Project page. Methodology, data, and practical applications.
About the Project
“Digitalizing the Cultural Code” is an interdisciplinary research project at the intersection of Digital Humanities, the psychology of self-determination, and literary studies. Its goal is to develop and validate a methodology for the automated extraction of life scripts, worldviews, and retrospective life evaluations of literary characters from Russian-language texts (1800–2025) using large language models (LLMs).
The project rests on the premise that fiction functions as an archive and simulator of cultural models of life: writers reduce behavioral patterns to pure semantic structures, which society unconsciously uses as matrices for self-determination. By digitizing these structures through a system of measurable scales, we obtain a map of the cultural landscape of life scenarios — suitable both for scholarly analysis and for practical application.
This page collects the materials that supplement the main publication: full prompt texts, the validation protocol, the description of the practical tools, the open dataset, and technical documentation.
Project Chronicle
Current status: prompts v3.2 · Pilot-300 (300 profiles, validation underway) · Researcher Application (P-300) · 1 methodological case report (Medara) · “Scenario Contrast” tool (validation by educators and psychologists).
The project is evolving: prompts are calibrated, the pilot sample is expanding, and validation cases are being published. The chronicle records versions, decisions, and open questions — so that the methodology remains transparent and the results reproducible.
Go to the chronicles (in Russian) →
Main Publication
Title: Digitalizing the Cultural Code: Automated Extraction of Life Scripts from Russian-Language Literature (1800–2025) Using Large Language Models (Methodology, Pilot-300, Validation Infrastructure). Version 5.4.
Author: Anton U. Smirnov, Director of Research, Autonomous Non-Profit Organization “Center for Scientific Research in Career Guidance and Occupational Psychology”.
Abstract (abridged). The article presents a methodology for the automated extraction of life scripts, worldviews, and retrospective life evaluations from Russian-language literature (1800–2025). The architecture comprises four independent LLM agents (socio-situational, behavioral, psychological, and sociocultural), each of which works with the full text of a book within a single context window following a two-stage scheme (observation journal formation → discrete scoring), plus a fifth synthesizer agent that performs cross-validation and computes an integral reliability index as the minimum across scales. Ten measurable scales with multidimensional operationalization are proposed (declaration / behavior / inner position, domain-based crisis analysis, ambivalence of the authorial verdict, horizon of intent). The article reports results from a methodological pilot on six cases (including two crash tests) and Pilot-300: 300 character profiles, 78% passing the reliability threshold of 2.8; Gap_Flag is present in 65% of profiles. A three-tier validation system (text / expert / author) is introduced. The open infrastructure is described: a dataset on Zenodo (1901–1925 slice, CC BY 4.0), a researcher application with blind validation, and “Scenario Contrast” for adolescent career guidance. Limits of applicability, a plan for 10,000 texts, and cross-cultural portability are discussed.
Zenodo publication: 🌐 https://doi.org/10.5281/zenodo.23560692 (previous version 4.5 is available in the project archive).
Architecture and System of Scales
How the Pipeline Works
The analysis of a single work is performed through five sequential passes. The first four agents — socio-situational, behavioral, psychological, and sociocultural — work independently, each with the full text of the book in its own context window (1M+ tokens). Each agent follows a two-stage scheme within a single window: first, it forms an observation journal with mandatory quotations; then it assigns discrete scores on its assigned scales without returning to the raw text. The fifth agent — the synthesizer — does not read the book: it works with the journals from previous passes, performs cross-validation, computes the integral reliability index, and produces the final profile.
The refusal to use intermediate summarization is a deliberate design decision: any summary loses psychological subtext. The mandatory quotation base makes every claim verifiable, and cross-validation turns discrepancies between journals from errors into diagnostic signals.
