Release notes

v0.2.50-beta · 2026-11-02 (weekday slot)

• Money section: up to five funding, deal and valuation stories (a deal word plus a stated amount) get their own "Money" block in the briefing, email and site, between Recommended reading and In brief.
• One chart a day: the write stage asks for a 3-6 bar chart from one Must-read story; every value must appear in that story's own sources or no chart is made. Shown under the story in the email and on the site, labelled "Figures as reported in this story's sources." (analysis.chart: false turns it off.)
• Dashboard phase 3: Briefings shows email time and count, takes status and corrections per day, with the rebuild action folded under Maintenance; Sources leads with failing sources and folds the queue and history; Runs shows the last 14 with a health line and folds older runs; Subscribers has one channel summary and folds manual sends and list management.
• Fixes found by the ship check: gift links carrying any *_token parameter (a Washington Post pwapi_token JWT reached the staging site and the privacy scan stopped the release) are cleaned like other tracking parameters; the chart accepts a story number returned as text ("2") and logs why no chart was made.

v0.2.49-beta · 2026-11-02 (weekday slot)

• Core beliefs, first refresh (Steven approved all six proposals, 2026-11-01; doc "Core Beliefs: Proposed Updates"): thesis 16 says containment is conceptual until proven; view 5 adds founders' liability; view 15 warns against writing off incumbents on early cannibalisation; new view 22, "Human provenance or ruin"; perspective.md adds robotics and AI-driven science, agent-to-agent authority, and a liability-and-insurance diligence rule.

v0.2.48-beta · 2026-11-02 (weekday slot)

Quality fixes from the 10-31 review: - Story fatigue: when today's lead is the same story as a lead in the last two editions and brings no new primary source or report, the next Must read leads instead (editorial.stale_lead; the S-1 led five days running). - A capability slot: when no Must read is about what AI can now do and a well-backed release or result sits in Worth knowing, it moves up (editorial.capability_slot). The summary prompt rates capability news 3+, 4 for a first or a new generation (DeepMind's Genie 4 had been rated 2 and never reached the edition). - Verdict labels that discriminate: Reported / Self-reported / Single source instead of a blanket "Unverified" (all five Must reads on 10-31). Machine phrasing in evidence lines is put into plain words. - The board-seat scrubber no longer removes "a disclosure." from the middle of a sentence (it left "not But" in the 11-01 deep dive); the deep-dive check lists any sentence that looks cut short. - A held-back edition, sent later, also posts to Slack (the 11-01 post had to be sent by hand).

Dashboard refresh (Steven, 11-01: "pertinent information front and centre"): - A new home screen: Needs you (approvals, protected-file changes, an unsent edition, held takes, fact-check items, new feedback; "Nothing needs you right now" when calm), Today (run, site, email, Slack, takes, deep dive and corrections as status pills, with the day's top lines), Tomorrow (the deep dive to be drafted and what follows) and Health (critic score and rank, the release, failing sources, a trend chart against the top three shows). - Five menus instead of twenty tabs: Today, Editions, Sources, Audience, System. The old status page is /status. - A release restarts the dashboard, so it always runs the released code (11-01: a stale process failed every page).

v0.2.47-beta · 2026-11-01 (fix lane)

• A re-recorded podcast keeps Steven's takes (2026-11-01): remaster rebuilt the briefing audio from the take-free script, so the re-records of 10-30, 10-31 and 11-01 dropped his takes from the episode. It now splices the published takes back in (the same step the morning run uses) and stops if it can't.

v0.2.46-beta · 2026-11-02 (weekday slot)

• Protected files change only through a release (2026-11-01: a house-views edit committed straight to production held the morning edition and its email until the gate was approved):
• the pre-commit hook refuses a commit on production's main branch that touches a security-sensitive file (code, tests, config.yaml, house_views.md, newsletters.yaml, ...); releases and rollbacks pass (NP_RELEASE=1);
• the run queue watches for protected files that differ from the last approval, including uncommitted edits, and says so at once: a notice, and an approval job so Steven gets the prompt on his phone right away instead of finding the edition held at 05:00 (no extra job when a release is already queued: it approves them).

v0.2.45-beta · 2026-11-02 (weekday slot)

• Steven's take leads (Steven, 2026-10-31: "yes on take"): the day's take sits under the title on the site, in the email header (Top of mind's take when there is no day take) and right after the podcast intro.
• Typo fixes in takes: obvious slips are fixed before publishing (single words, within two letters, never a rewording); every fix is listed, with the original, in the takes note to Steven so he can undo it.
• Running trend charts (Steven, 2026-10-31): the dashboard's Quality page shows every show's weekly critic score over time against ours (overall or any characteristic), and our daily edition metrics as line charts (pipeline/trends.py); the weekly critic email adds the field's scores by week.
• House views 18-21 (Steven's standing theses from his takes of 10-27 to 10-31) are in house_views.md; Our read builds on a thesis used in the last few days instead of restating it.

v0.2.44-beta · 2026-11-02 (weekday slot)

• Hub page in the brand (Steven, 2026-10-31): the project page on the public hub shows the author photo, the tagline and the teal accents (this page only; the hub's shared style is untouched).
• Deep-dive fact check checks facts only (2026-10-31: 12 items held on 10-31, most of them framing): sentences that frame a reading ("the AI-realist reading says..."), imagine an example ("Picture a startup...") or give our view are no longer checked; in "How people read it", only a sentence that attributes a view to someone is.
• The release soak covers the deep dive: ./np staging run --force --deepdive rewrites the day's deep dive on the same topic with the dev code, and staging compare reports it (status, words, sources, checked / held).
• A rebuilt Saturday deep dive keeps the two-voice flagship format (10-31's rebuild came out in one voice).

v0.2.43-beta · 2026-11-02 (weekday slot)

