Coverage, and the pitch — spanning the news, and two ways a writer meets a topic coverage BUILT 2026-09-09 · the pitches BUILT 2026-09-09

The owner's answer to the budget-meeting study: the world is big and the harvest sees a patch of it; and a writer can be given a topic by an editor, or can want one. Sections 1 to 3 are coverage. Section 4 is the two doors. Nothing here is built.

1 · Coverage, in four words

Three of the four are already the brew's own. A peg is a candidate story with its link. The harvest is how pegs are gathered each morning. The slate is the day's sixteen commissions. The fourth is the owner's: the space is everything the paper could be about, and it has three sides: where it happened, what it is about, and what kind of thing happened (a ruling, a death, a discovery, a match, a strike, an anniversary, a small strange thing). Coverage is the share of the space today's pegs came from. Nothing else is needed.

2 · The patch, measured

The 09-09 harvest, 82 pegs, placed on two sides of the space: the beat, and where the outlet lives. The third side, the kind of event, the harvest cannot see at all, because it searches by beat name.

Thirty-three searches such as「technology AI news today」and「今日体育新闻」, 15 outlets, the New York Times alone 17 pegs. Two thirds of the pegs are American; Culture is a third of them because book, film and art sections open wide; Sports, Markets and China have two pegs between them and「everywhere else」has none. A search by beat name asks Google what is loud, and the loud centre of every beat is the same five outlets.

3 · The harvest, redesigned

Two ways in, then two rules, then a count.

How the aggregators sample, and what we take from it

The common mechanic is cluster first, then rank the cluster by how many sources carry it. We already fold duplicates by title overlap; the change is to count what we fold instead of discarding it. A story that arrives from forty outlets in nine countries is important whatever its topic.

Every part touched: the checklist

The specimen's sixty sub-field searches were a first cut, and the owner's question (art? films? games? what else?) is the right one: touched-at-least-once has to be a property of a list, not of luck. The list below is the checklist the search layer walks, about 130 lines under the eight beats plus the kinds of event. Each line is one free Google News search; the rota puts every line through at least once a week and the count on the brew log names the lines a week has not touched. The list is a table in the repo, edited by hand when a gap shows.

Against this list the specimen's sixty searches missed: religion, weather, crime and trials, military, energy, oceans, chemistry, neuroscience, animals, agriculture, poetry, classical music, art crime, fashion, design, comedy, podcasts, comics, language, philosophy, celebrity, bonds, currencies, commodities, banking, luxury, airlines, media, scams, baseball, boxing, cycling, swimming, winter sports, chess, horse racing, women's sport, friendship, grief, sleep, fitness, diet, pets, travel, hobbies, faith, disability. About fifty of a hundred and thirty. A list is the only way to know that.

Important whatever the topic: three signals, no model

The sources: what already aggregates the world

Nobody has to search for the span; several services already gather it, and the useful ones are free. The commercial ones sell the same search shape we have now over a wider index, which fixes the outlet problem and not the query problem.

Gone or closed: Bing's news API (retired 2025), Ground News (no public API). Chinese one-liners come from Bocha and the Chinese outlets' own pages; a Chinese story only Chinese media carried arrives as a Chinese headline, which the meeting reads.

Speed and cost, a day

The current pegs stage on 09-09 took 90 seconds and cost 0.147 USD, all Serper fees; no model gathers pegs. The redesign is faster and about as cheap, because everything wide is free.

Three to five minutes of wall time, 0.06 to 0.10 USD. Time is invisible: the text pass starts at 03:00 and the meeting waits for the harvest. GDELT throttles bursts (a 429), so it runs spaced and the harvest proceeds on the other layers when it is slow, as the peg ladder already falls through to Wayback. The box in Singapore reaches all of these directly.

Tags from birth: what a picker and a recommender will need

A published piece already carries a beat and the free tags the meeting invents that morning (russia · ukraine · naval-war), and the app records each reader's verdicts and how far they read. The reading side of a recommender exists; the supply side does not: sixteen pieces from a patch, tagged in words that change from day to day.

The wide harvest, as a picture

Green is counting and rules, auditable from the files; blue is the cheap model, reading as a persona or picking from a shortlist with thinking off. The owner's rule (09-09 evening): ten minutes for this step means the design is wrong. The one giant call (154 pegs · the pitches · 90 roster lines · the policy → 20 commissions with questions, thinking on) starved on every model; the meeting is now a shortlist made in code, one fast pick, twenty small question calls and repairs in code, and the door stays open through the ten unpitched stories by attention. Same inputs, 24 seconds instead of nine minutes, no problems left. What the owner's read of 09-09 found: the pitch door held; the assignment door let a name pun, a brand and a register through, because the validator only checks that a reason exists.

i · The editor's match-making, kept for the holes

The meeting stays, but it stops being the only door. It is the right tool when the day needs a story and no one wanted it: the war on its fourth day, the budget, the thing that happened at 5 a.m. It gets the fixes from the study: the bench first, a reason that cites a line of the profile, no professional adjacency, a quota on how many of the day's slots it may fill by assignment. What it loses is the job of inventing a writer's interest, which is the job it did worst.

ii · The persona's spontaneous motivation, the pitch

The pitch pass. One cheap call per persona (Flash-class, ~4k tokens in, ~200 out). The prompt is the day's headlines first, as the shared, cached prefix, then the persona's motivation sheet, then one instruction: read the morning's paper as yourself; if something in it is yours, say which, and why, in two sentences, and in what form; if nothing is, say so. Silence is a valid answer and most days most personas are silent. Cost at 92 personas: roughly 0.1 to 0.2 USD a day on DeepSeek Flash with the prefix cached, less than one piece's writer call.

What a pitch is. Three lines, machine-readable: the peg (by id), the reason (which must name something in the profile: a memory, a value, a place, a person), the form. Two examples of what the pass should produce, written by hand here:

The motivation sheet. The pass needs something the profile does not have. The card's writes: is the language a persona writes in, not its subjects; nothing on the card says what a persona would want to write about this month. The sheet is a derived block, ~300 words, made once per persona from Values & seams and Specific memories and refreshed when the profile changes: the five or six things this person cannot walk past, each tied to the profile line it comes from, plus the things they are following (a persona may follow a story across days, which is how series happen). It is not authored by hand for 92 personas; it is extracted by one call each and cached beside the card, and it is the only thing the pitch pass reads besides the headlines.

What still needs an editor

5 · What this changes upstream