Privacy & Data Handling Policy

How this system collects, stores, and uses content for hate-speech monitoring in Gilgit-Baltistan.

What data is processed

This system processes three categories of data:

Submitted text — sentences or posts manually entered by an analyst for classification.
Ingested public content — the text of posts on monitored public pages and of the public comments left under them, the public display name of each page and commenter, and links to the original posts, collected through a licensed data provider. Location is not collected: each monitored page is assigned a district by an officer, and content from it is counted there.
Review decisions — the reviewing officer's name, decision, and notes recorded against a case file.

No passwords, payment details, or government identification numbers are collected or stored by this system.

Why it's processed

Content is classified and, when flagged, mapped to relevant provisions of Pakistani law (Pakistan Penal Code, PECA 2016) and the Constitution, to support timely and consistent identification of hate speech and coordinated harassment campaigns in Gilgit-Baltistan. See the Legal Framework page for the specific provisions referenced.

Human review, always

No content is ever acted on automatically. Every item that clears the classification threshold is written to a case file with requires_human_review set unconditionally — this is enforced structurally, not just as a policy, in the pipeline's storage layer. A human reviewer must confirm, dismiss, or escalate every case in the Review Queue before any downstream action is taken outside this system.

Retention & storage

Analysis logs, case files, and review decisions are retained in the system's internal database for as long as needed for review, audit, and reporting purposes. This dashboard is an internal analyst tool — it is not public-facing, and data within it is not shared with third parties outside the reviewing agency.

Known limitations

In the interest of transparency, this system does not currently:

Reliably determine the district a piece of content originated from.
Use a trained sarcasm-detection model — sarcasm scoring is a rule-based heuristic and should be treated as a weak signal, not a verdict.
Guarantee a legal citation for every jurisdiction — where no verified provision has been mapped (for example, some GB-specific local statutes), the system says so explicitly rather than guessing.

Questions or concerns

For questions about this policy or a specific case, contact the system administrator for your agency's deployment of this dashboard.

Last updated: 2026-08-25