• Public pages never show a local address (2026-10-31): release notes and the hub page turn a URL on the Mac (the X tool's) into plain words, so the hub export is no longer refused.
• Technical deep dives (Steven, 2026-10-31): the daily suggestions now include at least one "technical" candidate (world models, robotics, scaling laws including test-time compute, surprising results...; deepdive.technical_themes), and a technical episode is written as an explainer for investors and business people.
• Watchlist skips newsletter sales emails (Steven, 2026-10-31): an Exponential View "24 hours to get 50% off" email surfaced as a watchlist company mention. Titles that read as a discount or subscribe deadline are no longer matched.
• Credentials never pass through links (2026-10-31): a Stratechery feed's per-reader access token reached the 10-31 page and email. Token, key and signature parameters are now stripped from every link, and the publish privacy scan refuses any page carrying a JWT or a token parameter.
• X reposts skipped (2026-10-31): xsukax lists reposts in an account's feed; a post whose link is another account's is no longer credited to the person who reposted it.
• Links from more Slack channels (Steven, 2026-10-31): slack.channels (first: #discuss-generative-ai), read by the AI Briefing bot. Linked content only; the message is a private relevance cue (mentions and links stripped, never quoted or attributed); posters never stored; each pull logged with its time in data/state/slack-pulls.jsonl.
• Subscribers page (was Email): one row per person with every channel they get the briefing through (email, Slack #cto-ai-newsletter; podcast listeners are anonymous). pipeline/audience.py.
• One brand template (Steven, 2026-10-31; docs/BRAND.md): brand/brand.yaml (teal palette with a dark mode, Inter, the one author image, the waveform, the taglines) and pipeline/brand.py. The website moves from navy and mint to the teal brand: pale reading surface (deep teal in dark mode), teal masthead with the author image, wordmark, tagline and waveform on every page, deep-teal masthead on deep-dive pages, the day's Slack card as each page's link preview, "Today in 30 seconds" on the home page. The email: pale surface, white cards, teal buttons, Inter. Slack cards, the channel banner and the covers read their colours, taglines and author image from the brand (./np brand build [--covers]); the author image is the glasses photo everywhere (also the site's og:image and bio). Feeds: author "Steven Woods"; the RSS feed has the cover and the main tagline. Dashboard: brand colours and header, and a Brand tab showing the template and where it appears.

v0.2.42-beta · 2026-10-31

• Monthly promo (Steven, 2026-10-31, option A): on the first Monday of each month, once the edition is out, one short post in #cto-musings pointing to #cto-ai-newsletter (a channel link: one click opens it with Join), the website and email sign-up, with the channel banner. ./np slack-post promo [--preview|--now], queue recipe slack-promo.
• brand/: the author image master (author-master.png, the glasses photo cut from the podcast cover) and the Slack channel banner (slack-channel-banner.png), the first files of the brand template (see the brand plan).

v0.2.41-beta · 2026-10-31

• Fix: the bot couldn't find #cto-ai-newsletter by name (public, bot invited); slack.post.channel_id is now used directly when set (C0CDBLD1JT1), so the morning post and Earlier editions don't depend on the lookup.

v0.2.40-beta · 2026-10-31

• ./np slack-post archive FROM TO [--preview] (queue recipe slack-archive, chat): one "Earlier editions" message in the post channel, a line per day with its first lead and briefing and deep-dive links, posted once. For the days posted to #cto-musings before the move; Steven deletes the old bot posts there himself.

v0.2.39-beta · 2026-10-31 (weekend exception, Steven's yes in chat)

• The morning Slack post (card + deep-dive thread) goes to #cto-ai-newsletter instead of #cto-musings (Steven, 2026-10-31). Links are still read from #cto-musings.
• Fix: a page found by the fact check's search joins the Sources (or is linked) only when it carries words of the fixed sentence itself (a 5-word run or its quotation); on 10-31 an unrelated TechCrunch story was added to the rebuilt deep dive's Sources because the fix was supported by the dive's own sources (removed by hand).
• Docs: ARCHITECTURE, FACT-CHECKING and DEEP-DIVES brought up to v0.2.38 (pre-send and deep-dive fact checks, v0.2.34 search and second-judge rules, release lanes and soak, chat approval for ship/release, auto-tier remaster and deepdive-rebuild, X via list members and a local RSS generator); run-queue, release-flow and fact-checking diagrams updated.

v0.2.38-beta · 2026-10-31 (fix lane)

• Fix: deepdive --topic on a topic already in the queue (the dive being rebuilt) found no item and wrote nothing, and the rebuild recipe then recorded a correction anyway. The queued item is reused, and deepdive --topic exits 1 when no dive was written, so a rebuild stops before recording a correction.
• Fix: the subcommand-argument tests read config.yaml, which the public export doesn't have (public tests failed).

v0.2.37-beta · 2026-10-31 (fix lane)

• Fix: ./np deepdive DAY --topic "..." --force stopped at the argument parser (--topic unknown, --force taken as the top-level flag), so the deepdive-rebuild recipe failed; ./np staging run --force lost its --force the same way, so the ship's soak refused before 08:30. Everything after a subcommand that takes arguments now goes to it.
• Approvals up front (Steven, 2026-10-31): a ship/release carrying Steven's quoted yes from chat runs on its own after tests and the soak pass, with no phone prompt; rollback, code push and secrets keep the tap. remaster and deepdive-rebuild no longer need a chat quote (auto tier). See docs/security/RELEASE.md.

v0.2.36-beta · 2026-10-31 (fix lane)

• Fix: the deep-dive fact check (v0.2.33) keyed a two-voice script by section, and a flagship repeats section ids on every turn, so every turn of a section came back as that section's last turn. The 10-31 flagship repeated whole paragraphs. The check now works turn by turn; a test covers a two-voice script.
• deepdive-rebuild queue recipe (chat): rewrite a shipped day's deep dive on its topic, record the correction and publish. A rebuild keeps the take Steven already gave the dive.
• Fix: the ship's soak (staging run) refused before 08:30 even after the edition was out; it now passes --force, since a ship only runs once the edition is out and no run is busy.

v0.2.35-beta · 2026-10-30 (Steven asked to bring X forward for the weekend runs, an exception to the slot rule)

