Short answer: an admin writes what they want to read, in their own words; the desk turns it into ticked sub-fields and a few custom searches; the admin confirms; from the next morning the brew reads only that. One scope, shared by every admin. Clear it and Ink reads the whole world again.
Every brew keeps the whole harvest of its morning as pegs-all.json: about 2,400 stories a day, from Wikipedia's Current events, 37 Google News country editions, six topic feeds in five editions, 59 outlet feeds, and 189 free Google News searches (172 sub-field rows + 17 kinds of event). Ten days of pools were read for this page: six on the box (09-09 → 09-13, 09-20) and four on localhost.
The taxonomy a story is born with. One tree and three flat axes:
The checklist's eight「World › Africa / Latin America / …」rows are regions dressed as sub-fields; the scope treats them as regions and never as a topic, so the topical list is 164. The free-text tags an editor writes on a commissioned piece (["buffett", "succession"]) are a different thing — subject words for the repeat check, not a taxonomy.
| Fact | Measured | What it forced |
|---|---|---|
| half the pool carries no sub-field | 47% — every story that came from an edition, an outlet feed or a topic feed has a beat only, and an edition's beat is「World」by default | a filter on today's tags would see half the pool, and the personas write mostly from the other half: 63% of the 182 pieces commissioned on the box came from a peg with no sub-field |
| the sub-field counts are a quota | every checklist search returns 8 lines, so all 172 rows sit at 7–8 a day; only history (26) and obituaries (12) differ, because Wikipedia feeds them too | a distribution read off the raw tags is flat by construction — the pool has to be tagged before it can be read |
| Pass | How | Cost | Recovered (7 days, 16,704 stories) |
|---|---|---|---|
| ① the harvest's own tags | a story a checklist search found carries that row's sub-field from birth; the fold merges it onto the same story seen in a feed | free | 51% |
| ② keyword rules | each checklist row is a query a OR b OR c; split on OR, match every phrase as a whole word against title + snippet; a hit gives beat and sub-field at once | free | 13% |
| ③ one Flash pass | the rest, 40 headlines a call, the closed list in the system prompt; one sub-field and one region each, thinking off, JSON mode | $0.25 for the seven days · ~3.5¢ a morning | 36% |
| still empty | 1% | ||
Region: 43% from the harvest (a named place in the title, an edition's country, a regional checklist row), 32% from the Flash pass, 25% genuinely placeless — science, tech and life pieces mostly name no country.
The raw tags said World 29%: edition inflation. Tagged, World is 22%.
With the untagged half tagged, the range runs from 6 to 76 a day. A sub-field near 8/day is one only the checklist search reaches (90–100% of its stories came from that search); one well above it is what the world's editions and feeds push on their own.
| The world pushes these | /day | from the search | The search-only tail (~8/day, 90–100% from the search) |
|---|---|---|---|
| crime and trials | 76 | 9% | sleep · sex · grief · disability · faith and meaning · poetry · language · art crime · architecture · esports · quantum · drugs and addiction · private equity · shipping and logistics · philanthropy · winter sports · women's sport · cycling · the diaspora · Chinese sport |
| elections | 65 | 11% | |
| diplomacy and summits | 57 | 13% | |
| public health | 56 | 13% | |
| war and ceasefires | 42 | 18% | |
| celebrity · military and defence · disasters | 38 | 20% | |
| history · medicine · courts and rulings | 32–37 | 22–71% | |
| football (world) · pop music · cricket · schools · film | 25–29 | 26–31% |
Per beat, the heads: World → crime, elections, diplomacy · Science → public health, medicine, astronomy · Culture → celebrity, history, pop music · Markets → airlines, commodities, economics research · Tech → AI models, AI in use, gadgets · Sports → world football, cricket, American football · Life → schools, housing, travel · China → policy, trade, Taiwan. Beat × region: World is spread across every region (Middle East 65/day, Western Europe 56, Southeast Asia 47); every other beat is North America plus placeless.
