A Better Newspaper

The podcast desk's alarm, the search behind From the Wider Web, the archive of the knowledge base, and the paper's improvement desk each had a way to fail that read as a quiet day; each now says so.

• Behind the scenes: the alarm that watches the podcast desk no longer counts an episode the paper gave up on as proof the desk is working. The search that feeds From the Wider Web now raises an alarm when it cannot run or every search fails, instead of printing as a day on which the web had nothing new, and each search has a time limit. The archive of the knowledge base has the time it needs, reports when it could not be saved, and leaves a record a later check confirms. The improvement desk asks once more for a reply it cannot read, and reports a failed read of the Critic's grades, instead of proposing nothing. Nothing changes for readers.

When the day's reading cannot be handed to the knowledge base, the publisher now hears about it the same day; the sorting step stops in time on a heavy news cycle; and a knowledge-base update that comes back with unusable details keeps the page's existing summary and ratings instead of failing, or writing a rating off the paper's scale.

• Behind the scenes: the step that sorts each day's articles into sections also notes what each article says about the people, companies, and storylines the knowledge base tracks. When those notes could not be passed on, the articles were still marked as read and quietly never reached the knowledge base, with every alarm green. The publisher is now told the same day. On a heavy news cycle the sorting step now stops starting new work in time and leaves the rest for a later pass, instead of being cut off mid-way with no record. A knowledge-base page update whose summary or ratings come back malformed keeps the page's current summary and ratings rather than being given up on, and a rating that comes back above the paper's scale is held at the top of the scale instead of written as it came.

The paper's weekly improvement desk now reads more of what the knowledge base has learned about AI and newsrooms, reads the same pages for the same data each time, and no longer re-proposes an idea already turned down.

• Behind the scenes: the paper's weekly search of its own knowledge base for ways to improve now finds the pages that bear on AI and newsrooms even in a period dominated by war and market news, gives the same answer for the same knowledge base every time, and remembers the ideas the publisher has already declined, so they are not proposed again. Nothing changes for readers.

A section whose material could not be loaded, a desk reply the paper could not read, or a Sunday column that came back unusable now raises an alarm, and is asked for again where that can help, instead of printing as a quiet day.

• Behind the scenes: when a desk's reading material cannot be loaded, the desk is reported as failed rather than writing "No significant developments today", so the publisher hears about it the same morning. A desk reply the paper cannot read as a list of items is asked for again once, instead of being printed as an empty section.
• The Analyst's Sunday column is held to the same standard as the daily desks: a column that came back cut off or unreadable is asked for again once, and the publisher is told when no column could be stored, instead of the paper quietly rebuilding it later in the day.
• A podcast episode whose highlights could not be read is no longer marked as done with nothing to show; it is tried again on a later pass, a few times, before the paper gives up on it and says why.
• The Ideator and the front page now run under the same morning time limit as the desks. On a morning when the desks have used that time, the paper ships with a plain front page and without the Ideator, and the publisher is told, rather than the morning's work being cut off with no record of what happened.

The paper checks once a day with its mail service whether each edition reached its readers and keeps that as its delivery record; it no longer raises an internal alert asking for a second, live delivery feed.

• Behind the scenes: since October 10 the paper has asked its mail service each day whether the day's sends arrived and kept the answer as its record of delivery. An internal alert that kept asking the publisher to connect a live delivery feed on top of it is retired; the daily check is the record. Nothing changes for readers.

Summaries, context lines, and the mass-tort desk's next-step notes were reaching the Critic cut off mid-word; the Critic now reads each item as the reader does, and a desk that leaves a sentence unfinished is asked to finish it or has the block cut back to its last complete sentence.

• The Critic grades what the reader saw. Its copy of each item had been shortened to a fixed number of characters before grading, so it was marking the paper down for mid-word cuts that never appeared in the email or on the site. It now reads each summary and context line in full.
• A desk whose summary or context line stops mid-sentence is asked once to complete the final sentence. A block that still ends mid-sentence after that is cut back to its last complete sentence rather than printed as a fragment, and the cut is recorded where the publisher's daily checks can see it. The Ideator's business idea is held to the same standard.

