the single largest shock would put the life on the other side of zero in one case in seven when

    shocks are thin-tailed and in almost one in four when they are fat-tailed. We call the pattern the

    Chance-Choice Asymmetry: chance may open a story, but choice has to close it. In these films

    the rule is a modern one. It took hold around 1989, and in the films since then it has had no

    exception.

    • chance
    • luck
    • coincidence
    • fate
    PDF DOI
  2. Препринт

    The Turning Point in Youth Happiness and Mental Health Across 75 Countries

    Drobyshev D., Chertovikov A., Erokhin A.

    For a decade the evidence on young people's mental health pointed one way. Symptoms of depression and anxiety rose from the early 2010s, most steeply among girls, and life satisfaction fell. In 2025 and 2026 several surveys reported the first improvements. This study asks whether the decline has turned, where, when, in which measures, and whether the turn came before or after the wave of school phone bans and social media age limits that governments adopted from 2023 onward.

    We harmonized 1,129 annual and wave series (12,223 data points) from 14 sources covering 155 countries into one database, and followed 572 series that reach 2023 or later in 93 countries. Each series received a joinpoint trend with up to two breaks, a wild-bootstrap interval for its turning year, and a test of whether its latest value lies below its peak by more than sampling error plus unexplained year-to-year variation. We sort series into three clocks by how the data are generated: experience (what young people report), care (what services record) and mortality (suicide deaths).

    The clocks disagree. Of 263 experience series that rose, 208 (79%) have turned and 20 (8%) are still rising; the median turning year is 2022. In PISA, the life satisfaction of 15-year-olds rose between 2022 and 2025 in 65 of 66 countries and economies, by 0.45 points on a 0 to 10 scale (60 increases significant). Girls gained 0.65 points and boys 0.21; girls gained more than boys in all 55 countries with data by sex, and the gap between boys and girls shrank from 0.72 to 0.29 points. Across 31 countries observed in all four waves, girls' life satisfaction went from 7.16 (2015) to 6.41 (2022) to 7.06 (2025). In the United States, persistent sadness among high school girls fell from 56.6% in 2021 to 43.4% in 2025, and major depressive episodes among girls aged 12 to 17 from 29.8% in 2021 to 21.8% in 2024. The care clock did not follow: of 35 care series that rose, 13 (37%) turned and 14 (40%) are still rising. Antidepressant users among Swedish girls aged 15 to 19 went from 36.9 per 1,000 in 2010 to 2013 to 104.6 in 2025, and ADHD stimulant users from 17.6 to 92.3. The mortality clock is the slowest: 28 of 77 youth suicide series that rose (36%) have turned, and the median suicide rate of women aged 15 to 24 across 102 countries rose from 2.50 per 100,000 in 2010 to 2014 to 3.24 in 2019 to 2023.

    Of 16 national series that turned in countries with a national or first state-level school phone measure, 15 turned before the measure applied and one, in the Netherlands, in the year it applied. In PISA, life satisfaction rose by 0.44 points in six countries that introduced a national school phone measure between the 2022 and 2025 waves and by 0.40 points in ten countries that had none (difference 0.05, permutation p = 0.56).

    Youth is no longer the happiest age everywhere. In Canada, 95.9% of adults aged 18 to 34 were satisfied with life in 2015 against 89.7% of those aged 65 and over; in 2025 the figures were 82.0% and 88.2%, and the young have been below the old since 2022. In six US states that asked both periods, dissatisfaction with life among 18- to 24-year-olds rose from 5.5% (2013 to 2017) to 9.2% (2022 to 2024). The young now report lower life satisfaction than people aged 50 and over in 12 of 36 OECD countries, all English-speaking, Nordic or western European apart from Japan, while in eastern and southern Europe they remain happier by 0.6 to 1.5 points.

    We call this pattern the Three-Clock Model: a crisis that ends in what young people report years before it ends in what services record and what death registers count. The model makes six testable predictions for 2026 to 2028. Conclusions hold when the window starts in 2012 (97% agreement), with a single joinpoint (98%) and after re-weighting PISA 2025 for item non-response (mean rise 0.450 instead of 0.452).

    • youth mental health
    • adolescent wellbeing
    • life satisfaction
    • subjective wellbeing
    PDF DOI
  3. Препринт

    Birds of a Feather: A Multidisciplinary Review of Neural, Evolutionary, and Psychological Mechanisms Underlying Similarity-Based Interpersonal Attraction

    Drobyshev D. O.

