The brief line — the news brief is the product design · 2026-09-22

Short answer: Ink stops asking a persona to write. Each morning the scope's harvest is read, the day's stories are picked by the brief's shares, and every picked story becomes a block brief — the gist, keep reading, the questions, the numbers, the people, what came before, still unknown — on the list as a news item. No essay, no meeting, no rival reader, no art. About 3.7 cents a brief, about 50 cents a morning of twelve, off-peak; the essay line it replaces spent more than that on one piece.

Now we change the Ink design. We will no longer have the persona to write, for now; the brief is the end product. Describe the brief production line and mark the cost. — the owner, 2026-09-22

1 · The line

THE MORNING (03:00 CST, off-peak) cost · per morning of 12 1 HARVEST the brief's searches (Google News RSS, 20 a query, zh edition for zh) $0 search + the broad themes' SECTIONS read whole + the wild-card front pages $0 feeds → ~1,000 stories, ads out, folded (one story, many outlets) code 2 THE READ every story judged against the brief: in? theme · ● ●● ●●● · beat ~$0.05 Flash, batched → the pool, seated by the brief's shares (floor a quarter of equal) code 3 THE PICK N stories (the console's count): per-theme cap ceil(N × share) + 1, ●●● first by attention, one wild card if the brief reserves one, the same story never twice code $0 │ ▼ per story, 4 in parallel 4 THE DESK fetch the story (the full-text ladder), 2 searches, 2–4 pages read, ONE Flash distillation → the notebook: facts · numbers · people · quotes · disputed · unknown, each with its source $0.006 ($0.004 model + $0.002 search) 5 THE QUESTIONS ONE call, two lists: up to six the story itself raises + up to three 「what does this mean for me」only when it touches money, home, health, safety (the union, under nine) + four canned intents, never printed as labels (precedent · numbers · critics · mechanism) $0.001 6 THE LOOK-UP one search a question, 12 pages read once, notes per page, ONE answer a question merged $0.028 ($0.018 model + $0.010 search) 7 THE BLOCKS ONE Flash call reads the notebook + the look-up → the JSON page $0.002 → the LINTS in code: every figure in the records · digits in the gist · no「the records do not say」· a timeline line is an event · no aggregator among the sources; one re-ask on a fail +$0.002 sometimes 8 THE DEK one line under the title: what the headline leaves out, under 15 words; re-asked once when it reuses a third of the headline $0.001 9 PUBLISH the list item {kind: "brief", field, title, dek} + the page record (pegs.json, brief.en.blocks); the edition flips; yesterday's stays code $0 ONE BRIEF ≈ $0.037 · TWELVE ≈ $0.44 + the morning's shared read ≈ $0.05 A MORNING ≈ $0.50 off-peak · ≈ $0.90 at peak

Every figure above is measured, not estimated: the three real-estate briefs built from scratch on 09-22 (desk + look-up + blocks + dek, the dev call ledger by leaf) and the 09-22 scope harvest's read. The search fees are Serper's ($0.001 a query). Peak = DeepSeek's 09–12 and 14–18 CST at twice the model price; the search fees do not move.

2 · Where the money goes

StepModelSearchPer briefShareThe dial
the desk notebook$0.004$0.002$0.00616%none — the notebook is the floor; a brief without it has nothing to say
the questions + the look-up$0.019$0.010$0.02978%the one dial. Six questions + four intents = 10 searches, 12 pages. Four questions + two intents ≈ $0.018; no look-up = a notebook-only brief under a cent, and no questions block
the blocks + the lints + the dek$0.002—$0.0026%none worth turning
one brief$0.025$0.012$0.037≈ $0.06 at peak
the morning's read (shared)$0.05$0≈ $0.004 at twelvegrows with the pool, not with the count of briefs

For comparison, the essay line: the Nscale piece alone was $0.11 of writing and $0.017 of art, after a meeting, a pitch round and a rival reader that a brief never needs. The brief line writes twelve pages for what one essay cost.