Each item in the paper now carries a line saying which desk wrote it, from which source and date, and what was checked; knowledge-base pages point readers to related pages; and sixty-seven robustness fixes from an outside-eyes audit landed in one batch, with the publisher's sign-off reserved for decisions about what the paper is.

• A provenance line under every item, in the email and on the site: the desk that wrote it, the outlet and date it drew on, that the source link was checked against what the paper actually read, and the knowledge-base page behind its context line. On the site, items in an edition the Critic has reviewed say so.
• Every knowledge-base page lists up to five related pages that share sources or subject, so a reader who lands on one account of a story can reach the others. Near-duplicate pages can now be folded together, with the publisher's approval, keeping every page's history visible; nothing is deleted.
• Delivery is stricter about the truth: the paper is never sent with no sections, a day on which nothing reached anyone is never recorded as delivered, no reader can receive the same edition twice, and readers east of UTC receive the paper at their own morning hour. The Sunday Weekly reviews exactly one week, greets each reader by their own name, and reports a failed send instead of hiding it.
• Readers whose own hour falls late in the paper's day, such as a later morning in Hawaii or the central Pacific, are no longer missed.
• The knowledge base guards its memory harder: a page update that fails to save is never recorded as done, a returning story reopens its archived page instead of vanishing, new pages carry their real sources, and before creating a page the paper now checks whether the story already has one.
• Each desk's source links must match a page the paper actually read; an item whose link matches nothing, or has none, is sent back to its desk once to be re-sourced and otherwise left out rather than published with a plausible-looking link. A formatting check runs before publication, on the same standard the Critic grades against. A desk that runs long is cut short and named, so one slow desk no longer delays or repeats the whole paper.
• New knowledge-base pages now list the articles they were built from, and a story the paper already covers is added to its existing page instead of getting a second one. The morning audio edition is added later in the day when a paper was composed late, instead of being skipped.
• Behind the scenes: the paper's maintenance now runs on two lanes. Changes to how the paper is built flow through an independent reviewer and the test suite without a click; changes to what the paper is wait for the publisher. A Sunday receipts email lists every change, and a brake stops automated merges on its own if the Critic's scores fall or incidents climb. Sixty-seven robustness findings from the October audit landed in the same batch: checked writes, paged reads, capped retries, deadlines on every model call, and tests on the places where past incidents happened.

The paper-money desk starts again with its full stake invested on day one and a weekly deposit, scored against two benchmarks that receive the same money on the same days; the Ideator may add one simple market idea a day; and the paper's own operating room got a lock on the door.

• The Investor Desk relaunches with $100,000 of paper money deployed across six to ten positions on day one. A $500 paper deposit arrives every week, and the desk makes one new pick sized to it, plus up to two rotations. Positions from the desk's first run are closed at the relaunch; its record starts fresh.
• Two benchmarks receive the identical deposits on the identical days: the S&P 500, and cash at the 4-week Treasury bill rate. The desk reports both a time-weighted return, which is skill with the deposits taken out, and dollar profit, which is what a reader who shadowed every deposit would have.
• Every price on the desk carries its date and a link to the page it was read from. A new day-by-day page, /desk/history, lists every deposit and every market day with the holdings it was valued at. Shares are fractional, so a $500 pick can buy any stock. Each card says the day its target or stop was hit, or why it carries no verdict yet.
• The Ideator may now add a Market Idea between the business idea and the Stoic Thought: one simple, long-only position in a US-listed stock or fund that expresses a thread of the day's news, with a link to the story it grows from and what would prove it wrong. It is a paper exercise, not a recommendation and not investment advice, and the paper says so under every one. The business idea now focuses on something to build, acquire, or structure. The Sunday Weekly lists the week's market ideas.
• Behind the scenes: the publisher's private pages now sit behind a sign-in instead of a key in the address bar; the paper's database can no longer delete its own wiki, editions, subscribers, or records; and the scheduled sessions that maintain the code gained a way to run the whole test suite before they push, plus a written limit on what they may merge without the publisher.