The owner's own example: start-ups, tech, science, Bay Area real estate, Shanghai real estate, super-luxury real estate worldwide.
| Wanted | On the tree | Expected pool |
|---|---|---|
| start-ups | Markets › startups (+ IPOs and deals, private equity) | ~30/day |
| tech | the whole Tech & AI beat, 16 sub-fields | ~200/day |
| science | the whole Science beat, 20 sub-fields | ~310/day |
| Bay Area real estate | no fit — property × North America is a continent | — |
| Shanghai real estate | no fit — property × China is a country | — |
| super-luxury real estate, worldwide | no fit — property has no price tier; luxury and retail is shops | — |
The taxonomy's finest grain is a sub-field × a region, and a region is a continent. Cities and price tiers are below it. But a sub-field is nothing more than a name and a Google News query, and the search layer runs every row every day for free — so the scope can add rows:
A custom row behaves exactly like the built-in 164: the harvest searches it (a zh row in the Chinese edition as well), the fold tags stories with it, the Flash tagger sees it in the closed list so feed and edition stories can land in it too, and it is ticked. Each brings in roughly 8 to 20 stories a day, mostly from the search itself.
| Piece | Where | What |
|---|---|---|
| the store | settings.ink_scope | {text, subs: ["Beat › sub"…], custom: [{beat, sub, query, lang}], regions, updated, by} — empty or missing = no scope |
| the routes | run_room.py · /api/admin/ink/scope (GET · POST) · /scope/draft | under /api/admin/, so the central gate makes them admin-only; the console's /api/admin/ink carries the scope read-only |
| the draft + the tagger | ink_brew.py · scope() · clean_scope() · scope_draft() · tag_batch() | the eighteenth task leaf, app.ink_scope, on the Flash seat; a card on Models & APIs under Research |
| the harvest | ink_harvest.py · taxonomy() · scope_subs() · in_scope() · keyword_rules() · tag_pool() | model-free: the brew hands in the Flash call; without one the pass stops after the keyword rules. harvest(…, scope, tagger); the checklist rows carry a language |
| the page | room-ui.html · #inkScope · openScope | in DRAWERS, the back gesture, the wide-mode list; the ⊕ menu's first item on the Ink tab for admins; the reader's ⋯ after Feedback… |
| the console | Ink tab › The scope | read-only: what is set, by whom, when — it is set from the app |
pegs-all.json is the full, now fully tagged, pool. Widening the scope tomorrow costs nothing today; the tags are the record.pegs-all.json · pegs.json · coverage.json); /api/admin/ink/harvests + /harvest/{date}.The owner:「our harvest should be different. It's no longer pull all the sources and sort out what I need. It should be proactive search on my interested topic. Don't translate my text into tags from the existing database; do fuzzy and semantic search using my text on the web for news.」
As built (v1103). The page: your brief · Read it back (one Flash call: the readback line, the themes in your words, the number of searches) · Confirm / Clear · the harvest rows. The store: ink_scope = {text, summary, themes, queries:[{q, lang, theme}], updated, by}. The morning: ink_brew.stage_pegs → reading_scope() → ink_harvest.harvest_brief: every query searched (Google News RSS, 20 lines each, the Chinese edition for a zh query) → fold → judge_pool (the Flash seat reads every story against the brief: in · theme · closeness 0–3 · beat, 40 a call) → the stories in, ranked by (1 + closeness) × attention with the usual discounts → ~160 to the meeting; yesterday's pegs with closeness ≥ 2 → up to 10 follow-up queries (followup_queries). The harvest page shows ● ●● ●●● per story. The pick list, the tag recovery and the custom rows of v1101 are gone from the code; §2–3 stay as the study they were.
| The dry run, 09-21, the owner's brief | |
|---|---|
| queries drafted (6 themes: start-ups · tech · science · Bay Area · Shanghai · super-luxury; 12 in Chinese) | 45 |
| lines → stories after the fold | 772 → 641 |
| in scope after the read (●●● 110 · ●● 291 · ● 30) · out | 431 · 210 |
| by theme: tech 137 · start-ups 109 · science 62 · super-luxury 53 · Bay Area 38 · Shanghai 32 | |
| to the meeting · very close | 160 · 104 |
| time · cost (18 Flash calls) | 15 s · $0.016 |
The owner, reading the parsed searches:「biotech is not mentioned under tech. How can we solve this?」— and then:「I'm no expert in news. I just like the search to be reasonably well covered, not skewed towards a subset of the news space. I want the user to get what he wants with best effort.」
A finite list of phrases can never cover a broad theme; that was the tags' problem too. So Parse now decides, per theme, broad or narrow:
| What it is | What the morning does | On the Scope page | |
|---|---|---|---|
| broad | a whole section of the news — tech, science, health, sports, business, politics, world, culture, markets, environment, real estate, education, law and crime, China, lifestyle | reads the section WHOLE: every outlet feed it has, its Google News topic feed in two editions, its standing checklist searches — exhaustive by construction; the relevance read still decides what is close. 2–4 queries for the editor's particular angle, if the brief names one | 「tech · the whole technology section · 8 feeds · 17 standing searches」 |
| narrow | a place, a price tier, a company, a niche | enumerated: 10–15 searches covering the theme's sub-topics, players, places and phrasings | the searches, compact |
The sections map (lib/ink/harvest.json › sections, 15 sections, 92 feeds of which 33 new, 173 standing searches) is a second key over the same sources, by ordinary section names. It does not replace the eight beats: those stay the writing side's labels (hard news, the art director's accents). The eight were a lopsided map for reading — World and Markets were catch-alls, Health was split between Science and Life, Environment and Real estate had no home — so reading broad themes is keyed to sections instead. New feeds: Politico, The Hill, Guardian politics/law/environment/health/education/property, STAT, Medical Xpress, WHO, KFF, Nature, Science, Carbon Brief, Inside Climate News, Mongabay, The Real Deal, Realtor.com, Mansion Global, Inside Higher Ed, Chalkbeat, Times Higher Education, Caixin, China Daily, Courthouse News, ABA Journal, NYT style/travel/politics. A feed that fails is logged and skipped.