• ./np x probe [N] (queue recipe x-probe, auto): the first N accounts through the configured X instances, with post counts; nothing is archived.
• X through the local xsukax-X-RSS-Generator on Steven's Mac (Steven, 2026-10-30; his test returned 5 posts): an x.yaml instance may be a template with {handle} (a local address on the Mac);lists:adds the members of the named X lists (data/state/x-lists.json) as tier "list" accounts, after his own;delay_secondsspaces the requests;max_accounts` caps a run (120). Nitter mirrors stay as fallbacks.
• Fix: a remaster recorded as a correction published before the correction note existed, so the page lacked it until the next publish; it now publishes again after recording the note.

v0.2.34-beta · 2026-10-30 (fix lane)

• Fact-check fixes from the first dry run on the 10-30 deep dive: a search result that is a homepage or section front (anthropic.com, claude.com) is never a found source, and a found page must also carry a second name or a figure from the claim; the second judge checks one claim per call (the local model answered a batch with only its first verdict) and falls back to the first model; a rewrite is kept only when the second check confirms it, otherwise it is undone and goes to Steven ("rewrite not confirmed"); rewrites narrow the original claim, never swap in a different one, and may return nothing (then Steven decides).

v0.2.33-beta · 2026-10-30

• The fact check runs before the edition is sent (Steven, 2026-10-30, after "CDIC" was corrected the same morning): every factual sentence of the top section, What happened and Worth knowing checked against its sources; partly-right or contradicted lines rewritten (links kept) and re-checked; anything unsettled listed in the takes email. Acronyms no source contains are checked first. The day page ends with a Notes footnote: the fact check's time and every correction with date and time (ET). New queue recipe remaster (chat): re-record a day's podcast from its corrected text, recorded as a correction.
• Fact-check gaps closed (Steven, 2026-10-30): ranked excerpts with twice the room and a 12,000-character second look before "not found"; a search for a live page (government and regulators first) when the cited sources lack a fact, which the rewrite may cite with one new link that also joins the Sources line; a second, different model (llm.stages.judge2, local by default) re-checks rewrites and runs the after-send audit.
• The deep dive gets the same sentence check before it is recorded (factual sections only; spoken-sentence rewrites; a found page joins its Sources; second-model re-check). Open items go to the draft Doc ("Fact check: please look") and the takes email; the deep-dive page has the Notes footnote.
• ./np audit deepdive DAY (queue recipe audit-deepdive, auto): the deep-dive check on a recorded dive as a dry run; proposed fixes go in only as corrections.
• Release rules for day-to-day stability (Steven, 2026-10-30): ship has lanes. The default slot lane releases once per weekday from 13:00 (release.slot in config.yaml), after the edition and before the 15:30 draft; lane: fix is for broken things, any time in the release window; no slot releases at weekends. Every ship soaks first: staging run rebuilds today's edition with the dev code and staging compare checks it against production; a compare that needs review stops the ship before Approve unless accept quotes Steven's OK. The Approve shows the lane and the soak verdict. Schedule tab gains the release-slot row.
• Docs brought up to date with v0.2.10–v0.2.32: README (Steven's day, the run, new commands), OPERATIONS (the advance, runner, ntfy, Slack and Jev services), RELEASE (the three approval paths), DEEP-DIVES (drafts, takes, suggestions, Slack). Fix: ./np schedule printed the service status (an old alias); it now shows the timetable (show), and ./np timetable is the same.

v0.2.32-beta · 2026-10-30

• Jev shadow filter (Steven, 10-29/30): TypeSafe's Jev (a structured-decision model, sub-second, ~$0.04 per million input tokens) is asked keep/drop for every item the local filter judged; both answers are logged, the morning report shows the agreement, and nothing changes in the edition. Daily token budget (400k, about two cents). Public fetched items only; never in staging. One-time: the jev-setup job (a Mac dialog stores the key in the Keychain and makes one test call). ./np jev report [DAYS]; run-queue job jev-report.

v0.2.31-beta · 2026-10-29

• The advance deep dive (Steven, 10-29): at 15:30 (schedule.advance) tomorrow's deep dive is written and checked and saved to Drive as "<day> · Deep dive · <title> (draft)", with a nudge and an email. The Doc ends with two sections: "Steven's take" (read verbatim after the episode) and "Notes for the editor" (what to change). The 05:00 run reads them back: the take goes in; notes make the writer revise the draft, which is checked again; then it is voiced and shipped. Install the agent once with the service-install job (kind: advance). ./np advance drafts now; run-queue job advance. The Schedule page has the row.

v0.2.30-beta · 2026-10-29

• Schedule page (Steven, 10-29): the dashboard's new Schedule tab (and ./np schedule) lists what is produced through the day, when, where it goes, and the window for feedback on each piece and how it is picked up.
• The review page keeps your takes (Steven, 10-29: a page of typed takes was lost by tapping a source link). Every link on the page opens in a new tab; each box is kept in the browser as you type and sent to the mail script a few seconds after you stop (a later save for the same box replaces the earlier one); leaving the page with unsent text asks first. Submit and Done work as before.
• Pronunciations (Steven, 10-29): the table gains outlets and acronyms said by name (NYT is "New York Times", WSJ, FT, GAAP as "gap", METR as "meter", S-1 as "S one", PBC and LTBT spelled out) and more people and companies (Qwen is now "Kwen", Nvidia, Pichai, Nadella, Dwarkesh, Azeem Azhar, Stratechery, Zhipu, Mistral). Uncertain ones carry a comment. ./np pronounce lint [DAYS] lists what the voice may have garbled lately, for the next additions (docs/OPERATIONS.md).
• Fix: takes edited in the Drive Doc before the edition ships were not read, because the Doc is still the draft then ("held for your takes"); the draft is read when the published one isn't there yet.

v0.2.29-beta · 2026-10-28

• Readers suggest deep dives (Steven, 10-28): a "Suggest a deep dive" link in the email's deep-dive section (with the reader's own token) and on every deep-dive page on the site. The page asks for a topic, why now, and an optional name. Suggestions land in the queue as "suggested by <name>" (else "a reader") and show that way in the evening note, on the queue page and in the dashboard; when one runs, the deep dive's notes say who suggested it. Addresses stay in the private queue file. Capped per reader and per day. Needs the mail-script deploy (the ship does it).