    People tend to befriend and marry people who resemble them. The finding is old and well replicated, but its explanations are scattered: social psychology, cognitive neuroscience, behavioral genetics and computational social science have each built their own account and rarely cite one another. This narrative review brings them together around three questions: which mechanisms link similarity to attraction, when the effect weakens or reverses, and how digital platforms change who meets whom.

    A meta-analysis of 313 studies found a correlation of r = .47 between actual similarity and attraction, but the effect was largest when participants rated a stranger they never met, and it was not significant in existing relationships. Perceived similarity (r = .39) predicted attraction at every relationship stage. Partners resemble each other strongly in political values (r = .58), religiosity (r = .56) and education (r = .55), and weakly in the Big Five (r = .08 to .21). In three national samples totaling 23,250 married people, personality similarity explained about 0.5% of the variance in satisfaction beyond each partner's own traits. In a prospective fMRI study of 41 MBA students scanned within days of arrival, neural response similarity predicted who were friends eight months later. Friends show excess genome-wide similarity of roughly fourth-cousin size, although population stratification may explain much of it. An agent-based model shows that assortative mating can arise from heritable variation in traits and preferences, with no dedicated mechanism for detecting similarity.

    The evidence supports a multilevel framework with four levels: genetic substrate, neural alignment, neurochemical reinforcement and cognitive-motivational channeling. At every level, most of the evidence concerns how relationships form, not how they last. The open questions are which dimension of similarity matters, at which relationship stage, and through which pathway.

    • similarity-attraction effect
    • assortative mating
    • actual similarity
    • perceived similarity
    PDF DOI
  4. Препринт

    The unsupervised deployment gap in generative AI for mental health

    Drobyshev D. O.

    Evidence about artificial intelligence in mental health and the deployment of artificial intelligence in mental health concern two different objects, and the first is routinely cited to justify the second. The evidence base is forty randomized trials pooling to g = 0.31 for depressive symptoms and g = 0.28 for anxiety, with 35 of 39 trials at high risk of bias, significant publication bias, and only 8 trials testing a generative system. The one randomized trial of a fully generative therapy chatbot enrolled 210 adults for four weeks, ran on Falcon-7B and LLaMA-2-70B, had clinicians review every response, and excluded active suicidality, mania and psychosis, turning away 215 people for suicide risk while enrolling 210 in total. Recomputing that trial's effect sizes from its own published means and standard deviations gives standardized mean differences of 0.45 to 0.62 against reported values of 0.63 to 0.90.   Deployment is three orders of magnitude larger. The operator of the most widely used general assistant estimates that in a given week 0.15% of active users have conversations containing explicit indicators of suicidal planning or intent and 0.07% show possible signs of psychosis or mania. Against 800 million weekly active users that describes roughly a million people every week.   This paper proposes a taxonomy of four deployment classes defined by design intent, regulatory claim, population gate, supervision and exposure, and shows that evidentiary strength and population exposure run in opposite directions across them. It argues that sycophancy in this setting is a problem of reward specification rather than a defect of alignment: therapeutic benefit cannot be observed inside a conversation, user approval can, and expert clinicians agree on whether a response is desirable only 71% to 77% of the time. It identifies two regulatory inversions in the four United States statutes passed since March 2025, and proposes psychovigilance: version identity, reporting attached to function rather than claim, independent access to logs, and outcome measurement at the level of whole conversations.

    • large language models
    • chatbots
    • sycophancy
    • reward specification
    PDF DOI
  5. Препринт

    Neuroinflammatory Mechanisms of Depression and the Collapse of the Monoamine Hypothesis

    Drobyshev D. O.

    For over four decades, the monoamine hypothesis — and its serotonin-centric variant in particular — served as the dominant explanatory framework for major depressive disorder. The 2022 umbrella review by Moncrieff et al., synthesizing evidence from six major research domains, found no consistent support for the claim that serotonin activity or concentrations are responsible for depression. This paper traces the historical trajectory of the monoamine model from the serendipitous discovery of iproniazid (1952) through the formalization by Schildkraut (1965) and Coppen (1967), the STAR*D trial revealing ~28% first-line remission rates, and the formal repudiation by Moncrieff and colleagues.

    The paper then presents the neuroinflammatory framework as a mechanistically grounded alternative. Meta-analytic evidence demonstrates consistent elevations of pro-inflammatory markers (IL-6, TNF-α, CRP, IL-1β) in approximately 25–35% of patients with MDD. The kynurenine pathway is identified as a mechanistic bridge: inflammatory activation of indoleamine 2,3-dioxygenase diverts tryptophan from serotonin synthesis toward neurotoxic quinolinic acid, reframing serotonin depletion as a downstream consequence of immune activation rather than a primary deficiency.