System of Scales: 10 Measurable Parameters in Four Blocks
| Variable | English Name | Block | Values |
| Scale_of_Goals | Scale of life goals | I. Dynamics and Vector of Fate | 1 — local; 2 — social-status; 3 — global (+ Gap_Flag) |
| Social_Radius | Radius of social orientation | I | 1 — ego; 2 — micro; 3 — meso; 4 — macro (+ Perceived_Inclusion, Gap_Flag) |
| Scenario_Final | Scenario finale | I | 1 — collapse; 2 — stagnation; 3 — triumph (+ Closure) |
| Type_of_Agency | Type of agency | II. Kinematics of Action | 1 — passive; 2 — reactive; 3 — proactive |
| Crisis_Strategy | Crisis coping strategy | II | Five domains (ideological, love, existential, social, professional); in each: struggle / avoidance / conformism / NA |
| Scenario_Autonomy | Scenario autonomy | II | 1 — track; 2 — counter-script; 3 — alternative (+ Degree 0.0–1.0) |
| Locus_of_Control | Locus of control | III. Operating System of Personality | Internal / External (+ Gap_Flag) |
| Cognitive_Flexibility | Cognitive flexibility | III | 1 — rigidity; 2 — forced lability; 3 — conscious evolution (+ Evaluability) |
| Social_Validation | Social validation | IV. Cultural Echo | By group (family, estate, society): validation / rejection / ignorance / NA (+ Dominant, Salience) |
| Author_Verdict | Authorial verdict | IV | To character and to program: sympathy / irony / neutral (+ Counter_Verdict, Ambivalence_Level) |
The multidimensional attributes (in parentheses) are not an add-on: they allow the analyst to distinguish declaration, behavior, and inner position where a one-dimensional scale would have to choose a single label. Full definitions, examples, and the confounds table are provided in the article (Sections 2.5 and 3.2) and in data_dictionary.md (dataset, Zenodo).
Technical Specifications
| Parameter | Value |
| Model | DeepSeek-R1 (Chain-of-Thought) |
| Context window | 1M+ tokens (full book text in a single window) |
| Prompt version | v3.2 (entire Pilot-300, no intermediate changes) |
| Data schema version | v1.3 |
| Pilot size | 300 character profiles, 1800–2025, seven historical periods |
| Profiles with Total_Confidence_Score ≥ 2.8 | 234 (78%) |
| Profiles with Gap_Flag = true (at least one of three scales) | 195 (65%) |
| Dataset | 1901–1925 slice, CC BY 4.0, Zenodo |
| Validation type | expert panel (6 cases) + blind network validation (Pilot-300) + authorial (selective) |
Page updated: 22.10.2026. Prompt, schema, and dataset versions are fixed; the table is revised whenever any of them is updated.
Tools
“Scenario Contrast”: Career Guidance for Adolescents
A tool for supporting adolescent personal and professional self-determination, built on the Pilot-300 dataset. It is neither a test nor a diagnosis.
The adolescent configures an abstract hero along seven scales — in adolescent-friendly language, without psychological terminology: scale of goals, radius of social orientation, type of agency, crisis coping strategy, scenario autonomy, locus of control, cognitive flexibility. The three remaining scales (finale, social validation, authorial verdict) are not configured but observed: they are placed in toggle switches for endings and evaluations. The system finds “literary neighbours” — characters from various genres and epochs whose scenario profiles are close — and shows how, under similar settings, a path can end in collapse, stagnation, or triumph; receive recognition, rejection, or ignorance; evoke the author’s sympathy, irony, or neutrality.
The practical effect is threefold. First, the adolescent broadens their horizons: instead of a handful of “popular” books, they see dozens of works united by a scenario framework close to their own. Second, they begin to perceive life scenarios as variable and multiple — which is itself a tool for reflection. Third, they see their choices at scale and understand that identical settings do not guarantee an identical finale.
Key components: a configuration panel of seven scales with “what they say / what they do” discrepancy markers; a timeline of epochs (1800–2025) that can be toggled on and off; a distribution of outcomes in absolute numbers and percentages; toggle switches for endings and evaluations; an artifact — a configuration profile and a book list that is saved and used as material for a conversation with a psychologist; a specialist’s cabinet — the route taken and the final configuration, without interpretations or automatic conclusions.
The model is undergoing validation by career-guidance educators and psychologists.
🌐 Open Scenario Contrast → https://cc.iz-i.ru/
📄 Full description → https://cc.iz-i.ru/help.html
Researcher Application
An environment for reproducibility and collective validation. The application serves three purposes.
