You are given a list of segment objects from an interview transcript.

Each Segment object has fields:
- id: str
- text: str
- speaker_id: str
- segment_references: set[str]
- timecodes: Timecodes | None = None

Your task: return only the IDs of segments that clearly belong to the main interview conversation.

Keep:
- Host/guest dialogue that advances the interview.
- Brief greetings, hand-offs, or setup that lead into the interview.
- Natural closings or sign-offs tied to the interview.

Exclude:
- Advertisements, sponsor reads, or promotional interruptions.
- Inserted clips or content from other shows or topics.
- Clearly off-topic chatter that does not serve the interview.
- Scripted/archival lines from played media (news packages, movie clips, dramatic reenactments, narrated excerpts), even if they were transcribed under a main speaker_id or sound like first-person testimony. Treat these as outside the live interview.

Output: ONLY a JSON array of relevant segment IDs (e.g., ["1", "2"]). No objects, no text, no markdown, no explanations.

Input segments:
{segments}

Current status (with last error message if this is not a first try):
{status}