    Additional evidence is integrated from gut-brain-immune axis research (microbiome composition, intestinal permeability, vagal signaling) and metabolic psychiatry (the depression–obesity–diabetes triangle, insulin resistance, shared inflammatory pathways). Inflammatory subtyping using C-reactive protein as a stratification biomarker is discussed as a foundation for precision psychiatric intervention: patients with CRP > 5 mg/L show markedly better response to anti-inflammatory agents, while those with low CRP respond preferentially to SSRIs.

    The review evaluates emerging therapeutic approaches including celecoxib, minocycline, infliximab (in high-CRP subgroups), ketamine, exercise, and dietary interventions, with effect sizes ranging from 0.28 to 1.00 (Cohen's d). The synthesis argues that depression is not a unitary disorder of neurotransmitter deficiency but a heterogeneous syndrome in which immune dysregulation constitutes a major, and potentially treatable, pathophysiological dimension.

    • depression
    • neuroinflammation
    • serotonin hypothesis
    • monoamine hypothesis,
    PDF DOI
  6. Препринт

    Self-Destructive Drive as a Mechanism of Agonistic Flow in Extreme Human Performance

    Drobyshev D. O.

    This paper proposes agonistic flow as a distinct cognitive state accessed through sustained voluntary confrontation with physiological limits, integrating Csikszentmihalyi's flow theory, Dietrich's transient hypofrontality hypothesis, and Noakes's central governor model. We define diagnostic criteria, identify boundary conditions separating productive self-destruction from pathology, and review evidence across ultra-endurance sport, addiction recovery, cold exposure, and contemplative practice. The model extends to addiction's "rock bottom" phenomenon, reframing it as involuntary prefrontal breakdown producing the same clarity state that athletes access deliberately.

    • flow state
    • transient hypofrontality
    • self-destruction
    • ultra-endurance
    PDF DOI
  7. Препринт

    Verifiable Rewards in Language Model Training: A Taxonomy of Verifiability, GRPO Limitations, and Paths to Soft Verification

    Drobyshev D. O.

    We propose a five-level taxonomy of verifiability for NLP tasks ranging from exact match to subjective judgment. For each level we characterize the verifier type, its cost, and noise in the reward signal. We analyze GRPO sensitivity to noisy rewards and show that 15% noise in a binary signal degrades Pass@1 by 0.08. For levels 3-4 (soft verification) we describe three approaches: composite verifiers, claim-decomposition with NLI, and cascaded LLM judges. We compare GRPO, PPO, and DPO by noise robustness and propose a diagnostic pipeline for reward hacking detection under soft verification.

    • RLVR
    • GRPO
    • language models
    • verification
    PDF DOI
  8. Препринт

    The Inverted-U Under Scrutiny: A Meta-Analytic Test of the Yerkes–Dodson Law Across Cognitive Domains

    Drobyshev D. O., Maksimov D. A.

    The Yerkes–Dodson law — the claim that performance relates to arousal as an inverted U — is among the most cited principles in psychology. We test the claim by conducting an umbrella review of 36 meta-analyses and systematic reviews (covering over 1,200 primary experiments) on the relationship between acute stress, cortisol, or physiological arousal and cognitive performance. We then identified 40 primary experiments that manipulated arousal at three or more intensity levels and compared the fit of linear, quadratic, and null models using AICc. The quadratic model was preferred in 11 of 40 cases, concentrated almost entirely in working memory tasks under pharmacological cortisol manipulation (48% of pharmacological experiments vs. 5% of psychological-stress experiments). Episodic memory retrieval showed monotone impairment (d = −0.20 to −0.49). Executive control showed selective impairment: cognitive flexibility and cognitive inhibition declined while response inhibition improved (Shields et al., 2016, k = 51). Decision-making shifted toward reward-seeking (d = 0.17; Starcke & Brand, 2016, k = 32). Two neurobiological mechanisms account for the domain specificity: the mineralocorticoid–glucocorticoid receptor (MR/GR) occupancy balance produces a genuine inverted-U for cortisol acting on prefrontal working-memory circuits (confirmed by receptor-blockade studies: Rimmele et al., 2012; Deuter et al., 2024), while catecholaminergic modulation of neural gain via locus coeruleus–norepinephrine phasic/tonic modes (Aston-Jones & Cohen, 2005) produces monotone facilitation in subcortical systems. Critical moderators include sex (Liu et al., 2017, k = 34), hormonal contraceptive use (Jentsch et al., 2022), age (Lee et al., 2018), and timing relative to cognitive process (Shields et al., 2017, k = 113). None are captured by the Yerkes–Dodson framework. We propose replacing the broad law with a domain-specific, mechanism-grounded framework.

