The brew — how Ink is made, nightlyproposal

The operating half of Ink, answering the owner's audit (2026-09-02):「do we have all the design decisions?」— the reader-facing half is decided (the essays · the look); this page proposes the four that weren't: the nightly pipeline, the slate policy (how many · which beats · which formats · who writes how often), where editions live, and the console's Ink tab. Recommendations are marked; the owner's calls are collected at the foot.

The decision audit

1 · The nightly pipeline

• Pegs: the eight triggers materialize as queries — per-beat news queries (velocity-diffed against the last 3 days), the calendar file (data days · anniversaries · seasons), reply pegs (essays that provoked rooms), demand pegs (room topics, once step 1.5 has rooms). Peg cards freeze from the SERP hit (the walled-site rule).
• Budget meeting: one V4 Pro thinking call (≥24k output cap — the starvation lesson) sees every peg and the roster with cadence state, and commissions N+2 pieces with dek, format, beat, well hint, and a runner-up per slot; a code validator checks the slate against the policy below and re-asks once, then repairs by swapping runner-ups.
• Writers: parallel V4 Pro thinking calls, full profile + the dealt hand of 2–4 unburned anchors + the event facts; the first-hand boundary prompt (v2 wording); per-language length floors. Gemini failover on a moderation refusal.
• Tone gate: the antithesis meter in both languages + the de-tell net; a failing piece is dropped and its slot filled from the two spares — the persona「didn't write today」.
• Renditions: persona-voiced translation into the app's other languages (zh first; the Flash tier — V4.1 Flash since 09-09 — ~0.3¢ a piece). No piece is born in Chinese (owner, 09-02) — reversed 09-03: a Chinese writer's piece is born in Chinese and rendered into English (§3, Language).
• Art: DeepSeek art-directs (style per piece, no likeness, no text), Gemini batch paints 16:9 (submitted by ~21:30, a ~9h window against a「within 24h」promise that usually clears overnight). The cutoff rule: collection starts half an hour before the art cutoff; at the cutoff any piece still without a hero has its batch request cancelled (batches.cancel) and is painted sync at full price — 16 pictures in parallel ×4 is ~2 minutes, worst case +70¢ on the day. If sync fails too (a Gemini outage), the composer demotes heroless pieces to compact rows and picks the lead among the pieces that have art — the edition still flips at 07:30. A brew never misses the deadline for art.
• Compose + publish: ink_compose assigns roles by the day seed; the edition is written as edition-YYYY-MM-DD.json with prev chained, then edition.json flips to it (atomic rename) — the threshold archive builds itself.

2 · Where editions live

3 · The slate policy — how many, of what, by whom

The mechanism is the house one: these are constraints in the editor's brief (so it composes toward them) and checks in code (so a miss is caught, not hoped against) — one re-ask, then runner-up swaps. All numbers are console knobs, not constants.

3b · The rhythm — heavy · mid · light, and the kicker

The proposal. Register becomes a third axis of every commission (beat · format · register), briefed to the editor and enforced by ink_compose as a harmony constraint: H6 never two heavies adjacent, H7 the kicker is light, H8 one light front-of-book. Four light formats join the nine:

• Shouts 谐 — the persona's satirical short, ≤400 words, one premise pushed to absurdity (the New Yorker's Shouts & Murmurs). Munger on a「diversified」lunch; Banksy's terms of service.
• The kicker 尾 — ≤150 words, one observation, one turn; pegged to the day at most. The last thing you read before the band.
• Overheard 闲话 — two personas, comic, a dialogue that goes nowhere on purpose; the dialogue format's light twin.
• The riddle 谜 — a persona sets a puzzle from its own world; the door is play it with them — 20 Questions on the game device, which already exists. NYT's Play, with a host.

3c · Topics — coarse beats, fine tags

The proposal: keep the 8 beats closed and coarse, and mint an open tag layer now. The beat is the editorial constraint (what a day must cover) and the reader's rubric (World · Essay); it can split later — when a beat exceeds ~30% of a month's pieces, it forks (Tech & AI → Tech · AI). Tags are the fine layer: 2–3 per piece, chosen by the editor from a house vocabulary file (lib/ink/tags.json) that grows — a new tag is allowed but must be normalized (Fed, not Federal Reserve and the Fed), and a tag becomes reader-visible only once three pieces carry it. Tags cost nothing at brew time and are expensive to add later (retro-tagging every edition), which is why they are minted from day one even though no reader sees them yet. They feed, in order: search · the Metrics tab (which subjects draw rooms) · follow-a-topic · personalization, when there are enough readers to personalize for. Coarse now, finer in the future — with the data already there.