| The dry run, 09-21, the owner's brief, with sections | |
|---|---|
| Parse: tech → the technology section, science → the science section; start-ups, Bay Area, Shanghai, super-luxury enumerated at 15 searches each | 68 searches + 2 sections |
| lines: 68 searches 1,116 · technology whole 351 (26/26 sources) · science whole 368 (26/27) | 1,835 |
| stories after the fold · in scope after the read (●●● 220 · ●● 553 · ● 57) | 1,210 · 830 |
| by theme: tech 295 · science 185 · start-ups 155 · super-luxury 90 · Shanghai 65 · Bay Area 40 | |
| to the meeting · time · cost (32 Flash calls) | 160 · 33 s · $0.03 |
| Seen | Cause | Fix |
|---|---|---|
| the Shanghai theme was thirty developer sales-office notices from one portal, rated ●●● by the read | a listing is squarely「Shanghai real estate」to a reader that was not told ads are not news | ad-like lines are dropped before the fold (售楼处 · 官方专线 · 24小时响应 · sponsored · press release …, logged: 74 of 1,912) and the read is told a listing, a promotion or a press release is 0 however close its subject |
| after that, the Shanghai theme had twelve stories in scope and none at the meeting | three doors in a row: the tail seats handed each theme ONE story and stopped; the language limit (12 non-English) was the wide harvest's rule; and the brew's section-page guard counted a Chinese headline as one「word」and refused all twelve | the tail cycles the themes until its seats are used; under a brief a third of the seats may be non-English; the guard counts CJK characters as words |
| 「实得率超216%!杨浦滨江「外滩元境二期」推约110-174㎡风貌洋房」reached the meeting at ●●● — a soft ad: a launch report about one named development | it looks like news; the read's ad rule named listings and hotlines, not advertorial | two signals before the fold — a bracketed project name AND a sales term (实得率 · 推约 · 户型 · 均价 · 认筹 · 加推 · ㎡ …) — so a market report quoting prices survives; and the read is told the soft ad's tells: one named development quoting its own sales terms with no market fact and no event is 0, and a portal or self-media account is news only with a market fact from a source other than the seller |
| a theme's LPR / PBOC rate stories rated ●● for Shanghai real estate | the read is lenient on adjacency | left as is: rates move property; the meeting can still pass them over |
The owner:「will the distribution of the stories affect the distribution of the articles?」— not the pool, but the slate validator did, with the wide harvest's rules in beats: at most 4 pieces a beat, at least 6 beats covered. Three real-estate themes share the Markets beat, so together they could never pass 4 of 16; and a brief spanning four beats tripped「6 beats covered」every morning. Then:「is it a good idea to put hard code numbers?」and「I never said the total is 16」.
| Rule | Wide harvest (the world) | Under a scope |
|---|---|---|
| spread | at least STYLE.beats_min beats covered, no beat above STYLE.beat_max | every theme that has pegs at the meeting gets at least one piece; no theme above ceil(pieces ÷ themes) + 1 — computed from the morning's piece count (the console's knob) and the themes with pegs: 16 over 6 → 4, 8 over 6 → 3, 30 over 6 → 6 |
| the front page | STYLE.front_min of the pieces on a front peg when the harvest offers STYLE.front_offered | off — a scoped pool has no front page |
| the shape of a day | house style, named once in STYLE and read everywhere: light pieces 1–2, the kicker last, plain essays at most 6, Fable · Shouts · Overheard · Riddle once each. No bare literal in the validator, the trim or the policy text. | |
The first scoped meeting (09-22 on dev) showed the rule alone is not enough. The validator named the problems correctly — Bay Area and Shanghai had no piece, tech had five — but the editor could not mend them: the desk had reported the top 32 stories by attention (mostly tech), and the shortlist's open door held the top 10 by attention (all tech and science); two themes never reached the editor. So the same spread runs one level down too: under a scope the desk reports and the open door offers pegs round after round over the themes, each theme's own list by closeness then attention (spread_by_theme), the open door holding each theme's fair share (the cap × the themes). On the sample: the desk 6·6·5·5·5·5, the open door 4 a theme.