3 · What the old morning did that this one does not

Stage of the essay brewIn the brief line
the pitches — every persona reads the pool and pitchesgone; the pick is code, by the brief's shares
the meeting and the slate — the editor matches stories to writers, the match gategone; a story needs no writer
the archive — prior coverage, the precedent, the study, the numbers over yearsfolded into the look-up's canned intents (precedent · numbers · critics · mechanism)
the persona drives — reaction · life · angle · point · order · draft · mend · rereadgone
the rival reader and the mends, the rereadgone; the lints in code take the seat
the renditions (en ⇄ zh)open: a zh reader today gets the English blocks; one more Flash call a brief (≈ $0.002) would render them in Chinese — decide when a zh reader asks
the artoff (owner: not yet); $0.017 an image when it returns
the gate and the fine printgone; a brief that fails its lints twice is dropped, the morning says so in its log
the Background page behind an essayis the product — opened from the list item, the field in its head

4 · Built and owed

PieceStateWhere
the scope: the brief → searches, sections, shares, wild cardon the boxthe scope · ink_harvest.harvest_brief
the harvest + the read + the seats by shareon the boxink_harvest.judge_pool · select
the desk notebookon the box (the essay line's research)ink_brew.research
the look-upon the box (the essay line's; today given the persona's questions)ink_brew.look_up
the reader's questionsbuilt 09-22ink_brew.reader_questions
the blocks call + the five lints + the re-askbuilt 09-22ink_brew.brief_blocks · brief_lint (the prompt is BRIEF_SYS); a brief whose page is unusable after the re-ask is dropped and the log says so
the pickbuilt 09-22ink_brew.line_pick: the brief's shares walked (weighted_order), ●●● before ●●, never ●, a dry theme's seat handed on, the wild card's reserve, a story burned for three days after it is briefed, the same headline never twice
the dek · the list item · the page recordon mainink_brew.brief_dek · brief_item · peg_card
the list, the page, the bubbles, the text sizeon main (v1111–v1116)the Background page
the stage — pegs → the line (four stories at a time) → publish, one passbuilt 09-22, run on devink_brew.run_line · stage_line · line_one · publish_line; the CLI python lib/ink_brew.py line --n 12; the console's Brew now and the dawn slot run it while「the brief line」is on (the evening slot has nothing to do — the news would be stale by morning)
the zh rendition of the blocksbuilt 09-22 (owner: yes for Chinese)ink_brew.blocks_zh, one call a brief on the rendition seat; a Chinese reader gets the zh title on the list and the zh blocks on the page
the consolebuilt 09-22Ink tab: the switch「the brief line」(knob line: briefs | essays, default briefs) · Usage & cost ▸ the day's cost groups lead with the line's four (the harvest · the desk · the look-up · the brief) · Models & APIs: the seats' copy says what each does under the line and its measured price
the same story twice from two outletsowedthe pick refuses a headline that repeats a picked one by half its words; two rewrites of one Reuters story with different words still both pass (seen 09-22: two Anthropic-model stories) — the fix is a check after the desk on the notebooks' one-line, before the look-up spends