v0.2.28-beta · 2026-10-28

• The deep dive in the thread (Steven, 10-28): right after the morning post, a reply in its thread with the deep dive: title, summary, Steven's take on it and the links, with its own card in the deep-dive cover's colours. ./np slack-post deepdive DAY; run-queue job slack-deepdive (chat). Card text now ends at a full sentence or an ellipsis, never mid-sentence.

v0.2.27-beta · 2026-10-28

• The morning Slack post (Steven, 10-28): when the email goes out, one post in cto-musings from the bot user "AI Briefing": the day's three leads, one of Steven's published takes (his words, the one with the most bite that fits), the deep dive and the links, with the day's card attached: a 1600×900 image in the newsletter's teal with the masthead, photo, leads, the take and the deep-dive band (pipeline/card.py, Pillow + Inter, new in requirements). Set up once with the slack-setup job (a Mac dialog takes the bot token; scopes chat:write, files:write, channels:read; invite the bot to the channel). ./np slack-post preview DAY shows the text and renders the card; run-queue jobs slack-card (auto) and slack-post (chat). Security: a new outbound step, gated like the others; never in staging; bot posts are ignored by the Slack reader.
• Fix: the Drive check re-queued a takes-sync every 20 minutes after a plain re-publish (the Docs weren't rewritten, so nothing recorded our own last write). A take the commentary already has is never an edit, and takes-sync now rewrites the Docs.
• Podcast artwork refresh: podcast apps cache artwork by URL and never refetch it (Apple still showed the old cover a day after the new one went up), so the feeds now name a copy whose filename carries the image's hash. Apple Podcasts picks up the new covers at its next feed check.

v0.2.26-beta · 2026-10-28

• Fix: the runner's minute tick exited before the Drive check when the queue was empty, so edited takes were never noticed on their own. It now runs the check every 20 minutes regardless.

v0.2.25-beta · 2026-10-28

• Takes edited in the Docs count (Steven, 10-28: a take written in the morning boxes, then augmented in the car with Claude editing the shared Doc). The pipeline reads the published edition's "Steven's takes" and the deep dive's "Steven's take" back from Drive; lines that changed after its own last write are new takes, checked like any other. The runner looks every 20 minutes once the edition is out and queues a takes-sync job (auto tier) that applies them and re-publishes the site and email boxes; an edited deep-dive take re-records the deep dive and re-publishes its episode. No email is re-sent. ./np takes apply DAY does the same by hand.

v0.2.24-beta · 2026-10-28

• Phone notifications and one-tap approvals (Steven, 10-28: "more control from my iPhone"): through the free, open-source ntfy app. Approvals arrive as a notification with Approve and Reject buttons; one tap, no passphrase, no browser. The tap sends a one-time token the Mac checks, over private topics only the Mac's Keychain and the phone know. Nudges: the takes page is ready (tap opens it), the evening deep-dive note is out (tap opens the queue page), a run needs attention, a run-queue job finished. Turn on with the ntfy-setup job (a Mac Approve; the topic lands on the clipboard for the phone), then "ntfy test". The passphrase path stays as a fallback. Security-sensitive: a third approval path.

v0.2.23-beta · 2026-10-28

• Phone approvals: the job file now says why a phone answer wasn't accepted (no answer yet, the check failed, or the passphrase didn't match), and a passphrase that doesn't match gets one Mac notification, as promised.

v0.2.22-beta · 2026-10-28

• Calls we made (site page): every house-view call with what would change our mind and its check date. On Sundays the run scores the calls that are due against what we've reported since (right, wrong, mixed; too early gets 90 more days). ./np positions [list|score]; run-queue job positions-score.
• Accuracy audit: after the email, about five of the day's claims are re-checked against their own sources (supported, partly, not found, contradicted), and the consistency flags are counted. Partly right or contradicted claims send one alert email; the site's Accuracy page shows the daily counts and every correction. ./np audit [DAY]; run-queue job audit. Dashboard switch: Briefing structure → "Accuracy audit after the email" (on).
• Phone approvals (Steven, 10-27): mac-tier jobs (ship, release, approve...) can be approved from the phone. The runner emails a one-time link; on the page you type your approval passphrase and tap Approve or Reject. The passphrase never leaves the phone (the page sends a proof the Mac checks against its Keychain), so a link or a record alone can't approve. Turn it on once with the phone-setup job (a Mac dialog asks for the passphrase). Security-sensitive: a second approval path. Needs the mail-script deploy (the ship does it).
• Merged top, on (Steven, 10-27): Briefing structure → "Top of the edition" = merged gives one 5-bullet "Today in 30 seconds" with the diligence question under it, instead of the day in brief plus Top of mind (critic #7A). Audio, email, takes, Drive and the checks read both shapes.

v0.2.21-beta · 2026-10-27

• House views. Our read takes a position under one of the views in house_views.md (1–10 plus Steven's theses 11–17) and names it ("House view: …"). The lead story adds the case for, the case against, the deciding question and our lean, in place of its other view. Each position is logged in data/positions/ledger.jsonl with what would change our mind and a check date (a thesis's own, else 90 days). Dashboard switch: Editorial lens → "Take positions under the house views". Security: house_views.md steers the models, so it joins the files whose changes need release approval.

v0.2.20-beta · 2026-10-27

• Pending sign-ups (Steven, 10-27): one reminder email to people who signed up but never confirmed ("One click to start receiving it"), at most once a week each; 14 days after the reminder a still-pending sign-up is marked expired by the morning run (they can sign up again, and an old confirm link still works). Run-queue jobs email-remind and email-remove (chat tier). ./np email list counts readers without the owner's test copy. Needs a mail-script deploy (the ship does it).

v0.2.19-beta · 2026-10-27

• New taglines (Steven, 10-27): the AI Briefing is "Slowly, then fast. Every morning." and the deep dive is "One story, slowly. Down to the details." On the email masthead and deep-dive section, both podcast covers, both feed descriptions and the site.
• Speech: a name glued to its version number is split before the pronunciation table ("Qwen3.8" was heard as "QN3, 8"; now "Chwen three point eight").

