A report with a hundred accounts is useful but slow to read. So after a scan Oracul asks the Claude model to condense the findings into a few paragraphs: who the person is, what they do, what they write about, when they were active. The summary is prepared in two languages at once, Russian and English.
What the summary is written from
Only accepted accounts, the ones that passed the namesake and face filters. If the engine rejected an account, the model never sees it, so the summary does not blend a person with a namesake.
What goes to the model
- Names and location.
- Public bios and links.
- Interest zones and publication topics.
- Registration and activity dates.
What never goes
- Emails.
- Phones.
- Breaches.
These stay inside Oracul and are visible only in the report itself. The model does not need them for a portrait, and the risk of passing them on is real.
The summary does not decide
The summary is a retelling, not an assessment. It assigns no risk score, gives no "approve / decline" verdict and adds no facts beyond the data. Decisions rest on signals and your rules. Every statement in the summary has its source further down the report: an account, a publication or a date.