3d · The anecdote bank — designed; the backfill

Solved on paper, not in code. The five mechanisms on the essays page (the used-anchor ledger · the anti-echo block · method over memoir · the corpus grows in-world · the memory bank outside profile.md) are designed and two are proven in the Vibes engine; none runs for essays yet, because the brew doesn't exist. They ship as stage 0 of the brew, not as a separate project.

The corpus is mostly there; the bank is not. The repo holds 88 personas, 85 with a corpus/; the 27 app-built personas have none. So the backfill is not a research harvest — it is a distillation: run the existing vibes_extract anchor-bank pass over every corpus into anchors.json beside the profile (30–60 dated, sourced anchors each, tagged marquee / minor / method), one V4 call per persona ≈ 1–2¢ → ≈ $1.50 for the whole library, an hour off-peak. The three corpus-less repo personas and the app-built ones get the builder's research step (a few SERP calls) or, for invented personas, a frozen invented ledger — made once, canonical forever, so the character never contradicts itself.

Then the bank feeds itself. Each commission deals 2–4 unburned anchors (≤1 marquee), the ledger burns them with a 90-day cool-down, and the writer stage runs 1–2 SERP calls around persona × peg and files any new sourced fact into the bank — the corpus grows by writing. The arithmetic: 40 anchors, ~1.5 pieces a week at 16/day over 88 personas, ~2 anchors a piece → a first full cycle takes ~3 months, by which time the bank has grown and the persona's own published record is material (「as I wrote in September, and was wrong about」).

3e · The persona remembers what it wrote — proposal

Yes — and half the machinery exists. Two stores already carry a persona's past: the live persona memory (docs/live-personas.html row ③, shipped v931 — what a persona remembers across chats) and the brew's ledger (what it published). The proposal joins them with one new artefact, the stance gist:

• At publish, the brew writes a 2–3 line gist per piece — what I argued · my position · the date · the peg — into the persona's memory as a kind of its own (wrote), global to the persona (an essay is public; every user meets the same author), never per-user.
• In chats, the megaprompt's memory block carries the persona's recent stances the way it carries recalled facts today, prefaced as its own record:「On 3 Sept I argued in Ink that…」— so a user who read the piece meets the same mind at the table, and a user who disagrees can argue with it.
• In future pieces, the writer's anti-echo block becomes a stance block: the last N gists, with the instruction「hold these positions unless the world has moved; if it has, say what changed you」— which is exactly the Scorecard format's fuel, and the 商榷 reply's.
• Decay — the owner's λ. Each gist carries a weight w = λ^Δdays/30 (λ ≈ 0.8: a position is at half strength after ~3 months, a fifth after ~7). The prompt ranks gists by w × relevance to the peg, carries those above ~0.2, and words the older ones as「positions I once argued and may have revised」. Nothing is deleted — the ledger keeps the record; only the prompt's grip loosens. The same weight rides into the chat block.
• What must be measured first: a prompt change needs a baseline (the house rule) — a small A/B on chat turns with and without the stance block, judged on continuity and on whether the persona argues its published line rather than parrots it.

3f · The editor — the lint, the rival model, the persona's own sentences

The console's edition view shows each piece's lint score and the editor's before → after. The three proof editions were re-gated as the baseline (see the chat of 09-02).

3f½ · The writer, since 09-19 — the persona drives

3g · The quality score — the target function

IQS = 100 × (0.40·E + 0.40·V + 0.20·S), over pieces that pass the lint. The brew now commissions N+2, gates, scores every survivor, and ships the top N by IQS — a spare with a better score beats a weak main. The console's edition view shows each piece's IQS with its E · V · S and the editor's one-line verdict. Cost: about 0.7¢ a piece (four Flash calls; the traits once per persona).

