The public square does not only break when lies enter it.

It breaks when the conversation begins to misread itself.

Ithuriel Labs studies coordinated amplification.

Patterns, not people. Movement, not attribution. Method before findings.

What this is

Ithuriel Labs is an independent, non-commercial research project studying how narratives move across public platform conversations.

The project does not judge the truth of claims or identify the people behind accounts. It maps movement: similar language, shared links, timing, network structure, and public metadata used as screening signals.

This first statement describes the method and its limits. Findings come later, if baselines and measured precision support them.

How the pipeline works

  1. 01 / Collect

    Archived profile: Rising feeds from selected news/politics subreddits, top-level comments, 72-hour windows.

  2. 02 / Pair

    Semantic similarity and shared-link matching identify candidate narrative movement.

  3. 03 / Map

    Candidate pairs become a directed graph, clustered with Leiden/CPM, with component-level structure preserved.

  4. 04 / Screen

    Public metadata indicators are computed against configured thresholds and used only in measured combination.

The archived profile uses a fixed sampling frame so each run can be compared against later baselines. Semantic matching uses pretrained, inference-only models. No Reddit-derived content is used for model training without explicit platform consent. Metadata indicators remain screening signals only and are used in measured combination, never alone.

Publication discipline

Internal analysis uses permanent pseudonymous IDs, and a tested boundary prevents rendered artifacts from carrying real handles.

Every signal is a screening indicator, never a classification. No single indicator is cited alone. Combinations only, with their supports stated. A human reviews and approves every artifact before publication.

Calibration note

The archived research windows contain metadata signals such as account age, karma velocity, and username pattern. These are not findings. They are stress tests of the method.

One lesson from the archive is central: solo signals do not survive contact with reality. A signal can appear frequently in crisis conditions without meaning that it identifies coordination, inauthenticity, or wrongdoing. Prevalence is not false-positive rate.

For that reason, any future finding must be supported by measured combinations, control-window baselines, and labeled evaluation. Until those exist, indicator rates remain calibration material, not public evidence.

What we do not claim

We currently claim no findings about:

A findings brief will exist only when live collection is running under platform-approved access, control-window baselines exist, and every cited indicator combination carries measured precision from a labeled evaluation set. Until then, anything we publish is method.

Posture

Ithuriel Labs is independent, non-commercial research. Collection is conducted within platform rate limits and platform terms. Nothing here is a product pitch.

contact@ithuriel.io

@IthurielLabs