    • Yerkes-Dodson law
    • stress
    • cognition
    • inverted-U
    PDF DOI
  9. Препринт

    Semantic Segmentation of Aerial Imagery on a Consumer GPU with a From-Scratch Residual U-Net

    Drobyshev D. O.

    We present swath, an open-source pipeline for semantic segmentation of aerial and satellite imagery that covers the full workflow from raw data to georeferenced predictions. The system trains a residual U-Net from random initialization on the LoveDA benchmark, deliberately avoiding an ImageNet-pretrained encoder so that the architecture generalises to multispectral inputs with an arbitrary number of bands. Training runs on a single RTX 4060 Laptop GPU with 8 GB of VRAM, completing 30 epochs on 2,522 tiles in approximately 2.1 hours. The model reaches a mean intersection-over-union (mIoU) of 0.438 and an overall accuracy of 0.615 on 1,669 validation tiles evaluated at full 1024×1024 resolution through a sliding-window predictor with raised-cosine blending. We document the engineering decisions that make this possible on consumer hardware – mixed-precision training, adaptive accumulator placement, and self-describing checkpoints with safe training resumption – and report a domain gap between rural (mIoU = 0.379) and urban (mIoU = 0.471) scenes that published benchmarks typically obscure by averaging. Ablation experiments quantify the contribution of each memory-saving technique and the effect of class-weighting schemes on rare-class performance. The code, trained weights, and a FastAPI-based inference service are released under the MIT license.

    • semantic segmentation
    • remote sensing
    • U-Net
    • land cover
    PDF DOI
  10. Препринт

    Spatial Target Leakage in Location Intelligence: Quantifying Cross-Validation Bias from Spatial Autocorrelation

    Drobyshev D. O.

    Location intelligence models predict site-level outcomes using spatially derived features. Standard random cross-validation systematically overestimates their performance because nearby observations share spatially autocorrelated feature values, creating an information leak from test to train. We formalize this phenomenon as spatial target leakage, demonstrate it on a synthetic geomarketing task where random CV reports R² = 0.81 while spatially blocked CV reports R² = 0.43, and provide a minimal Python implementation of grid-based spatial blocking compatible with scikit-learn.

    PDF DOI
  11. Препринт

    stadion: A Benchmark for Operational Decisions with Classical Baselines and Exact Optima

    Drobyshev D. O.

    Most benchmarks used to evaluate autonomous agents share a structural limitation: nobody knows what the right answer was. A score of 61% on a browsing benchmark or an agent-versus-agent leaderboard says that one agent beat other agents on a particular task instance; it does not say whether the winning behaviour was close to what any reasonable procedure would recommend. We present stadion, an open-source benchmark for operational decisions in which the ground truth exists and is computable. Six tasks — inventory ordering, dynamic pricing of a perishable, admission to a finite-buffer queue, energy arbitrage against a daily price cycle, two-echelon supply-chain replenishment, and joint price-and-restock — are drawn from operations-research problems whose small state spaces admit exact solutions by backward induction. Every agent decision is scored against two references it cannot argue with: the classical operations-research method for the problem and the exact optimum. The result is a normalised score with a paired bootstrap confidence interval, on which indistinguishable from the classical method is a first-class outcome rather than a rounding error. We describe the scoring protocol, the six tasks, the harness self-verification procedure, and the headroom (0.4%–26.6%) between the classical rule and the optimum on each task — a distribution that argues against the common benchmark practice of selecting tasks where the classical answer is bad. The library is Python and treats reinforcement-learning policies and language-model agents through the same interface.

    • agent benchmarks
    • LLM agents
    • reinforcement learning
    • operations research
    PDF DOI
  12. Препринт

    A Systematic Review of Short-Form Video Effects on Attention, Memory, and Emotional Regulation

    Drobyshev D. O., Sviridov L. G., Maksimov D. A., Solodnikov E. A., Khoroneko L. A.

    This review synthesizes evidence from 47 empirical studies (2019–2026) on the effects of short-form video consumption on sustained attention, working memory, deep processing, and affective regulation. We identify four mechanisms: dopaminergic reward prediction error exploitation, attentional fragmentation, displacement of effortful cognition, and social comparison amplification. We propose the Attentional Fragmentation–Reward Habituation (AFRH) framework distinguishing passive scrolling from active engagement, and argue that the passive pathway drives the majority of observed cognitive costs.