The second scoped meeting showed the last limit, and it is not code. With the desk and the open door spread, the editor commissioned 15: tech 6 → trimmed to the cap, science 3, super-luxury 3, start-ups 2, Shanghai 1 (Xiao Pang on the 28 August policy package). Bay Area real estate still got none: every pairing the editor tried was refused at the match gate, rightly —「a tech investor's rate calls are not the Bay Area mortgage market」,「his trade is startup financing, not residential real estate」. The cast has no one whose trade is property. A theme is covered only as far as the roster has a persona for it; the fix for that theme is a persona (the pipeline), not a rule. The cap now also holds over what the fills bring in after the trim.
The fourth run, with the fills reaching for the themes still owed: 14 pieces — tech 4 (the cap), science 3, super-luxury 3, start-ups 2, Bay Area 2 (Xiao Pang on the all-cash craze, Thomas Feddo on a waterfront data-centre site), Shanghai 0 this run and 1 the run before. Run to run the editor's pick varies; the rules hold the shape. The code fills (from the unused pitches, then the spares) now go to the themes with the fewest pieces first and never add a theme already at its cap, so a fill is never trimmed straight back off.
The first box morning under the brief (09-21) found one more door. Nine pieces, none on real estate — and Xiao Pang, whose trade is exactly Shanghai home-buying, HAD pitched the Bay Area mortgage story with a verbatim line from his profile,「焦虑的是持币的,不焦虑的是持房的」. The pitch check demands a quote of five words and counted the Chinese line as one; the only real-estate pitch of the morning was dropped as「under five words」. Fixed with the same rule as the section-page guard, now one helper: CJK counted by character, three to a word.
And one more, the same morning: Max Rockbank — a London super-prime property agent, built in the app's Studio — was never at the meeting at all: the brew's roster walked the repo's persona folder only, and an app-built persona lives in the second root (the app's registry). The roster now walks both roots, the repo winning a slug collision, as the app itself does.
topic_rules(n, pegs, scope) derives the theme rules; policy_text tells the editor the same rule in words; validate_slate and trim_slate enforce it. The pieces per register (heavy · mid · light) are computed from n too.
Parse reads a share per theme out of the brief's own words: an explicit figure is kept as a band split among the themes it names; 「more on」, 「mostly」, 「especially」 mean a larger share; the unweighted themes split the rest evenly; an emphasis inside a theme (「esp consumer internet and AI」) goes into that theme's queries, not its share. The readback says the shares back in plain words; the Scope page shows each theme's percent.
The shares then run the whole morning, where the equal share ran before: the meeting's seats are given out by share — each theme fills its own seats from its own stories, closest first, no theme below a quarter of an equal share, and seats a thin theme cannot fill go back to the pool (on the sample: 43 · 43 · 43 · 10 · 10 · 11 of 160, the real-estate trio at 19%); the desk and the shortlist's open door do the same; the code fills reach first for the theme furthest below its share; the slate's cap is per theme, ceil(pieces × share) + 1. On this brief Parse gave 27 · 27 · 27 · 6 · 6 · 7 — the real-estate trio at 19%, inside the band — so at 16 pieces the caps are 6 · 6 · 5 · 2 · 2 · 2, at 10 they are 4 · 4 · 4 · 2 · 2 · 2.