The Critic's grades now feed back into how the desks write, a second reader checks the Sunday column before it runs, and the paper maintains its own code — with the publisher's sign-off on every change to how it thinks.

• Standing notes: when the Critic keeps finding the same shortfall at one desk, a short, versioned note tells that desk what to do differently. The publisher approves each note; the desk reads it from the next edition; its effect is measured on the same fixed standard, and a note that does not help is retired. The standard never moves; the desks do.
• The Analyst gets a second reader: before the Sunday column runs, an editor from a different model family checks it for overclaimed numbers, loose attribution, unearned confidence, entities priced as one, conflated concepts, and a steelman that dodges the strongest objection. The Analyst then revises.
• The paper files its own incident reports and fixes them through pull requests that run the test suite; improvements are proposed weekly from what the paper read and how it was graded, and built only after the publisher says yes.
• Readers are the number the system serves: open and click rates now reach the publisher and the Engineer every week.

The paper is free to read and costs money to make, so it opened a desk that looks for the grants and programs that fund work like this and drafts the applications itself — with the publisher deciding what goes out, and the whole record in public.

• Once a week the desk searches open grants, fellowships and startup programs, reads the eligibility rules, and scores the fit. For the best fits it drafts the application from the paper's live numbers: editions, knowledge-base pages, citations, subscribers, the Critic's grades.
• Every claim in a draft is tied to something a reader can check on this site. The desk writes in the paper's own voice: short sentences, exact numbers, no claims without receipts.
• Nothing is ever sent by the system. The publisher reads each draft and approves, passes, or submits it by hand. The public record at /grants lists every program considered, including the ones passed on, and will carry outcomes as they arrive.
• Site search now finds past stories as well as knowledge-base pages: a term that appeared in an edition but never became a wiki page is no longer invisible.
• A short “From the desks” note now closes each daily paper with how the Investor Desk and the Grants Desk stand, and introduces each desk for its first week.
• Every public page now explains itself to a first-time reader: what the wiki is and what its numbers mean, how to read the Investor Desk, what each section is, and — in the open — what the paper does not publish about how it is made.

The paper opened a paper-traded model portfolio: a public, timestamped test of whether its reporting and memory add up to good judgment — and a worked investing lesson every week.

• $100,000 of paper money, long-only, at most ten positions. Every decision is written as a card — thesis, target, stop, horizon, what would prove it wrong, and the lesson it teaches — before the fill price exists, and fills at the next day's close.
• Scored in the open against the S&P 500 in numbers anyone can check: what $100 at launch is worth now, the worst drawdown, and how many closed trades made money.
• The Investor reads the week's editions, the Analyst's column, and the knowledge base's most important storylines — the same memory the rest of the paper runs on.
• No real money, ever, and nothing on the desk is investment advice. It exists to build a track record first.

A Critic now reviews every edition against a fixed editorial standard, so improvement is measured rather than asserted — the first piece of a newsroom that gets better on the record.

• A daily Critic — a separate model from the desks, so it does not share their blind spots — grades each edition on seven fixed criteria: accuracy, citations, selection, overlap, continuity, writing, and front-page judgment.
• Every shortfall is attributed to the part of the newsroom responsible (sourcing, triage, a desk, dedup, the knowledge base, the editor-in-chief), not just scored.
• The Critic also sees what was available and did not run, so a missed story counts against the paper the same way a weak one does.
• Grades are recorded from today forward under a versioned rubric; when the standard changes, the scoreboard starts a new epoch instead of moving the goalposts.

The paper's memory became inspectable: context lines link into the knowledge base, every page carries a public revision history, and the brain's growth is measured in the open.

• “Context:” lines in the paper now link to the knowledge-base page they draw on — walk from today's insight to the accumulated page to its primary sources.
• Wikipedia-style page history: every knowledge page shows its revisions — when it was created, when new reporting updated it, and which articles drove each change.
• Live statistics on the wiki: pages tracked, source citations, revisions this week — real queried numbers, checkable against the histories themselves.

The paper's smartest seat moved to the newest frontier model, and the paper began publishing its own version history.