The first reading (09-02, the 17 proof pieces, 12.6¢): IQS ran 47–92. Buffett's letter led (E 1.0 · V 0.87 · S 0.84); Li Ka-shing's letter and Jobs's essay followed; the bottom were Dylan's Shouts (V 0.47, stylo 0), Chestnut's fable (trait match 0.0) and Twain's kicker. Three things the numbers said at once: ① specificity is the weak axis across the board (S 0.10–0.84, most under 0.5) — the reporter's notebook is the lever; ② trait match is the discriminating voice signal (Jobs, Mouratoglou, Banksy 1.0; Churchill, Darwin, Einstein 0.2; Chestnut 0.0) and the judge's own words agree —「Churchillian pastiche whose rhetorical elegance outstrips its content」; ③ the imposter test was too easy (1.0 everywhere): with two other-persona decoys on other subjects the judge was picking by subject, not by voice. Second cut, the same piece rewritten by an anonymous columnist: still 1.0 — the judge picked「the one with a voice」. Third cut (shipped): the decoy is the same piece rewritten in another persona's voice from that persona's own paragraphs, so the judge must tell Churchill from Dylan on the same facts. Re-scored, the test still passes every piece — but a NEGATIVE CONTROL (the key set to Dylan's prose, the texts Churchill's piece, its Musk-voiced decoy and Dan Wang's) answers 0.5 with「none of the options match Dylan」, so the judge is reading style, not guessing: the personas' pieces really are in voices consistent with their samples, and the test is a FLOOR (a piece that failed it would be badly wrong), not a ladder. V is weighted accordingly — trait match ½ (the ladder: 0.0–1.0 across the cast), the imposter floor ¼, the stylometric rank ¼.

How it becomes an optimisation: every knob the brew has — the writer's brief, the exemplar count, the editor's passes, the model, the register mix, the notebook — is changed one at a time and read on IQS over a week of editions, with the readers' verdicts as the check that the judges still agree with people. The persona × topic map (IQS by persona and by beat) is the first thing to read: it says which crossovers work and which personas write as imposters.

3h · The reporter's notebook, and the voice harvest

3i · Length — what the field runs

The engagement data agrees with the print numbers rather than with「write short」: Medium's completion peaks at a 6–8-minute read (1,000–1,200 words); Substack's most-engaged cluster is 800–2,500 words with 70–80% completion at 1,200–1,800 (Substack Course). The owner's ranking of the formats by read quality (09-03): essay ≈ numbered notes ≳ notes > letter ≈ fable ≫ kicker — recorded here; what it changes in the slate mix (fewer kickers, or none) is decided on the specimen brew at the new lengths. Ruling (09-03, late): the kicker stays for testing — dropped only if it keeps getting bad reviews. The specimen at the new floors (brew 2026-09-07, ten pieces): the writer does not obey a floor it is merely told — essays and notes landed at 700–1,060 words against 800–1,400, a scorecard at 570, a fable at 370; and the editor's passes trim further (Buffett 746 → 572). So a draft under 85% of its floor now gets one extension ask in the writer's own call — carry the argument further from the notebook (another reported fact, a second example, the objection and the answer), never pad; the longer draft stands (extended on the piece, shown on the brew log). Measured on brew 2026-09-08 (ten pieces): the drafts landed at the floors (Darwin 1,057 · Dan Wang 1,033 · Augustus 938 · Nolan 886 · Karpathy 875 words — against 700–1,060 the night before), but the editor cuts what the writer built (Nolan 886 → 631, Karpathy 875 → 701): a piece the editor trims under its floor is now extended once more and re-linted, where the words are last touched. The gate also dropped two edited pieces at exactly three soft tells (a contrast, a dash pair, a triple — what the majors carry at that length): the allowance is now three per thousand words inclusive, a fourth fails.

4 · The console's Ink tab

The admin console gains an Ink tab in its own list-row grammar. Six sections down the left; the landing is today's edition with the brew's health strip on top. Everything the pipeline froze is inspectable here, and every policy number above is a knob.