The owner:「what if I add "leave 1 article for wild card". Can it understand?」A wild card is the opposite of a theme: a piece from outside the brief, whatever the world put on its front pages that morning. Three small pieces:
| Parse | a reserve — 「leave 1 article for wild card」, 「one slot for any article you see fit」 → wild: 1; a PERCENT (「leave 5% for wildcard」) → wild_pct: 5, never a count, turned into pieces each morning from that morning's piece count (5% of 16 = 1, of 40 = 2; v1110) — kept apart from the shares (the shares still sum to 100 over the themes); the page shows a last line「wild card · 1 piece a morning · the world's front pages, outside the brief」 |
|---|---|
| the morning | beside the brief's searches, a small wide slice: the US and UK front pages and Wikipedia's Current events, tagged「wild card」and not read against the brief — the world's, by attention; it gets a quarter-share of seats per reserved piece |
| the slate | 「wild card」is a theme with a floor and a cap of the count: exactly that many pieces, from that slice, from whoever pitches or is matched; the code fills reach for it first until it is filled, then never |
| Date | What |
|---|---|
| 2026-09-30 · a theme for Chinese start-ups and AI | the owner:「our coverage for the Chinese start-ups and AI companies is lacking — check whether this is true and what caused it」. True: 4 of 48 start-up and tech briefs in six editions touched China, one of them a Chinese-sourced story about a Chinese company. Why, on the 09-30 harvest: ① the scope barely asked — 2 of 19 start-up/tech searches in Chinese, none on Chinese AI, and generic (「创业公司 融资」returns Chinese rewrites of US news); ② what it did find never reached the meeting — 27 relevant Chinese start-up stories, 0 of the theme's 42 seats. The cut inside a theme is not「the closest first」as §9 says: below 40 seats a theme is filled from the tail, by outlets, then newest, then input order — a Chinese source is always one outlet and lost the tie to ~100 same-day English one-outlet lines. (The 0.4 language discount touches only the seats above 40 — two that morning; it was first blamed and was not the cause.) The fix is the owner's, in the Scope: a theme of its own —「Chinese start-ups and Chinese AI companies」, 15, fifteen searches all in Chinese and all 中国-led — the way Shanghai real estate has always worked: its own seats, its own language. Dry-run of the 10-01 harvest: 89 stories in scope for the theme, 27 seated, 19 of them Chinese (36氪 · 极客公园 · 财联社 · 证券之星 …); the pick took two — DeepSeek's Huawei chip tools (Bloomberg) and 芯宿科技's A round (Chinese). owed the tail honouring closeness (§9's own sentence) — small, all themes, not needed for this. |
| 2026-09-22 | v1110: the wild card as a percent too; three phrasings checked (「5%」→ percent, 「1 slot for any article you see fit」→ 1, 「2 articles for anything interesting」→ 2). |
| 2026-09-22 | v1109: §11 — the wild card, a reserve of pieces outside the brief from the world's front pages. |
| 2026-09-22 | v1108: §10 — the brief's own weights, a share per theme, carried through the tail seats, the desk, the open door, the fills and the per-theme cap. |
| 2026-09-22 | §9: the slate under a scope — theme rules derived from the piece count and the themes with pegs; the house style named once in STYLE; the front duty off under a scope. |
| 2026-09-22 | v1107 (owner): the harvest page's counts are a TABLE — the funnel per theme: in scope · meeting · written (written traced from the edition back through the slate's peg ids; — until published). The owner's question「will the distribution of stories affect the distribution of articles?」— not the pool, but the slate's beat rules do (max 4 a beat · 6 beats covered), which under a scope should become theme rules: OWED. |
| 2026-09-22 | v1106 (owner): the harvest page opens with a count strip — one tile a theme (in scope · seats), the Stats page's tiles; a tap filters the list to the theme. The soft-ad rule (§8). The owner's questions answered on the page: the coral tile = read at the meeting, grey = in the pool only, the ring = front; ●●● squarely · ●● within · ● loosely · out; attention = outlets + countries + 2×front pages + Wikipedia, no model. |
| 2026-09-21 | v1105 (owner:「reasonably well covered, not skewed」): §8 — the sections map, broad themes read whole, narrow themes enumerated; dry run 1,210 stories → 830 in → 160, 33 s, 3¢. |
| 2026-09-21 | v1104 (owner): the Ink ⋯ holds TWO commands, each its own page — Scope… (the brief · Parse · the readback line · every search, compact by theme, Chinese ones muted · Confirm / Clear; no status block) and Harvests… (the mornings as drawer rows, a row opens the day). The owner's question「what does 51 searches a morning mean」→ the number of queries the brief was parsed into, each run against Google News every morning, 4–8 per theme. |
| 2026-09-20 · night | v1103: the brief-driven search harvest BUILT (§7) — the page is the brief + the readback; under a scope the wide harvest does not run. Dry run on the owner's brief: 45 searches → 641 stories → 431 in → 160 to the meeting, 15 s, 1.6¢. v1102: the Ink tab's corner is the ⋯ (one command, admin only); the harvest rows + the harvest page. The pick list rejected by the owner. |
| 2026-09-20 | the ten pools read; the recovery run over 16,704 stories ($0.25); the distribution above. Built the same night: the store, the routes, the draft, the tagger, the harvest under a scope, the Scope page, the console line, the eighteenth leaf. Owed: the owner sets the first scope on the box tonight; the 09-21 03:00 brew is the first scoped morning — read its log line「scope: N of M stories in scope」and the by-sub counts in coverage.json. |