Privacy and public evidence

Small excerpts.
Visible provenance.

We analyse public technical discussion to understand how models behave in practice. We do not build user profiles, sell personal data, or use collected material to train a model.

What we retain

  • A short excerpt, normally no more than 500 characters
  • The public source URL, title, username, date and engagement count
  • Thread identifiers used to prevent one conversation being counted repeatedly
  • Model, use-case, sentiment and evidence-quality classifications

What we do not do

We do not collect private communities, private repositories, email addresses, profile histories or direct messages. We do not reproduce full posts or articles. Public usernames appear only to attribute the quoted source.

Automated classifications can be wrong. They guide evidence review; they are not assessments of a person.

Deletion and correction

If the source disappears, the excerpt disappears.

We periodically check mutable sources. When an original item is deleted, we remove its excerpt, title, username and URL, and exclude it from recommendations. We retain only an anonymous source identifier, deletion date and one-way content fingerprint so the audit trail can explain why a citation was withdrawn without reconstructing the content.

For a correction or removal request, use the project contact listed in the repository. Include the public source URL so we can locate the record.

See methodology for source weighting and recommendation safeguards.