Multi-Platform Observational & Sociodynamic Analysis of Individual Configurations
This repository is a proof of concept. MOSAIC is a sociological framework for multi-platform behavioral analysis, built in papers before it was built in code. What you find here is its first test bench: the place where the theory meets real data, and where it is allowed to break.
Consider an illustrative case. The same person takes part in different digital environments. On GitHub, their contributions appear methodical and collaborative. On Reddit, their exchanges are more confrontational. In YouTube comments, they are shorter and more dismissive.
Most tools would ask which description best captures the person. MOSAIC asks a different question: what explains the variation across contexts, and what can legitimately be inferred from it?
Not who someone is. Where they stand.
The framework's central hypothesis:
Digital traces should not be interpreted as stable behavioral signals but as structured positional variance across socio-technical environments.
The variance between platforms is not noise to filter out. It is the signal. The framework proposes to read it through three components:
- What holds: what persists across contexts.
- What adapts: what varies coherently with the constraints and incentives of each platform.
- What slips: what neither of the first two fully accounts for.
These are analytical categories, not discoveries. "What slips" does not reveal a hidden "true self" behind the public presentation; it marks where the first two readings stop explaining the data. The three are read as a structure, never collapsed into a score.
This is a conceptual framework. Its hypotheses are not yet empirically validated, and the current code does not demonstrate them.
These distinctions are methodological commitments, not guarantees. Theoretical intent alone does not determine what a system does: the data it collects, the outputs it generates and the uses it makes possible matter just as much.
| Designed not to | Designed to |
|---|---|
| Infer personality | Analyze how expression varies across contexts |
| Predict behavior | Describe the moves a position makes plausible |
| Profile psychologically | Map a position within a field |
| Surveil | Produce a reading at a given moment, indexed in time |
Each commitment holds only if practice follows it. "Plausible" remains an inference, not a validated prediction. Describing a position is not in itself a psychological profile, but it can drift toward one if it attributes stable characteristics to a person. A time-indexed reading is not in itself surveillance, but it can serve it if collection, consent, retention and use do not follow the same commitment.
The grandmaster doesn't need to know the player. The position is enough.
| Status | |
|---|---|
| Theory: three-level architecture, signal decomposition, positional capital, epistemological commitments | Paper I — working paper, not peer-reviewed |
| Method: protocols for each component, grammar of analytical decisions, validation | Paper II — in preparation |
| Code: multi-platform collection, structuring, LLM-assisted reading | Alpha PoC — this repo |
The code is deliberately behind the theory. It collects, structures, and can produce LLM-assisted outputs. Those outputs are not yet a complete or reproducible application of the framework's protocols, which are being specified in Paper II. The current prompt templates predate the framework and will be rewritten to follow it.
MOSAIC provides a grammar for analysis, not a validated analytical procedure. The current code can produce LLM-assisted outputs, but the analyst remains accountable for interpreting them.
git clone https://github.com/Or1un/MOSAIC.git && cd MOSAIC
pip install -r requirements.txt
cp modules/config.yaml.example modules/config.yaml # optional: YouTube, Telegram keys
python3 mosaic.pyLocal LLM analysis (recommended):
pip install -r requirements_llm.txt
curl -fsSL https://ollama.com/install.sh | sh # Windows: https://ollama.com/download
ollama serve &
ollama pull mistral:7b-instructSources: GitHub, Stack Overflow · Reddit, Telegram, Mastodon, Bluesky · YouTube, Medium
Privacy-first, in practice
Privacy-first means local by default, for the analyst and for the person analyzed.
- MOSAIC collects public traces from the platforms you select and stores them locally, in
results/(git-ignored). - By default, LLM-assisted analysis runs locally through Ollama; prompts and analysis data are not sent to a remote LLM provider. Collection itself still queries the source platforms.
- A remote LLM is opt-in. If you choose one, the data you submit leaves your machine and is processed under that provider's terms.
- Retention is up to you: delete
results/when an analysis is over.
Privacy-first is a design default, not a guarantee: it holds only as long as the analyst keeps to it.
Who it is for
Self-analysis, sociodynamic research, and analyses with volunteer subjects. Knowing that an analysis exists is not the same as consenting to it, and consent alone does not settle proportionality, data minimization or reuse of results. For analyses involving volunteers, the conditions of consent and withdrawal will be specified with Paper II's call for participants.
Not for surveillance, stalking, harassment, profiling someone without their knowledge, or mass automated analysis.
Compliance. You are responsible for GDPR and local law. MOSAIC does not make a use compliant; this README describes the tool, not a legal guarantee.
- Theory: Paper I — Why (working paper)
- Article series: The Variance is the Signal, and what follows on Medium
- Coming next: Paper II — How (in preparation, not yet available), with a call for volunteer analysts and subjects
The goal is not to confirm the theory, but to see which parts survive contact with real data. Challenge the assumptions, the method, the ethics. Point out where it fails.
Issues and Discussions on GitHub · or1un@proton.me