Topical deep dives

A separate episode of about ten minutes (deepdive.minutes 10, up to max_minutes 13 when an important story needs the context), also published as a readable article on its own page (site: <day>/deep-dive/, and the front page's Deep dives section), after the daily briefing, on the day's focal point: a breakthrough and its controversy, a concept suddenly everywhere (the "third scaling law": scaling compute at inference time), a change in the trajectory of the field, or a story that has matured over several days.

• Who decides: Claude picks the topic from today's briefing, the past week and recent deep dives, and may say there is none today. One a day (deepdive.max_per_week 7, since 2026-10-19), no repeat within 21 days unless something material is new. ./np deepdive --topic "..." sets the topic yourself.
• How it is written: an explainer for the ear, in six parts: why today → the idea plainly → what is established vs claimed → how the AGI-forward / AI-realist / market / policy readings differ (only where the facts allow) → our read and one diligence question → what would settle it. Claude may search the web for the paper, the lab's post and the strongest critique.
• Checks before it airs: every source link is opened (dead ones dropped; fewer than 2 live → no episode); quotation marks survive only around words found in the archived or fetched text; sentences with a $ or % figure found in no source are cut; private names are removed; markup or banned words fail the episode. The summary and "Why today" lines are then checked against the finished body (pipeline/consistency.py, since v0.1.9): a sentence whose figure, company or "not disclosed" claim the body doesn't support is dropped. Since v0.2.33 the script's factual sentences are also checked against the dive's sources, turn by turn, before it is recorded, and rewritten or left for Steven in the draft Doc (./np audit deepdive DAY is a dry run). See docs/FACT-CHECKING.md (4e).
• No story twice (v0.2.5): when the deep dive covers one of the day's Must reads, the briefing keeps that story to a sentence and a pointer to the deep dive, in the written edition, the email and the podcast.
• In the podcast (2026-10-19): the deep dive follows the briefing in the same episode. The briefing host ("I'm Michael" in the first line) closes by promoting today's deep dive and where to find it (the web page, with the written edition and every past episode), hands off to Bella, and the deep dive's sign-off closes the episode. The feed carries one file per day (briefings/<day>.podcast.mp3, "… · with deep dive: <title>"); the web page keeps separate players for the briefing and the deep dive. If the briefing audio was made without the hand-off (a deep dive made later by hand), the deep dive stays a separate bonus episode.
• Where it goes: briefings/deepdives/<day>.md (show notes + sources), .script.md, .mp3. The podcast feed lists it as a bonus episode "Deep dive: <title>" right after that day's briefing; the day's page links to it.
• The weekly flagship (v0.2.18): Saturday's deep dive is a two-voice episode of about 22 minutes on the week's topic most worth understanding (the first topic in Up next, drawing on the whole week's Must reads). Bella hosts; Michael is the sceptic, whose job is to push back on the evidence. Same evidence rules and checks as the daily deep dive; the article reads as a conversation under the usual headings. The evening note on Friday says so. deepdive.flagship in config.yaml (weekday, minutes, names, the sceptic's voice) or enabled: false. We review it after four weeks with the weekly critic.
• Its own podcast (v0.2.17): the deep dives also have a feed of their own, deepdive.xml ("AI Briefing: Deep Dive", with its own cover). The daily show keeps the deep dive after the briefing, through the host's hand-off. The site's Deep dives section, each deep-dive page and the email link to the deep-dive feed.
• When: planned during the morning run before the briefing audio (so the briefing hands off to it), written and voiced right after, then published and mentioned in the email.
• Picking from the evening note: your private note (sent by the daily run after the subscriber email) gives every topic its own buttons: tomorrow's (Add your take, Move down, Dismiss), the rest of Up next (Make next, ↑, ↓, Add your take, Dismiss), each new suggestion (Make next, Add to Up next, Add your take, Dismiss), and "Suggest as a deep dive" on today's deep dive and each watchlist mention. Each button opens the private queue page with that one change ready to confirm (an email link never changes anything by itself); the page has the same buttons and a take box on every topic (signed owner links, 8-day expiry).
• Your take on a deep dive (v0.2.11): type it on the queue page or on the dashboard's Deep dives page, any time before the run that writes it. It is read verbatim as "Steven's take" at the end of the deep dive (its own section on the site and in the podcast), after the same privacy check as morning takes; a take that fails it is held back and stays private. Takes are opinion: never rewritten by a model, outside the fact check.
• Queue: dashboard /deep-dives: Claude adds up to 4 suggestions each morning; queue, reorder, dismiss or add your own. Up next goes first; otherwise the day's best (strength 4+). One a day; Up next is never left empty (the best suggestion is moved up if you queued nothing).
• Voices: briefing am_michael ("Michael", ~131 words a minute), deep dive af_bella ("Bella", ~157), changeable on the Tuning page. Script length is set from each voice's measured speed (audio.voice_wpm).

Since v0.2.29: drafts, takes, suggestions and the thread (2026-10-28 to 10-30)

• The advance draft (v0.2.31): at 15:30 tomorrow's deep dive is written and checked, and saved to Drive as "<day> · Deep dive · <title> (draft)" with a nudge and an email. Two boxes at the end: Steven's take (read verbatim after the episode) and Notes for the editor (what to change). At 05:00 the take goes in and notes make the writer revise the draft, which is checked again; then it is voiced and shipped. Only the two boxes are read.
• Takes edited in the Doc (v0.2.25): after publication, a changed take in the deep-dive Doc re-records the deep dive and re-publishes its episode within 20 minutes; no email is re-sent.
• Readers suggest topics (v0.2.29): a link in the email and on every deep-dive page; suggestions enter the queue as "suggested by <name>" (or "a reader") and the episode credits them when it runs.
• In Slack (v0.2.28): the deep dive is a reply, with its own card, in the thread of the morning post.
• Rebuilding a shipped dive (v0.2.36–v0.2.38): the deepdive-rebuild queue job (auto tier from v0.2.37) runs ./np deepdive DAY --topic "..." --force on the same topic, keeps the take Steven already gave, records a correction and publishes. A topic already in the queue is reused, and --topic exits 1 when no dive was written, so a failed rebuild records no correction.