5 · Status

DateWhat
2026-09-30 · the shelf's memorythe owner, reading the box's 09-30 edition:「there are many repeating topics compared to yesterday」. Short answer: 7 of the 16 briefs had run on 09-29 — Instinct, AMD–World Labs, Starship, the Shanghai presale rules, IIT Madras (its third morning running), the microblazar, SiMa.ai. Cause: the pick's three-day guard burned URLs, and yesterday's story comes back today from another outlet (Instinct: Yahoo Finance → Bloomberg · AMD: MarketWatch → WSJ · Starship: Reuters → the Guardian), so no URL ever matched. The same-story look of 09-23 only compares a morning's candidates with each other. Fix: the pick now reads what the shelf already told the reader — the briefs of the last three published editions (recent_briefed, the record itself) — and drops a candidate that reports the same event: the word test on titles always, and one look (story_repeats, Flash, ≈0.1¢ a morning) for what words cannot see (「Cheers go up as huge rocket circles Earth」 = Starship reaches orbit). Event, not subject: a new development in a running story stays — yesterday's Anthropic brief was the Trump dinner, today's IPO filing is news. Replayed on the box's own 09-30 pool (152 stories, the ledger as it stood before that morning): word test alone — 6 of yesterday's stories picked again; with the look — 0, their seats going to EliseAI, OpenAI's funding round, Marissa Mayer's Dazzle, rogue black holes, the fruit-fly study. The look marked 26 of the top 80 candidates as already briefed; read one by one, all 26 were true repeats (three of them of 09-27 and 09-28 briefs). Fails open: a dead look leaves the word test.
2026-09-30 · …and its two holes, the same eveningA dry run of the 10-01 pick (the new scope's harvest, below) showed the guard leaking twice. ① The siblings: the look marked the Guardian's line on Anthropic's IPO filing, and the pick took Unite.AI's line on the same filing — one story, many headlines, one of them marked. A repeat now takes its whole story with it: its same-day group (the groups are read on the pool before the repeats leave) and any headline the word test pairs with it. ② The reach: the look read the top 80 of 160 candidates; a second Larry Ellison story at rank 101 walked through, and the word test missed it because a brief's title is short (_told_before: nearly all of the shorter title's words, at least four). The look now reads the whole pool, forty at a time (four small calls, ≈0.2¢; shorter lists also read truer). Re-run twice: the same 16 both times, none of them a story the shelf had told; the Shanghai seat went to a real follow-up (「新政落地24小时」— the land bids the day after the rules). (A repeat's group first left with it whole — and the groups are by subject, so a real development would have been lost; see the next row.)
2026-09-30 · a sibling is judged: said again, or a developmentthe owner:「if the LLM is smart we can distinguish a news said again from a development」. Right — dropping a repeat's whole group was a shortcut. Now only a near-identical headline (the word test) leaves outright; a sibling by group is read against the brief we ran, its summary line included (story_developments): the same news again leaves, a development — something that happened after the brief — stays. How the judge was made to hold still, measured on seven real pairs with known answers, five runs each: ① a bare verdict in one list — it flipped on the hard pairs (「Starship set to fly」→「reaches orbit」read as the same news twice in three); ② made to write the new fact and whether it happened after, with thinking on — six of seven stable; ③ given the brief's summary line — all seven stable, one wrong every time; ④ one pair a call — seven of seven, five runs of five (in a list the Shanghai land bids were「the same news」; alone, the judge named the new fact —「within 24 hours China Merchants won two plots at a 20% premium」— and kept it). Siblings are few (four on the 10-01 dry run), ≈0.1¢ each. On the dry-run pool: the two Anthropic IPO siblings leave both runs; one borderline headline (国常会 + the Shanghai rules + the stocks' reaction — half new, half told) went each way once. An unjudged sibling (a dead call) leaves.
2026-09-22 · a weak answer, tracedthe owner:「What is driving SF prices so far above asking?」came back as three Compass figures — the good answer touches the inventory and the AI rich. Traced: both WERE in the look-up's notes (the Guardian's AI boom, CBS's agent on AI pay, a buyer's guide on scarce single-family inventory); the merge kept one page's figures under its 130-word cap and dropped the other two, then the blocks rule「one fact with its figure」dropped the causes that had no number. Fixed at both seams: a WHY question is merged causes-first from every page, figures after; on the page a why question gets one cause per bullet, named. The mortgage brief re-run.
2026-09-22 · deployedthe owner:「I think we can now deploy」→ the box at v1130: the line, the company, the pictures. The box's line knob reads briefs by default from here. Owed on the box: OPENROUTER_API_KEY in its .env — until then Jev is unreachable there and the seating falls to the tag shortlist's top two.
2026-09-23 · Ink is users onlythe owner asked for the stories to be viewable by all users with the ⋯ (the Scope, the harvest) admin-only — which was already so: the Ink tab and every story surface have no role gate, the ⋯ paints only for an admin, and its data rides /api/admin/ink/*, refused to any other role. The one gap found on the way: the Ink files under /ink/ (editions, the briefs' records, comments, pictures) were static and served to anyone, signed in or not — the session gate covered only /api/. The owner:「ink is users only」→ the gate now covers /ink/ too; an anonymous fetch gets 401, a signed-in user 200.
