Coverage of HPSP

How tones are assigned

For each source, the title + the first ~1500 characters of raw content go to an LLM (cheap call) with a prompt asking it to pick exactly one tone from the canonical five:

• Favourable — positive framing, advocacy, sympathetic, PR-like
• Neutral — descriptive reporting without editorial lean
• Critical — adversarial framing, accountability journalism, investigative
• Mixed — substantial both-sides framing in a single piece
• Primary record — government filing, court docket, SEC submission, raw transcript — framing doesn't apply

The classifier runs on a boot-time backfill (small batches, high-priority entities first). New ingested sources also pass through it. The model name (and run date) is stored per source — a future upgrade can re-classify with a stronger model.

A high critical % means the coverage in this archive is adversarial. It does NOT mean the subject is guilty of anything. A high favourable % means the coverage leans positive — it does NOT mean the subject is benign. The distribution shows the editorial frame; verify against the linked primary sources before citing.

Machine-readable: JSON