The budget meeting — what it commissions, whom it picks, and why study · 2026-09-08 · no code yet
The owner's three complaints after five editions on the box, each checked against the 62 published pieces (09-04 to 09-08) and the code that produces them: the topics repeat and cluster; a handful of personas write while most never get a slot; the persona-to-topic match swings between the obvious and the absurd. This page is the study and a design sketch. Nothing here is built.
Our topics are repetitive and only focus on several areas. Only a handful of the personas got the opportunity to write; many get no chance. The topic-persona matching is sometimes too stereotypical and sometimes too far-fetched — asking Federer to give a pep talk to a football team?the owner, 2026-09-08
1 · The topics repeat, and the harvest is why
| Subject cluster, 62 pieces | pieces | Examples |
|---|
| AI · Nvidia · OpenAI · Anthropic | 9 | Hinton twice, Wolfram, Dan Wang, Rothman, Brownlee, Ada Lovelace twice |
| film · books · art | 8 | Kurosawa twice, Hamaguchi, Conan Doyle, Koons, Banksy twice, Stan Lee |
| the Fed · rates · bonds · IPOs | 6 | Munger twice, Buffett, FDR, Moritz |
| the Iran war · Hormuz · the Black Sea | 5 | Churchill twice, FDR, Orwell, Lincoln |
| space · astronomy | 4 | Hawking, Leonardo, Feynman (Mercury), Euler |
| the Chinese economy · the FT report · CFIUS | 4 | Feddo, Li Ka-shing twice, Dan Wang |
| parenting · relationships | 3 | Chapman twice, Chestnut, Ephron |
| sport | 2 | Nadal, Federer — in five days, with a Sports beat on the list |
| everything else | 21 | |
The top four clusters are 45% of the paper. The slate policy's own cautionary example, a reader meeting the Iran war three times in four days, happened five times in five. Sports got two slots in 62. Beats look balanced on paper (Culture 14, World 9, Life 9, Science 9, China 7, Tech 7, Markets 5, Sports 2) and are not, because a beat is not a subject: Science was space four times.
the causeThe harvest asks the same thirty-three questions every morning. The news search runs eight English queries and six Chinese ones, one per beat, plus nineteen per-beat extras, phrased like「technology AI news today」and「science news this week」, plus a per-beat list that gives AI and China two extra queries each and Sports one. A generic query returns the same outlets' front pages every day; since Phase 1.5 the harvest opens those pages and takes their stories, which made the pegs real but not different: the same dozen sites, the same five running stories. The shelf's subject cap only stops an exact repeat of a peg within four days, and「prefer the beats least covered in the last 3 days」is a soft line the meeting reads after 140 pegs. Nothing counts a cluster, so the AI story comes back every day under a new headline.
2 · Forty-six personas have never written
| Roster | 92 |
|---|
| wrote at least once in five editions | 46 |
| wrote exactly twice | 16 |
| never wrote | 46 |
The twice-used: Churchill, Twain, Lincoln, Munger, FDR, Hinton, Dan Wang, Chapman, Confucius, Kurosawa, Banksy, Li Ka-shing, Ada Lovelace, Bob Dylan, Ke Yu, Asimov. Never used, a sample: Karpathy, Jensen Huang, Elon Musk, Steve Jobs, Jack Ma, Zhang Yiming, Zhang Xiaolong, Keynes, Joan Didion, Lu Xun, Seneca, Kant, Newton, the Buddha, Oprah, Ang Lee, James Cameron, Liu Cixin, Esther Perel, Phil Knight, Jony Ive, Michael Porter, Masayoshi Son, Sa Beining, Liang Ning, Xiao Pang.
the causeThe meeting picks the names it knows best, and the only brake is a rest after two. The roster reaches the model as 92 one-line entries: name, role, tagline, and a cadence word (fresh · wrote 1 this week · resting). Faced with a war peg it picks Churchill; a Fed peg, Munger; a marriage peg, Chapman: the most famous name for the most obvious subject, every time. The cadence rule rests a persona only after two pieces in seven days, which is exactly where sixteen of them sit, so it is the one thing that stopped a third Churchill. Nothing prefers the fresh, nothing rotates, and nothing on the card says what a persona would want to write about: the card's writes: is the language it writes in, not its subjects, so the meeting has no field to match on beyond role and tagline. Six personas landed on 09-07 (Orwell, Keynes, Ephron, Didion, Lu Xun, Seneca); two were used, two were written and dropped at the gate, two were never picked.