2026-09-23 · the same story never twice — the lookthe owner, reading the box's 09-23 edition:「the recent stories have 2 Heidi and 2 Muse — what has caused the duplicate?」The pick's only duplicate test was a word test on headlines (more than half of the longer headline's words in the other). Two outlets rarely phrase one story alike: Bloomberg's「AI Startup Heidi Doubles Valuation to $900 Million」and SmartCompany's「Heidi raises $475 million, hits $1.26 billion valuation」share four words (score 0.44, the bar is 0.5) — one funding round quoted in US and Australian dollars, so the two briefs even contradict each other on the figures; the NYT's first-person Muse review and Yahoo's JPMorgan note share one word (0.08). The three-day burn is by URL and could not help. Now ONE look before the pick: Flash reads the pool's top 80 headlines (site, snippet) and groups the same story or the same subject — a round in two currencies, one product from a review and an analyst note; the pick takes the strongest peg per group and refills the seat. Run on the box's own 158 pegs: 8 groups, both Heidi pegs in one, all seven Muse pegs in one, plus Opus 5.5 ×2, Chamelio ×2, the fire amoeba ×2, New World ×2, Iambic ×2, the S.F. bidding wars ×2 — for 0.08¢. No key or a failed look → the word test alone, as before. Owed: when a group carries two currencies, the desk should say which one the brief quotes.
2026-09-23 · the picture, under the title (v1131)the owner:「the hero picture is not under the title and the dek — move it so that it's in line with previous Ink articles」. The essays' lead card runs title · dek · picture; the brief page had the picture as a mast above the title. Now the brief page runs title · dek (newly shown, the lead card's italic) · date and language pill · picture with its credit · Comments · the gist. Still edge to edge, still unclipped.
2026-09-22 · the meter's clockthe owner:「make sure the cost numbers are accurate」. A recompute of the day's ledger against the price sheet disagreed by a quarter. Traced: the brew stamps each call with Shanghai time and the meter's peak test read that stamp on the UTC clock — so a 15:00 CST pass (07:00 UTC, peak) was booked off-peak, and the box's 03:00 CST dawn brew (19:00 UTC, off-peak) was booked at the peak rate, since 09-13 when calls began to be priced as of their own hour. Fixed at the seam (cost.is_peak and the dated rows normalise an aware time to UTC; the ledger row now carries the call's own hour), a smoketest check pins 03:00 CST off-peak and 10:00 CST peak, and scripts/reprice_ledger.py re-derives booked rows from their tokens and hour — run on dev (5,700 Ink rows since 09-13: $6.51 booked → $8.34 true; the dev brews ran in CST afternoons, i.e. at peak). The box owes the same run after deploy; its dawn rows will come DOWN. The figures on this page stand: they are rate-based (off-peak sheet), not clock-based. Jev's rows are booked flat from OpenRouter's own cost line and were right — a probe shows that line is exactly input tokens × $0.042/M, output free — and since 09-22 they also carry their token counts, so the console's token columns no longer read zero for Jev. Jev also has its row on the console's price sheet (Models & APIs › prices, under TypeSafe): $0.042/M on every input column, output 0 — the sheet OpenRouter's line has matched, and the brew's fallback rate when a response carries none.
2026-09-22 · the story's own picture (v1128)the owner:「50 cents for 12 personalized stories plus the masters — the only thing missing is a good hero picture, but at 3¢ it carries little info」→「use the source picture if available and meaningful」. The outlet's photograph is free and IS the story: the peg page's og:image (then the first fetched sources'), pulled once through the unfurl guard (the reader never hotlinks the origin), the code drops logos, icons, strips and tiny files, and ONE look by the eyes (V4.1 Flash, multimodal, ≈0.1¢ — Jev reads text only, so it cannot judge a picture) says whether it is OF the story or stock art, a logo, a reporter's headshot. A brief with no usable picture has none. It rides the list row and the top of the brief page, credited「Photo: CBS News」. Owed: a per-outlet opt-in before a launch (rights).
2026-09-22 · the questions' unionthe owner read the crater's「who pays to protect NASA's moon base?」→ a comp on four briefs (the checklist A · 「after reading this」on the notebook B · the same on the finished brief C, docs/ink-background.html): A wins on the reader's own stakes, B and C on what the story raises, B ≈ C. The owner:「A is worth keeping — how does this impact me; B and C are very good — the union, cheaply」→ ONE call returns two lists (the story's six, the reader's three only when the story touches money, home, health or safety), the look-up's page budget holds (≈ +0.3¢ a brief), and the four canned intents never reach the page as their labels — a found fact is written under the question a reader would have asked. The crater and the mortgage briefs re-run on dev.
2026-09-22 · built on localthe whole line in ink_brew (the pick · the desk · the questions · the look-up · the blocks with five lints and one re-ask · the zh blocks · the dek · the publish), the console's switch, the seats' copy, the smoketest's three checks (the lints, the pick, the wiring); the first dev morning for 09-22 run at 12 under the dev scope. Not on the box.
2026-09-22 · the designthe owner's turn:「the brief is the end product」after reading twelve on the dev server. The line above; the cost measured on the 09-22 set. Next: the build of the stage, then a dev morning under the box's scope for the owner's read, then the box.
design 2026-09-22 · siblings: the Background page · the scope · the brew (the essay line)