v0.2.18-beta · 2026-10-27

• The weekly flagship deep dive (Steven, 10-27: "lets do this as outlined"): on Saturdays the deep dive is a two-voice episode of about 22 minutes on the week's topic, with Bella as host and Michael as the sceptic who pushes back on the evidence. Same checks as the daily deep dive. Shown as "Weekly deep dive" in the feeds and "This week's deep dive · two voices" in the email; Friday's evening note says so.
• Fix: the new audio check is ./np audio-check (and run-queue audio-check); ./np listen already played the day's MP3 on the Mac.

v0.2.17-beta · 2026-10-27

• Hear what was published (Steven, 10-27): after the email, the day's episode is downloaded from the public feed, transcribed locally (Whisper, free) and compared with its script. A missing or garbled section (like 10-27's intro) sends one action-needed email. The as-published transcript is published with the next publish (WebVTT, podcast:transcript in the feed, a link on the day's page). ./np audio-check [DAY]; run queue audio-check; setup (chat tier) installs the new requirement (mlx-whisper, Apple silicon).
• The deep dives as their own podcast: deepdive.xml, "AI Briefing: Deep Dive", with its own cover. The daily show still flows from the briefing into the deep dive. The deep-dive feed is promoted on the site (Deep dives section, each deep-dive page, the footer) and in the email ("▶ Listen to the deep dive", "subscribe to the deep dives").

v0.2.16-beta · 2026-10-27

• Podcast brand matches the newsletter (Steven, 10-27): new 3000×3000 cover in the teal of his photo, with his photo, "AI Briefing", the tagline and "Curated by Steven Woods". Podcast apps pick up new artwork on their own schedule (Apple can take a few days).
• The show is named in the first seconds (Steven, 10-27: "the audio today has no introduction"): the cold open now starts "On today's AI Briefing." Each recording also gets a .sections.json with every section's length, so a section that comes out silent or clipped is visible.

v0.2.15-beta · 2026-10-27

• Fix (mail script): rewriting a Drive doc that ends with a list (Steven's takes) failed in clear(); an empty paragraph is added first (10-27). Needs a mail-script deploy (the next ship does it).

v0.2.14-beta · 2026-10-27

• Fix: the public export's tests failed because the photo isn't exported (the public code push stopped).
• Fix: ./np release … --force never reached the release (the command line took --force for another command).
• Fix: the release itself still waited for 08:30 when the run queue let it go early; it now follows the same rule (once today's edition is out).
• Shares: a link shared in Slack is labelled "Shared internally today" (it said "Shared in Steven's team channel"); Steven's own Bluesky posts say "Posted by Steven on Bluesky" (Steven, 10-27).

v0.2.13-beta · 2026-10-27

• Fix (mail script): rewriting an existing Drive doc (the draft becoming the published edition) failed with "Can't remove the last paragraph in a document section" (10-27). The tidy-up step is now optional. Needs a mail-script deploy.
• Voice: scripts/voice/transcribe.py transcribes Steven's recorded talks locally (Whisper), from the public record manifest.
• Approvals keep moving in the morning (Steven, 10-27: "get everything moving as soon as I am done"):
• Approve early: a release queued before the morning is done asks right away, and runs as soon as the edition is out. The click is kept as a signature made with a key in the Mac's Keychain, which a file alone can't fake; it's used once and only that day.
• The morning window ends when today's edition has gone out and no run is busy, not at a fixed 08:30.
• One Approve covers the release's follow-through: the public code export, and a mail-script deploy when the live script differs from the release.
• The approval dialog stays up 19 minutes and comes back right away, with a sound, until it's answered.

v0.2.12-beta · 2026-10-26

• Email masthead (Steven, 10-26): his photo and colour (the teal of the photo), "AI Briefing", the tagline "What happened, what holds up, what it means." (email.tagline overrides) and "Curated by Steven Woods · for early-stage investors". The photo is served from the site (assets/steven-woods.jpg). Visual only.
• Drive docs: the deep dive no longer repeats "Why today", and flag emoji (garbled in Docs) are left out.
• Fix: the run queue's approve job stopped at its own gate check when only the approval was missing (10-26).

v0.2.11-beta · 2026-10-26

• Your take on a deep dive (Steven, 10-26): a box on every topic on the queue page (from the evening email) and on the dashboard's Deep dives page. The take is read verbatim as "Steven's take" at the end of that deep dive, on the site and in the podcast, after the same privacy check as morning takes.
• Every topic in the evening email has its own buttons: Make next, ↑ / ↓, Add to Up next, Add your take, Dismiss, and "Suggest as a deep dive" on today's deep dive and watchlist mentions. Each opens the queue page with that change ready to confirm. The queue page can now also add a suggestion to the end of Up next. Needs a mail-script deploy.
• The edition and deep dive in Drive (Steven, 10-26): Google Docs in the shared Briefings folder, named by date. With the pre-release takes email: "<day> · AI Briefing (draft, held for your takes) · <lead>" and "<day> · Deep dive · <title>" (the takes email links to the draft). Once published, the draft becomes "<day> · AI Briefing · <lead>" with Steven's takes, and the deep dive doc is rewritten. ./np drive save [DAY] / run-queue drive-save. After the mail-script deploy, run authorizeDrive once in the Apps Script editor to grant Drive and Docs access.

v0.2.10-beta · 2026-10-26

• Fix: the takes email linked to a page that doesn't exist yet (Steven, 10-26). The day's site page is only published after the cutoff, so the link now opens a private review page served by the mail script: the held edition with a box after each section and one Submit ("Done: send the edition now" ticks it out early). The pipeline uploads the page just before the takes email (setreview, key only); the script shows it only with Steven's signed takes token. If the upload fails, the email goes without the link (the reply buttons still work).
• The text version of the takes email numbers each story by its headline again.
• Email takes: a section name as a header ("Worth knowing:", "Top of mind:", "Must read 2:") now counts like "W:", and a signature (a bare link or "-- " line) ends a take (10-26: a Worth-knowing take was missed).
• Includes the v0.2.9 mail-script fix below (that release was not shipped on its own). Needs a mail-script deploy: deploy-mailer, Approve on the Mac.