5 · The build — lib/ink_brew.py (2026-09-02)

python lib/ink_brew.py evening --n 16            # 20:00 CST (deploy/mad-ink-evening.timer)
python lib/ink_brew.py dawn --cutoff 07:15 --publish-at 07:30   # 06:45 CST (deploy/mad-ink-dawn.timer)
python lib/ink_brew.py brew --n 16 --sync-art     # dev: both passes now, art painted sync, publish at once
python lib/ink_brew.py status --date 2026-09-03   # what the day's brew holds

• Where: MAD_INK_DIR — the box's /var/lib/mad/ink (set in the box's .env beside the DB and rooms paths); dev launchers set rooms-dev\_ink. run_room serves /ink from it and seeds an empty dir from the repo's lib/ink once, so a fresh box shows the shipped editions until its first brew. The two systemd timers ride in deploy/.
• The anchor bank in practice: the Vibes voice card ($MAD_ROOMS_DIR/_vibes/cards/<slug>.json, 20–30 anchors) is the bank; a persona with no card gets one extracted at first commission (~1¢). The ledger burns used anchors for 90 days and the writer is dealt at most 3, told to use at most 2 —「method, not memoir」.
• The first run (dev, 2026-09-02 14:41–14:47 CST, at PEAK rates): n=6 for the 09-03 edition — 110 pegs · the budget meeting on V4 Pro in 3 min (8 commissions, zero policy problems, 5.5¢) · 8 pieces written in parallel in 2½ min (119–640 words) · gate: all clean, antithesis 0–2 per piece, the three light pieces judged「landed」by the cold reader · 6 zh renditions · 6 heroes painted sync in 25 s · composed L r r W r r (Churchill leads on Hormuz, Buffett's letter takes the well, Twain's kicker closes). 33¢ all in, 7 minutes. Two traps met and coded around on the way: V4 Pro's「Server Overloaded」503 at peak (→ the Flash failover) and Flash STARVING on the 24k thinking budget (→ an empty answer flips the retry to thinking-off JSON mode).
• The dedupe (owner's catch, 09-02:「why Churchill on Hormuz again?」): the ledger had only ever seen what the brew published, so the two hand-made editions were invisible to it. Now the ledger folds in every edition file on disk, and the validator refuses a subject repeat (a persona back on a subject they wrote about in the last 14 days — shared title/provenance/tag tokens name it) and a repeat lead (the same persona leading twice in 7 days) — re-asked, then trimmed. Re-validating the 09-03 slate against the shipped editions now catches it:「winston-churchill wrote on hormuz/iranian on 2026-09-01」. And the shelf's cap (owner, 09-04:「the third article on the Iran war in four days」): the per-persona rule let three different personas take the same war on three days; a subject any persona wrote on in the last four days is now refused whoever the writer (shelf_repeat, in the validator and in the editor's brief) — the way back to a running story is the second-day form, a Scorecard or a Response, or a continuity/reply trigger that names what moved.
• The owner's read of the first edition (09-02) → six fixes: ① no ✦ in a body — the writer never emits one and publish strips any ornament line · ② the reader sees only words: the editor may not commission a piece that hinges on an image/video/clip the reader can't see (the Twain kicker rode a deepfake nobody could look at), and the writer states in one sentence what such a peg shows · ③ stale pegs: the extra beat queries ride Serper's news vertical (past week) and every peg passes a freshness gate — undated or older than 7 days is out (Darwin's piece rode a 25 April report) · ④ the cliché anecdote: anchors whose source is the profile are never dealt — the profile rides into every chat, so its stories are exactly the ones everyone has heard (Buffett's) — and the hand is「at most one」· ⑤ the needless disclaimer: 「I haven't touched the phone」is a rule now (never announce what you haven't done) AND a gate regex that sends the piece to rewrite · ⑥ the antithesis threshold drops to 2 (owner: still strong at 3) and the one rewrite ask covers both tells. Kept as is, on the owner's word: Churchill's opinionated Hormuz piece —「sounds confident and logical, that's what we need; a reader who wants another Churchill builds version B in the Studio」.
• Not yet: the reader's Language switch for the renditions (the ⋯ menu item + the honesty pill are mocked, the data now ships in every edition) · the console's Ink tab (§4) · the console's Models-tab card for app.ink · the 12-piece light specimen read (funny is still unproven) · the door that seeds a room (step 1.5).

The owner's calls