КурсГайдыСтатьиПодкастGitHubОбо мнеНачать бесплатно

Научные работы

Публикации

Евгений Арсентьев, PhD. 6 рецензируемых работ, 69 цитирований по данным OpenAlex на 2026-09-10. Сейчас — измеримая экономика и надёжность ИИ-агентов.

11
работы
69
цитирования
27
журнала
8
реестра
10
членства
16
ID автора

Идентификаторы

Отчёты с DOI

2026·Zenodo - technical report (not peer-reviewed)

No Request With a Known Crawler User-Agent Reached llms.txt: Fifteen Days of Logs at One Origin Server

Fifteen days of logs at one origin server: 46,155 requests presenting known search and AI crawler user-agents, none for llms.txt, while robots.txt was requested 698 times in the two days after the file went live. Corrects the numbers in draft-arsentev-llm-context-discovery-00.

DOI: 10.5281/zenodo.23018429 · серия: 10.5281/zenodo.23018428 · PDF

2026·Zenodo - technical report (not peer-reviewed)

AI Content Engine Economics: The Cost of Short-Form Video Production by AI Agents and by People

One company's AI content engine produced 1,497 short-form videos across 67 channels in 101 days for $4,313.76 in subscriptions: $0.52 per 1,000 organic views and $2.88 per video. The same volume from a human studio would take 42–78 people and $179,000–1.21 million a month; buying the reach from influencers would cost 3.8–57 times more. Built by one person with no technical background; code written by Claude agents. Company under NDA, figures and calculations only.

DOI: 10.5281/zenodo.22802229 · серия: 10.5281/zenodo.22802228 · PDF

2026·Zenodo - technical report (not peer-reviewed)

Clear Every Third Task: A Measured U-Curve in the Context Economy of Coding Agents

Corrected 2026-09-15. How often should an agent's context be cleared? 36 runs, six session-length policies, six replicates each, and no failure among the tests the agent wrote (4086 in total). Cost is U-shaped in session length: clearing after every task costs 33.5% more than clearing every third (p = 0.0022), while never clearing costs 9.7% more, which is not significant (p = 0.046, Holm-adjusted 0.14). Three, four and six tasks per session are indistinguishable, so the low region is a plateau. The U decomposes into cache writes falling and context per call rising, with 1-hour cache writes priced 20:1 against cache reads.

DOI: 10.5281/zenodo.22759217 · серия: 10.5281/zenodo.22699667 · PDF

2026·Zenodo — technical report (not peer-reviewed)

The Context Economy of Agentic LLM Sessions: Where the Money Actually Goes

Corrected 2026-09-15. 722 agent sessions, 150,902 model calls, 34.6 billion tokens. Context handling accounts for 87.8% of modeled cost and generation for 12.2%; by estimate, 80% of the spend comes from 2.9% of sessions. Includes two log-deduplication traps that change the answer by a factor of two in either direction, and the de-identified dataset.

DOI: 10.5281/zenodo.22759216 · серия: 10.5281/zenodo.22688704 · PDF

Препринты

2026-09-14·Qeios — article, preprint v1 (open peer review)

A Suite That Does Not Discriminate: The Case for Result-Set Fields in Agent Conformance Records

A set of conformance checks can consist entirely of sound, correctly reported, individually falsifiable checks and still support no conclusion, because no check changed state across the conditions compared. That is a property of the set, and no per-check field can carry it. The paper proposes a result-set field recording whether any check discriminated and which ones; it documents the adjacent case in a published corpus of 36 coding-agent runs, where a verdict of 4,086 tests and zero failures held with zero variance across six conditions whose modelled cost differed by 25.3% (a figure computed with the incorrect prices of report versions 1.0 and 1.1; 33.5% at the corrected prices of version 1.2, DOI 10.5281/zenodo.22759217); it reports as negative the tempting explanation that longer sessions simply did less work; and it identifies a case the proposed field cannot answer — where the checks are authored by the run itself, check identity does not survive across runs and the comparison-level question is undefined rather than false.