v0.2.9-beta · 2026-10-25 (folded into v0.2.10)

• Fix (mail script): site takes were saved, but Sheets turned the day ("2026-10-25") into a date cell and the pipeline's read-back compared it as text, so it found none. Takes are now written as text, and the read-back compares raw cells, so the 10-25 test takes are found too. Needs a mail-script deploy (the first via clasp).

v0.2.8-beta · 2026-10-25

• Takes on the site, fixed for an open page: the takes email's site link adds ?v=<day> so the browser loads the page fresh, and the boxes also appear when only the #takes= part changes (on 10-25 a cached page showed none).
• Run-queue jobs for takes: takes-test DAY (the takes email marked [TEST], to the owner), takes-check DAY (what would publish now: email, site, Slack, with the privacy check; saves nothing), takes-apply DAY (chat tier: save the day's takes and publish; for a shipped day only the written takes change).
• The mail script via clasp (pipeline/clasp.py, ./np mailer status | setup | deploy): mailer-status checks the live script against the repo, deploy-mailer (Approve on the Mac) pushes Code.gs and updates the same deployment, so mail-script changes stop needing copy and paste. One-time setup by Steven: docs/EMAIL-SETUP.md.
• The site script is a raw string (Python's escape warning is gone).

v0.2.7-beta · 2026-10-25

• Takes on the website: the private takes email has a "Read and comment on the site" link to the day's page. With Steven's signed token (after the #, valid two days) the page shows a box after the day in brief, Top of mind, each Must-read story and Worth knowing, and one Submit (with "Done: send the edition now"). Visitors without the token see nothing different. The mail script checks the token and stores the takes; the pipeline reads them during the hold together with email replies and Slack, and the same privacy hold applies.
• Needs a redeploy of the mail script (scripts/apps-script/Code.gs: new takes / takes_pull actions, owner token scope t, a {{OWNERT}} placeholder for owner sends). Until then the site link shows no boxes and everything else works as before.

v0.2.6-beta · 2026-10-25

• Takes email: bold lead phrases were pale blue on white (the subscriber email's dark-mode colour) and now read dark; links are blue. A "W" slot covers Worth knowing (button in the takes email, a take box under Worth knowing in the subscriber email and on the site, and read after the quick hits in the podcast).

v0.2.5-beta · 2026-10-25

The critic's first batch (review of 2026-10-24), plus an easier way to add takes. - No story twice: after the deep dive is written, a Must read it covers shrinks to one sentence and a pointer, in the written edition, the email and the podcast. On 10-24 this would have cut the PHP story that took about 11 of 19 minutes. - The briefing agrees with its deep dive: "not disclosed" lines the deep dive answers are dropped from the day in brief and Top of mind; "unverified" lines about the deep dive's story are flagged for review. - Verdict labels match their evidence: "Holds" is no longer used on an evidence check that says the claim is self-reported or unverified; it reads "Unverified". - Source gate: a homepage link, or a sponsored or webinar page, can't be a Must read. - "For early-stage": a new line on each Must read on what the story means at seed and Series A (written edition, email, podcast). - Audio: a cold open with the lead story's sharpest line; loudness normalised to −16 LUFS at 96 kbps; "K3.1" is said "K three point one", not "Kthree"; a lint logs words the voice may garble; developing stories get a short "what's new" segment. - Takes email: it now carries the whole edition, with an "Add your take" button after each section that opens a reply with the section's number filled in, and a link to the page on the site. - Not in this batch: merging the day in brief and Top of mind (critic item 7), which changes the email, site, takes slots and audio together; it needs a preview first.

v0.2.4-beta · 2026-10-25

• Fix: v0.2.3's public code push was refused because the new critic page (docs/CRITIC.md) wasn't in the public export, so the exported tests couldn't find it. It's added, and a test now checks that every page on the site's docs is in the export.
• ./np sources add says "already a source" (and closes the queue entry) when the feed is already wired in, instead of "no feed answered".

v0.2.3-beta · 2026-10-25

• Fix: on Sun 10-25 the Mac went offline during the takes hold, so the publish at the 09:15 cutoff failed and the email was held. The only retries were at 06:30 and 07:30, before the cutoff, so nothing sent it later. With takes on, the schedule now also retries at 09:30, 10:30 and 12:00 (schedule.after_takes).
• Fix: v0.2.2's release was approved at 20:02, after the release window, and failed. The run queue no longer shows a release dialog that would be answered after 20:00, and an approval that lands outside the window now expires cleanly instead of failing.
• The run-queue job send-held (chat tier) runs ./np daily to publish and email a held edition.
• Includes v0.2.2 (the weekly critic email, ./np sources add), which wasn't released on 10-24.
• Docs: a new page, "The critic" (docs/CRITIC.md, also on the site): the rubric and its weights, the benchmark shows and why they were chosen, how a weekly run works, the email and the record, the limits, and the latest results (generated from data/critic/ at every publish: ranking, us against the field by characteristic, our trend). The architecture page gains a section on the weekly critic.

v0.2.2-beta · 2026-10-24

• The weekly critic: every Saturday the briefing is scored against the latest episode or issue of ten benchmark shows, with a 10-part rubric weighted for an investor daily. A private email to the owner gives our score and rank, bar charts by characteristic against the field, our trend week by week, what shipped, and the top suggestions with recommended options (pipeline/critic.py, ./np critic preview | email | list, the run-queue job critic-email). Results are kept in data/critic/<date>.json; the first is 2026-10-24.
• ./np sources add "Name" URL [URL] [note] and the run-queue job source-add: verify a feed live on the Mac, then wire it in. Used to add the ten benchmark shows as sources.

v0.2.1-beta · 2026-10-24

• Fix: the first takes morning (10-24) never sent the takes email. The takes step called the mailer with positional arguments it doesn't take, so it failed straight away and the edition shipped on time without takes, as designed. Fixed, with a test that checks each call against the mailer's real signature.