Analysis. Filtering by epoch, genre, character status, and confidence threshold; profile cards with a full quotation base; comparison in absolute and normalized values; dynamics across the entire timeline with epoch selection; CSV export. Any filter configuration is fixed in a persistent URL that can be sent to a colleague — they will see exactly the same view.
Blind validation. A standardized form: an expert rates a work on 10 scales without seeing the LLM’s scores or other experts’ responses. Responses accumulate by work; aggregates are opened once a accumulation threshold is reached.
Transparency. Aggregate validation statistics (without individual responses) are available in the application. This allows the robustness of conclusions to be tested at different confidence thresholds — the reader can reproduce any configuration and see how the sample changes.
A fixed slice is available on Zenodo for citation (1901–1925 epoch).
🌐 Open the application → https://ls.crcg.iz-i.ru/for-researchers.html
Appendices
Appendix A. Prompts v3.2 (Five Passes)
Full texts of the instructions for the four agents and the synthesizer, with a description of the two-stage scheme (observation journal → discrete scoring), rules for filling multidimensional fields, and the JSON contract format.
🌐 Download prompts v3.2 (PDF) → https://crcg.ru/boxes/ls/Prompts(v3.2).pdf
Appendix B. Expert Validation Protocol
Instructions for experts, scale definitions, scoring rules, the procedure for resolving disagreements (majority consensus for categorical scales, median for ordinal scales, an independent arbiter in case of divergence), and the report cross-check form.
🌐 Open the protocol PDF and cross-check prompt → https://crcg.ru/boxes/ls-appendix_app3_protocol.pdf
Appendix C. The Medara Case: Full Methodological Report
Three-tier validation (LLM / expert / author), a table across 10 scales, diagnosis of divergences, and open questions. Includes the author’s response (with permission) and the expert annotation JSON.
🌐 Go to the Medara case report → https://crcg.ru/ls-project-log/medara-case
Appendix D. Researcher Application: Description and Functionality
Filters, profile cards with a quotation base, the blind validation form, aggregate statistics, persistent URLs, and CSV export.
🌐 Open the application → https://ls.crcg.iz-i.ru/for-researchers.html
Appendix E. Dataset (Zenodo)
1901–1925 slice, CC BY 4.0 license; data/*.json, dataset.json, manifest.csv, schema.json, data_dictionary.md.
🌐 Go to the dataset → https://doi.org/10.5281/zenodo.23526136
Appendix F. Version Mapping: Prompts v3.2 → Data Schema v1.3
A technical table mapping schema fields to prompt instructions; confirmation that the entire Pilot-300 was produced on a single prompt version without intermediate changes.
🌐 Go to Appendix F (PDF) → https://crcg.ru/boxes/ls/Appendix_Zh.pdf
🌐 Go to Appendix F (Markdown) → https://crcg.ru/boxes/ls/Appendix_Zh.md
Project Archive
- Initial publication. “Automated Extraction of Life Scripts from Russian Literature Using Large Language Models (LLMs): Methodology and Limits of Applicability” (version 4.5). DOI: https://zenodo.org/records/22825133
- Prompts v1.0. Full texts of the system prompts for the five passes (socio-situational, behavioral, psychological, sociocultural, synthesizer). 🌐 Open PDF
- Prompts v2.0. Full texts of the system prompts with calibration: three-level operationalization, domain-based crisis analysis, ambivalence of the authorial verdict. 🌐 Open PDF
- Dataset (P-6, v1.1). Open dataset on Zenodo: JSON, all methodological pilot cases (P-6), including crash tests. 🌐 Go to the dataset (Version 1.1).
Contacts
Project author: Anton U. Smirnov, Director of Research, Autonomous Non-Profit Organization “Center for Scientific Research in Career Guidance and Occupational Psychology”.
Email: smirnov@crcg.ru
Project page: crcg.ru/ls-appendix-eng/
For collaboration: we invite psychologists, linguists, literary scholars, and text-processing specialists to participate in expert coding and interpretation of the results.
Link to the Russian Version
Русская версия страницы: https://crcg.ru/ls-appendix/
Full appendices (prompts, validation protocol, dataset description, version mapping) are available on the Russian-language page.