DOI: 10.32388/0BV3Z8

Программы с DOI

2026·v0.2.1 · MIT

contextburn

Open-source meter built from the context-economy experiments. Reads the transcripts Claude Code already writes locally and reports run efficiency: the share of paid tokens that became model output rather than a re-reading of context already sent, by tokens and cost-weighted. Makes no network calls.

DOI: 10.5281/zenodo.22712985

В каталогах: OECD.AI Catalogue of Tools & Metrics · Research Software Directory

Открытые данные и видео

2026·Zenodo · CC BY 4.0

AI Tools Radar: GitHub and Hugging Face projects selected by arsentev.ai (monthly)

490 open-source AI projects selected and reviewed by the arsentev.ai radar since June 2026, with daily GitHub and Hugging Face API snapshots of how they evolve after selection. A new version is released every month.

DOI: 10.5281/zenodo.22730450

2026·Hugging Face · CC BY 4.0

Context U-curve: 36 coding-agent runs under six context-clearing policies

Run-level token counters, modeled cost, wall clock and test outcomes for every run behind the U-curve report. Counters only - no prompts, no model output, no paths.

DOI: 10.57967/hf/10366

2026·Hugging Face · CC BY 4.0

Context economy: modeled cost of 722 coding-agent sessions

Per-session model calls and modeled cost at list prices for the 722 sessions behind the Context Economy report (v1.1). Counters only - no prompts, no code, no project names.

DOI: 10.57967/hf/10649

2026·Kaggle · CC BY 4.0

Context U-Curve of Coding Agents (36 runs)

The same run-level dataset on Kaggle, with a data card and per-column profiles.

2026·OSF · CC BY 4.0

Clear Every Third Task - report and data on OSF

Open Science Framework project holding the report PDF, the run-level dataset and its card.

DOI: 10.17605/OSF.IO/5QTWY

2026·YouTube · Zenodo · Internet Archive · CC BY 4.0

contextburn demo: run efficiency of 36 coding-agent runs

Twenty-second terminal recording of contextburn over the experiment's session logs: 168 sessions, 94.4% of paid tokens re-read context already sent. Archived at archive.org/details/contextburn-demo-context-ucurve-2026.

DOI: 10.5281/zenodo.22713920

Рецензирование

Опубликованные рецензии

Письма в журналы (eLetters)

Публичные комментарии регуляторам и парламенту

Членства и стандарты

  • W3C

    Участник сообщества и рабочих групп

    Участник, публичный список группы Cognitive AI Community Group

    профиль →
  • W3C

    Редактор версии 0.1 формата отчётности, Agent Conformance and Benchmarking Community Group

    Группа разрабатывает воспроизводимые методы измерения соответствия ИИ-агентов. Редактором версии 0.1 назначен участниками в открытой рассылке 17–18.09.2026; версия 0.1 выпущена в открытой рассылке 30.09.2026; документ группы, не стандарт W3C

    профиль →
  • IETF

    Автор двух интернет-драфтов: метрики прогонов агентов и обнаружение контекстных файлов

    draft-arsentev-agent-run-metrics · draft-arsentev-llm-context-discovery

    профиль →
  • MLCommons

    Участник рабочих групп

  • AI Standards Hub

    Член сообщества (Alan Turing Institute, BSI, NPL)

  • FORCE11

    Член сообщества научной коммуникации

    профиль →
  • Research Data Alliance

    Член

    Членство с сентября 2026, участник 13 рабочих и тематических групп

    профиль →
  • OpenInfra Foundation

    Член

    профиль →
  • Open Source Initiative

    Член

  • Internet Society

    Член, в том числе AI Special Interest Group

Ранние рецензируемые работы

6 рецензируемых работ 2017–2019 годов, 69 цитирований. Индексированы в Scopus, Web of Science и PubMed. Тема — экономика и организация здравоохранения; к нынешней работе об ИИ-агентах отношения не имеет, поэтому полный список вынесен в реестры.

Полный список — ORCIDПолный список — Google ScholarПолный список — ScopusПолный список — ResearchGate