v0.2.0-beta · 2026-10-23

• Steven's takes (My takes, phase 1): the owner's own view in the edition, verbatim, labelled "Steven's take".
• Once the edition is written, a private "AI Briefing · Your takes" email lists slots: 0 (the day), T (Top of mind), and 1 to N (each Must-read story). Reply "1: ..." from a phone; "done" sends early, "skip" means none.
• Your own words on a link you post in the team Slack channel count as your take on that story.
• The edition waits until 09:15 or "done" (./np takes done, or takes-done in the run queue).
• Each take is checked for private names, owner addresses, personal links, disclosure phrases and the confidential marker. One that fails is held back and listed to you, never published.
• Takes appear on the site and in the email, and the host reads each one after its story in the podcast. Only the briefing audio is re-rendered.
• Retries leave a waiting run alone; a failure ships the edition without takes.
• ./np takes status | email | preview | done.
• Fix: the Slack reader closed or leaked its HTTP client when it found your own LinkedIn posts.

v0.1.14-beta · 2026-10-23

• First-lines check for the daily briefing: the day in brief and Top of mind are now checked against the edition's own body, as deep-dive summaries have been since v0.1.9 (consistency.check_briefing). A figure belongs to its whole story, so a line can name the company in its lead. A sentence that ties a figure to the wrong company, or says something wasn't disclosed when the body gives it, is dropped.
• Across days (flag): a "not disclosed" claim is compared with the previous 14 editions and flagged for review if an earlier day gave the figure for the same company.
• The "not disclosed" test is stricter everywhere: a question asking for a figure, or a sentence saying it's lacking, no longer counts as giving it, and it must be about the same company.
• ./np consistency [DAY|all] (also a run-queue recipe) runs the checks over published editions as a dry run. Over all 18 editions and 8 deep dives it found one misattributed figure (10-22's day in brief) and two cross-day flags (10-23).

v0.1.13-beta · 2026-10-23

• Fix: the run queue's launchd tick never started a job. Its quick idle check listed both queue folders with ls, which fails whenever one of them is empty. It now checks each file, and a test runs the real ./np runner tick.

v0.1.12-beta · 2026-10-23

• Run queue: releases, corrections and checks can be started from the Claude chat instead of pasted into a shell (pipeline/runner.py, ./np runner, docs/plans/run-queue.md). Claude queues a named recipe and the Mac runs it. There are three tiers: tests, staging and verify run at once; corrections and publish need Steven's yes in the chat; releases, rollbacks, approvals and the public code push need Approve in a dialog on the Mac. The click grants presence for that job only (pipeline/presence.py), and no file can stand in for it. Release invariant 10 guards the runner. One-time setup: ./np runner install.
• Docs stay current: release-gate tests now require every public module to be described in a hand-written doc and every diagram to be used. Each release lists the modules that changed while their docs didn't ("docs to review"), in the approval dialog and in data/state/docs-impact.json, for a weekly docs refresh.
• Docs: Fact checking is rewritten as "Grounding and fact checking": six layers, a new diagram, why the first-lines check exists, corrections and known gaps. Architecture gains sections on grounding, the run queue (with a diagram) and keeping docs current, plus every smaller module.

v0.1.11-beta · 2026-10-23

• Fix: v0.1.10's public code push was refused again, correctly: a code comment and the changelog named a portfolio company. Both are reworded, and the dev test now builds the whole public export and fails on any private term, not only in test files.

v0.1.10-beta · 2026-10-23

• Fix: v0.1.9's public code push was refused, correctly, because a new test used real portfolio names. The test now uses made-up names, and a test fails in dev whenever any test file names a private term, so the release can't be the first place it's caught.

v0.1.9-beta · 2026-10-23

Fix: on 2026-10-23 the deep dive's body was correct, but its summary and "Why today" lines, which readers see first in the email, on the site and in the podcast notes, said a legal-AI company disclosed no ARR and that "a vendor" claimed a $21M saving. In fact the company disclosed ARR, and the $21M was an investor's post about a different company. The day before, an unsourced Agenda line had called that investor the other company's "founder". - The why-today line comes from the checked body: it's written from the deep dive's checked "why today" section, no longer copied from the previous day's topic suggestion. - A header check (pipeline/consistency.py) for the summary and why-today lines: a sentence is dropped if a figure isn't in the checked body, if it names none of the companies or people the body ties that figure to, or if it says something wasn't disclosed when the body or a source gives it. Spelled-out figures in scripts count. If nothing in the summary survives, the checked opening is used. - Agenda lines can't assert unsourced roles: "founder", "CEO", "investor in" and similar must appear in the item's own source text; otherwise the line becomes neutral. - Deep-dive text is guarded like the briefing: a shipped deep dive changes only through a correction, and its page shows the correction note. - ./np correct DAY "what changed": records a hand edit to a shipped edition as a correction, shown on the day page and the deep-dive page.

v0.1.8-beta · 2026-10-23

Fix: the 2026-10-23 edition was held at 05:26 and again at 06:30. The release gate re-ran the whole test suite during the morning run, and one release test that checks for a running daily run saw the real one and failed. The same check blocked the backup's git push. - Released code isn't re-tested in the morning: once a code and settings fingerprint has passed the full gate (tests included) and is approved, later checks accept it without re-running the tests (data/state/release-passed.jsonl). Changed code is still fully checked. - Tests never read the live machine: test runs set NP_TESTING=1, run_in_progress doesn't look at real processes under it, and the release tests mock it too. - The evening preflight tests tonight what runs tomorrow: if the code hasn't passed the full gate yet, the preflight runs it at 20:30 and emails the owner if it fails, instead of finding out at 05:00. - The pre-push hook uses the same fast check (--release-check require), and ./np release reinstalls the hooks.

