research-companies

Follow the rules in `CLAUDE.md` and the spec in `PROTOCOL.md`. 1. Confirm scope: all active companies, or the subset the user names. 2. For each company, run every active signal through its declared provider (the `detection` block in its `signal.yaml`). No signal without a provider; ask the user to complete the definition instead of improvising sources. 3. Append each occurrence to `research/raw-signals.jsonl` (occurrence ID `signal-event.{company}.{signal}.{YYYY-MM-DD}.{source-key}`, source, status `unvalidated`). The ID is the dedupe key — never append one that already exists. 4. Cut the noise in `research/distillation.md`: which detections were junk and why, which are real, and how the real ones rank. Detection returns coincidences, stale news, and same-name companies — never pass raw hits through unfiltered. 5. Write the survivors into `research/signals.md` in priority order: what fired, the evidence, a suggested angle. 6. Report back: companies checked, occurrences found, strongest signals first.