3 · The matching: obvious, or absurd
| Mode | Pairings from the five editions |
|---|
| Home ground, to the point of stereotype | Feynman on the path integral · Hinton on AI regulation, twice · Munger on the Fed, twice · Nadal on the Williams sisters · Churchill on a naval war, twice · Chapman on marriage, twice · Hawking on a star census · Darwin on a fossil gap · Asimov on a brain study and an AI drug · Brownlee on a gadget company |
| Far-fetched | Federer writing to an Indiana college-football freshman about the AP poll · Lincoln writing to a passenger on the「Peace Express to Ukraine」· Harari reviewing a 2023 Polish election as news · Bob Dylan on the August jobs report, pegged to nothing · Banksy taking Kim Kardashian round a museum · Twain's beaver-and-moose fable on Canadian tariffs |
| The ones that worked | Bourdain's lunch-counter dispatch (the owner's favourite: inspired by a topic, not a paraphrase) · Kurosawa on a Hamlet film · Dan Wang on Nvidia from a Chinese lab · Orwell on the Black Sea · Ephron on sleep training · Tao's letter to a teenage mathematician: expertise with a personal reason |
the causeOne sentence in the policy licenses both extremes.「Match persona to peg by genuine expertise or a sharp, unexpected vantage — an outsider's angle is welcome when the reason is stated.」The first half is read as「the famous name for the famous field」; the second as「anyone for anything, if you can write a sentence about it」, which a model always can: Federer knows what a ranking is, so Federer gets the football poll. The meeting is never asked for the persona's reason to care, only a reason it could speak; and it has nothing to find one with, since the roster line carries no memories, no values, no subjects. The pairings that worked share one thing: the persona had a documented, personal connection to the subject, not a professional adjacency.
4 · A design sketch no code yet
the harvestRotate the questions and cap the outlets. A pool of five to eight queries per beat, a different pair each day, phrased around what happened rather than a beat name (「a court ruled」,「a study found」,「a match ended」,「a company shut」). A per-host cap in the harvest so one outlet cannot be a third of the pegs. A cluster cap in the slate, not just a peg cap: the same running story at most twice a week, counted by the tags the meeting already writes.
the rosterShow the meeting a bench, not a list. The personas who have not written in fourteen days come first, each with a derived line of subjects and one line from its profile that says what it actually cares about; a quota that at least a third of the day's slots go to personas with no piece in fourteen days; and the cadence rule turned around: rest after two stays, and「fresh」becomes a preference the validator enforces, not a word the model may ignore.
the matchTwo modes, named, with a quota, and a reason the validator can check. Home ground, the persona's own field, at most two thirds of the day. A real vantage, the rest: the persona has a documented personal connection to the subject, and the meeting must cite the profile line that gives it (a memory, a value, a place, a loss), never a professional adjacency. A pairing whose reason is a pun on the persona's fame, or that crosses a sport, a country or a century for no reason in the profile, is refused. Federer writes about a football poll only if his profile says something about polls and being ranked; it does, but the reason must be said, and then the piece is about ranking, not about football.
what this costsNothing in model spend: the harvest is fetches, the roster shape is a prompt change, the quotas are validator lines. The one real cost is that a bench of unfamiliar names produces worse pieces at first, because the meeting has less to go on; that is the price of the roster being used, and the profile line in the bench is what pays it down.
the owner's answerThe owner read this and set a different course (09-08): the harvest must
span the news space, not ask better beat queries; and a writer meets a topic two ways, the editor's match-making and the persona's own motivation. Developed on
coverage and the pitch.