Turn the attached interview into a short, source-grounded article and three social excerpts. Complete the work without asking about routine style choices. Use only the interview for event claims.

First read every turn. Keep each turn's ID. Write the article yourself; do not substitute a script that fills a prose template. Aim for 200–300 words and a clear factual headline. Each social excerpt must stand on its own and preserve the key qualification behind its claim.

If the input is synthetic, visibly label the article and each social excerpt fictional. Never present the sample as real reporting. For a real interview, preserve its actual provenance and do not invent permission or independent verification.

Attach source-turn IDs to every article paragraph, headline and social excerpt. Record each direct quote's verbatim span and source turn. Record every numeric claim with its value, meaning, source turn and exact supporting span. Do not confuse items with households or people. Preserve missing measurements, uncertainty and unconfirmed plans. Do not invent quotes, savings, emissions reductions, dates, success percentages or follow-up results.

Deliver interview.json unchanged, outputs.json with structured provenance, article.md, social-excerpts.md, and a readable local HTML preview with source references. Use a local standard-library validator to check turn IDs, all quoted spans and numeric-claim coverage; retain its result. Review every paraphrase yourself because mechanical checks cannot establish that the prose follows from its source.

If validation fails, correct the unsupported output or source mapping, then rerun only the relevant check. Never alter the source to make a claim pass. If the interview does not support a claim, remove or qualify it. Preserve the best supported output when a detail is unavailable.

Do not publish, send messages, access external accounts, or call a paid API. State actual execution and cost boundaries: assistant-authored output in the current session; no separately paid API invoked; subscription or usage allocation unmeasured. Do not call this an independent model benchmark or promise a free end-to-end workflow.