• The Analyst's Sunday column now runs on Claude Fable 5.1, Anthropic's newest frontier-tier model (released September 1).
• This page: a public version history, so readers can see exactly how the paper has evolved — and which models sat at the desks when.

The knowledge base stopped just growing and started deepening — new reporting now stacks onto the pages it belongs to, and the paper's own analysis became part of its memory.

• Total capture: every edition — including the Sunday Weekly and The Analyst's column — is now permanently archived. Nothing the paper produces is ephemeral anymore.
• The Analyst gained memory: it reads its own past columns and advances its theses week over week instead of starting from scratch each Sunday.
• Two-stage research for The Analyst: it surveys the complete knowledge-base index (~2,000 pages), chooses what to study in depth, then writes — genuine whole-library access every week.
• Editorial memory lives on its own shelf, deliberately separate from source-tagged facts, so the paper's opinions can never masquerade as reporting.
• Smarter knowledge routing: new articles about a known entity or storyline now update its existing page (name-variant matching, aliases, similarity search) instead of spawning near-duplicates.
• Spring cleaning: 73 groups of duplicate pages were merged into single, deeper canonical pages — with full revision history preserved.
• Months of articles were re-mined into the knowledge base to recover updates missed during a summer defect.
• Knowledge-base update cadence tripled: twelve passes a day.

A month spent making every silent failure loud. The paper had alarms for its worst-case scenario; now it has them for everything.

• Every failure mode now announces itself: missed deliveries, a failed morning compose, a dropped section, API trouble, a stalled knowledge base, a missing weekly.
• Self-healing delivery: if the morning composition fails, the hourly delivery pass rebuilds it automatically.
• Robust output parsing across every AI agent — malformed responses are repaired or retried instead of silently costing a section.
• A defect had quietly capped the knowledge base at 1,000 pages for ten weeks; it was found, fixed, and the archives immediately jumped to their true size (nearly double).
• The weekly knowledge-base archive now includes every page — active, stale, and archived.

The daily became the same paper seven days a week, and Sunday gained a second, more ambitious edition built for connecting dots.

• The Sunday Weekly: a separate edition led by The Analyst — an original synthesis column written from the knowledge base and the week's reporting, on the strongest Claude model the paper runs.
• The daily Stoic Thought became human: real, verified public-domain quotations (Marcus Aurelius, Seneca, Epictetus, Thoreau, Montaigne…) — the AI selects, never writes, and every quote is verified verbatim against its source text.
• A quote harvester that reads Project Gutenberg editions and admits only passages that match the original text exactly.
• Public signup opened, with daily and weekly-only subscriptions.
• Front-page headlines link straight to the underlying source article.
• Reader-facing section order redesigned: hard news first, discovery in the middle, reflection to close.

The paper learned to look beyond its fixed reading list — into markets, preprints, government data, and the open web.

• Prediction markets (Polymarket, Kalshi) attached to stories as verified odds with day-over-day movement — the crowd's pulse alongside the reporting.
• New source classes: research preprints (arXiv, bioRxiv), federal contract awards (USAspending), and earnings-call transcripts.
• The Scout: a discovery agent that searches the open web daily for what the fixed sources missed — and can only cite what the search actually returned.
• Podcast pipeline rebuilt around full transcripts: timestamped highlights mined from every episode, not summaries of video titles.
• Estate Intelligence: a dedicated trusts, wills, and estates practice feed.
• Every claim's provenance policy formalized: no citation the system wasn't actually given.

A personalized daily newspaper, composed by an AI editorial team from dozens of hand-picked sources — and, from day one, a knowledge base that remembers.

• The daily edition: a sectioned morning briefing (AI & technology, science, business & markets, legal, world news, podcasts, classifieds, and The Ideator's daily business idea) delivered by email and on the web.
• Every claim linked to its source — the founding rule.
• The knowledge wiki: pages for entities, developing storylines, and cross-domain patterns, built and updated from each day's reporting with source-tagged provenance. The paper was designed to get smarter as this grows.
• Editorial context from the paper's own memory rendered separately from source reporting, so readers always know which is which.