v0.1.7-beta · 2026-10-22

Shipped editions don't break. Once a day's email has gone out, that day is shipped (pipeline/shipped.py), and: - Publish guard: before every production publish, the new site is checked against what readers already have. It refuses (the live site is untouched and the owner is emailed) if any of these hold: - a shipped day page or deep dive is missing; - a shipped edition's text changed without a correction; - a shipped MP3 inside the retention window is missing; - the home page or a feed is missing, or the podcast feed would shrink. - Corrections are explicit: rebuilding a shipped day (rebrief, reprocess, remaster) needs --correct "what changed", and the page shows "Corrected <date>: …". - Remasters keep the old audio: the earlier MP3 is carried over from the live site, so no audio URL breaks. - After-publish check: once the site is live, the newest shipped pages, both feeds and the latest MP3 must answer. If not, the previous site is put back automatically, the email is held, and the owner is emailed. ./np publish restore does the same by hand. - The email waits for a good publish: if the publish fails or the after-publish check fails, no subscriber email goes out. Before, it went anyway. - Audio retention: 30 days on the site and in the podcast feed, up from 14. 60 days would be about 1 GB, the GitHub Pages limit. - Slack: links any member posts in the team channel now count, links only. Who posted is never stored, and their text is never used. - Staging preview: carries only the newest day's audio. The first full preview push timed out. - Site pushes are sturdier: a push that stalls (under 1 KB/s for 60 s) fails and is retried once over HTTP/1.1, instead of hanging the morning run. Pushes no longer use one huge buffered upload, which is what GitHub was timing out with HTTP 408.

v0.1.6-beta · 2026-10-22

• Development preview: ./np staging publish [setup] publishes the dev checkout's site to its own Pages repo (sgwoods/ai-briefing-staging). The preview has a banner on every page linking back to the published site, is kept out of search engines (noindex, robots.txt), and has no subscribe form and no podcast feed. Staging can publish only there, and production never publishes there.
• ./np staging compare [DAY]: staging against production for the same day: sections, counts, words, source links, the evidence mix, fact-check counts and audio length, plus the privacy scan of the preview and a diff of the edition. Writes logs/compare-DAY.md and flags what needs review.
• Fix: a staging run no longer copies its MP3 into the owner's iCloud folder. The first staging run had overwritten that day's copy. It no longer deletes iCloud files either.
• Tests always run as production (NP_ENV=production), in either checkout.

v0.1.5-beta · 2026-10-22

• Fix: the scheduled jobs had been firing exactly one hour early every day (the 05:00 run started at 04:00, the 20:30 preflight at 19:30). launchd now only ticks every half hour, and the Mac's own local clock decides when a scheduled time has come (pipeline/schedule.py, ./np scheduled). Times must be on :00 or :30.
• Releases refuse checkout-local files (.venv, .env, .staging) committed on dev; .venv is ignored as a link too.
• Fix: rebuilding an edition (./np rebrief, ./np staging run) failed with KeyError 'critique', because the evidence check on Must reads needs the critique model, and a rebuild had stopped loading it. Found by the first staging run.
• Git safety: every git call in the export, hub, dev and release code ignores GIT_DIR / GIT_WORK_TREE from the environment, and the tests that build throwaway repositories clear them. A leaked GIT_DIR once pointed a test's git init at the production repository and wrote a wrong core.worktree into its config. Caught and fixed before any run used it.
• The first release built on dev and promoted with ./np release.

v0.1.4-beta · 2026-10-22

• Development checkout: ./np dev setup creates a second working copy of the repository on branch dev, in staging mode. Inside it:
• ./np staging sync brings in production's latest data and fixes;
• ./np staging run [DAY] rebuilds an edition with the development code while outbound stays off, and --email sends the review copy to the owner only;
• ./np staging reset throws away what staging wrote.
• Staging outputs can't be committed: the pre-commit hook refuses data paths in a staging checkout, archive commits are skipped there, alerts aren't sent, and services can't be installed. Staging runs keep out of the 03:30–08:30 production window.
• Releases and rollback:
• ./np release VERSION "reviewed: ..." merges dev into production between 08:30 and 20:00, needs one approval, then tags the release and pushes the public code, the backup and the hub status. It refuses if production has stray code changes, a run is in progress, dev's version or changelog don't match, or the staging marker was committed.
• ./np rollback "reviewed: ..." returns the code to the previous release.

v0.1.3-beta · 2026-10-22

• Documentation on the site at /docs/, rebuilt at every publish from the same markdown as the public code, with the diagrams embedded. Each page shows the running version and when it last changed. Linked from the footer and the hub page.
• Generated sections in the architecture document: current facts and a component table built from each module's own description. Drift tests fail a release if the document names a module that no longer exists, or doesn't name the current version.
• Release check: the tests line shows unittest's own summary.

v0.1.2-beta · 2026-10-22

• Staging mode: a development checkout (marked by an untracked .staging file) can build everything but can't reach readers. The release gate refuses every outbound action there. The mailer allows only reads and review copies to the owner, marked [STAGING]. A new release-gate invariant keeps every outbound function behind the gate.
• The hub project page shows the latest release notes and links to the architecture document.
• Architecture document with four diagrams: system overview, the daily run, repositories and privacy boundaries, and the release flow (docs/ARCHITECTURE.md).

v0.1.1-beta · 2026-10-22

• AI Briefing on the public project hub (sgwoods.github.io/public) as an exporting project: after each publish, ./np hub push writes its status (19 fields, naming this released public commit) and its project page with the hub's freshness panel, and pushes only those two files. It refuses when the running code isn't the release.
• The site's old /hub/ copy of the status files is retired.
• ./np log <command> saves any command's output under logs/.

v0.1.0-beta · 2026-10-22

The system as it runs today, labelled the first beta: a small number of outside readers.

• Daily edition: fetch from feeds, sitemaps, Bluesky lists, newsletters and Slack links; filter, summarize, rank, cluster into stories, critique, verify and fact-check; Top of mind, Must read, Worth knowing, Open questions.
• Evidence-weighted sources: every source has a profile, and a Must read needs primary or reporting evidence.
• Podcast: a ~10-minute episode read by the host, followed in the same episode by the day's deep dive.
• Public site, podcast feed and RSS on GitHub Pages; a subscriber email sent after the site is live.
• Release gate on every outbound step: tests, secret scan, privacy invariants and Steven's approval of security-sensitive changes; owner alerts when action is needed; a cloud watchdog for missing editions.
• Reader feedback collected from email replies and triaged daily.
• Public code repository: an allowlisted, scrubbed export (./np push-code), MIT licence.