    • short-form video
    • TikTok
    • attention
    • dopamine
    PDF DOI
  13. Препринт

    Computing Nearest-Amenity Walking Distance for Every Node in a Street Network in a Single Shortest-Path Pass

    Drobyshev D. O.

    Urban accessibility analysis requires computing the shortest walking distance from every address to the nearest instance of an amenity class. The naive approach runs one shortest-path computation per amenity location at cost O(k · n log n). We describe the fictitious super-source node technique from operations research that reduces the problem to a single Dijkstra pass at cost O(n log n), independent of the number of amenities. We extend the method to directed street networks via graph transposition, provide a complete Python implementation using OSMnx and NetworkX, and benchmark it on the street network of Moscow (680,000 nodes, 12,400 pharmacies), where the approach completes in 8 seconds versus over 14 hours for the naive method — a speedup of 6,000×.

    • shortest path
    • Dijkstra
    • street network
    • urban accessibility
    PDF DOI
  14. Препринт

    mlango: A Django-Inspired Framework for Machine Learning Lifecycle Management

    Drobyshev D. O.

    We present mlango, an open-source Python framework that applies Django's design philosophy — convention over configuration, batteries included, don't repeat yourself — to the machine learning lifecycle. The framework introduces four declarative families (Dataset, Model, Agent, Eval) built on a shared metaclass engine, content-addressed dataset versioning, three training backends (scikit-learn, PyTorch, Hugging Face Transformers), version comparison with McNemar's exact test, LLM agent development with cassette-based testing, a Django-style migration system, server-rendered admin UI, and FastAPI-based production serving with shadow deployment. The framework comprises 111 Python source files (~21,600 lines of code) with 1,320 tests and 84% coverage.

    • machine learning
    • MLOps
    • framework
    • model versioning
    PDF DOI
  15. Препринт

    decisionrl: A Reinforcement Learning Library for Operational Decisions with Classical Baselines

    Drobyshev D. O.

    We present decisionrl, an open-source Python library that provides (i) nine applied environments modeling canonical operational decisions (inventory management, dynamic pricing, energy arbitrage, queueing control, supply chain coordination), each paired with the best classical operations-research baseline; (ii) 32 reinforcement learning algorithms spanning tabular, value-based, actor-critic, model-based, offline, evolutionary, and RLHF methods, all sharing a uniform four-method interface; and (iii) a rigorous evaluation framework using interquartile mean with bootstrap confidence intervals. On stationary problems, RL matches but does not exceed the dynamic-programming optimum. On non-stationary, coupled, or partially observable problems, RL outperforms the best classical heuristic by 14–79%. The library is dependency-light (NumPy-only core, optional PyTorch), ships with ONNX/FastAPI production deployment, and is available under the MIT license.

    • reinforcement learning
    • operations research
    • inventory management
    • dynamic pricing
    PDF DOI
  16. Препринт

    dokimos: Practical Mutation Testing for Python via Per-Test Coverage Selection, Diff Scoping, and Verdict Caching

    Drobyshev D. O.

    Line coverage certifies that a line executed; it says nothing about whether any test would notice if the

    line were wrong. Mutation testing answers the second question directly by modifying the program

    under test and asking whether the suite complains. Its adoption has been limited by a straightforward

    cost — one full suite run per mutant — which for a project of moderate size turns a nightly rebuild into

    a multi-hour job. We present dokimos, a mutation-testing tool for Python that combines three cost

    reductions that individually are known but rarely used together in the Python ecosystem: (i) per-test

    coverage selection, running each mutant only against the tests that execute its line; (ii) diff scoping,

    mutating only the lines changed on a pull request; and (iii) verdict caching keyed on the tuple that

    determines the outcome. Mutations are applied through an import hook rather than by rewriting files on

    disk, so an interrupted run leaves the working tree byte-for-byte unchanged. On a representative

    fixture, dokimos executes 126 mutants in 14.0 s across eight workers and in 0.9 s with a warm cache,

    versus 397.6 s of a comparable serial baseline. The tool is open source, ships as a GitHub Action, and is

    applied to itself: on its own source it reports a mutation score of 64% [55%, 72%] over a 120-mutant

    sample. We discuss prior work on selective mutation, why the classical cost has kept the technique out

    of Python continuous integration, and the failure modes that survive the reductions.

    • mutation testing
    • software testing
    • coverage
    • test-suite adequacy
    PDF DOI