AI use and corrections

Where AI is used

• Classifying records against the fixed data center topic vocabulary.
• Planning Policy Desk retrieval and synthesizing answers from retrieved evidence.
• Drafting daily briefing language from tracked changes and verified upcoming items.

Where deterministic rules are used

• Policy mechanism signals, legal state buckets, identifiers, source labels, and change logging.
• Source retrieval, citations, record relationships, counts, exports, and evidence receipts.
• Safety boundaries for predictions, private information, legal conclusions, and unsupported local claims.

What AI output does not prove

Classification does not establish legal effect. A retrieved association does not establish support, influence, causation, or policy similarity. A summary does not replace the original document.

Verification behavior

Consequential factual claims should resolve to an inspectable source. Primary government text, tracker metadata, contextual reporting, voluntary guidance, and model inference are labeled separately. When the evidence is weak, the Policy Desk should narrow or abstain.

Human review

Briefings have a dedicated review queue and are not represented as substantively reviewed until that pass is logged. The selected oversight letters are manually reviewed. Most tracker classifications are automated and remain open to correction.

Questions and retention

Policy Desk questions and operational answer metadata are stored to support retrieval evaluation, feedback, cost control, and short lived answer caching. A random first party browser identifier measures unique, active, returning, and engaged use; it is converted to a one way pseudonymous key before storage. The system does not retain IP addresses, browser fingerprints, or user agent profiles for analytics, and pseudonymous activity rolls off after 400 days. Feedback text and any name, organization, or LinkedIn URL a person voluntarily provides are stored for owner review; optional identity fields are removed after one year. This is a public source research product, not a confidential workspace. Do not submit private project, client, or sensitive personal information.

Product usage measurement

The private staff dashboard measures daily anonymous browsers, completed product actions, and return usage. It stores country-level approximate location when available and a broad first referral category, without retaining IP addresses, precise location, raw referral URLs, or analytics copies of research questions. Cleared cookies and different devices may count separately. Staff use and marked automated tests are excluded. The public product has no customer sign-in, so anonymous browsers are not verified people. Personal MCP use is excluded by default. An external MCP installation can explicitly opt in using a separate revocable token; only tool name and success or failure are reported. MCP installations are counted separately from website browsers, and their location is not inferred.

Report a problem

Use the answer level helpful/not helpful control for structured feedback. For a source, classification, or product correction, open a public issue with the affected URL and the evidence supporting the change.

Published corrections

No material public corrections are currently listed. Future corrections should name the affected record or output, the original error, the corrected information, evidence, and correction date.