The writer as an agent — why the writer needs material rather than rules, and the six skills that supply it built 09-18 · all six skills built or tried (⑥ not adopted) · the point's two tests · the defect track (r31) on the gate page, 09-19 · IN THE MORNING BREW 09-19 (recipe r33)

Your argument, with the measurement that settles it. A human writer has years of experience and the ability to look things up; our persona has a few thousand tokens of condensed life and only the source text it was handed. Asked for a thousand words, it plays "invent the missing links" under a rule that forbids inventing, and the result is the same point restated six times in different clothes, or strained rhetoric. The redesign is the writer as an agent with skills, each with an expected effect it can be held to. The gate is on its own page: the gate design.

the answer in one paragraphThe writer is handed a median of 904 words of reporting and asked for 700 to 1,200 words. That is about one to one. A columnist writes a thousand words off twenty or fifty thousand they never show. We are not asking our writer to compose; we are asking it to inflate. The fix is not a better prose rule, it is ten times the material: an archive of what came before, a retrievable life instead of a dealt hand, a witness search for names and quotations, a check, a memory of the persona's own past positions, and a scope step that decides format and length after the material is in — including "no piece". Then a claim step, so that what comes out of the material is an insight and not a summary. The bar, from the density read: a claim stated once, supported by facts the source page does not carry, from a vantage the reporter does not have.

1 · Why the writing fails: the arithmetic

Short answer: we hand the writer about as many words as we ask it to produce. There is nothing to compose from, so it inflates.

what the writer gets, per assignment (median of the twelve)	words
the reporting — every fact it is allowed to use	904
the persona profile — who it is and how it writes	2,855
what it is asked to produce	700–1,200
a human columnist's ratio of material read to words published	20–50×

Three instructions are given at once, and they cannot all be obeyed: invent nothing, write a thousand words, here are nine hundred words of facts. A writer with a life and a library resolves this by bringing the other forty thousand words. Ours has neither, so it resolves it the only two ways left open: say the same thing again in new clothes, which the density read counted at three to six restatements a piece, or dress a thin thought in rhetoric, which is the strained figure and the wise line. Both are what you have been marking all week. They are not habits of style. They are what a vacuum sounds like.

The profile is not material. It is three times the size of the reporting and the wrong kind of thing: instructions about voice, plus ten or so anecdotes dealt at random. A dealt anecdote that does not fit today's news either gets forced in, which reads as a costume, or gets ignored, which leaves the vacuum. Kurosawa's brother and the ink in the rain tanks worked because they genuinely rhymed with the theft of a painting. That was luck, not retrieval.

2 · The six skills

Short answer: in the order I would build them. The first two close most of the gap. Each carries the effect we expect from it, in a number we already measure, so each can be judged on its own.

① The archive — what came before this news

The single biggest difference between our pieces and the majors'. The Times piece on women and doctors is dense because the writer brought a study of two thousand patients, research on diagnosis times, and the four words that get written in a chart. None of that was in the day's news; she went and got it. Our writer is handed one story and nothing else. The skill: before writing, search for the prior three stories on this beat, the last time this happened, the numbers over five years, the study everyone cites. Three or four searches, two reads.

expected effectEvery piece carries at least one fact the source page does not, checked by code (a fact in the piece, absent from the peg). Names per hundred words move from Ink's 3.2 toward the bank's 7.4. Restatements per piece fall from 3–6 toward 1, because there is a second thing to say. The reader's "source dump" marks do not rise, which is the test that selection held.

built · first run · 2026-09-17 · the workbenchBuilt as ink_brew.archive() and recipe r21 (three drafts, beats, repair, plus the archive), run on two commissions the owner already knew, the $45,000 gifts (Kurnit) and Florida's dengue (Brown), before = r20 on 09-14, after = r21. The archive found what it was meant to: the 1995 disclosure rule and a 2012 Congressional Research Service report on the President's gift rules; for dengue, that 2025 reached 50 local cases, and a 2024 study giving the 2009–2021 median as seven a year. On the first write the writer ignored all of it — the block said "use at most one, only where it changes how the event reads", and it took the permission not to. Made a requirement ("the piece must carry one of these, with its date"), and both rewritten pieces carry it.

per piece	gifts · before → after	dengue · before → after
a fact from before the day's news (the reader's check)	a lived memory → 1995 rule + the CRS report	none → the 2009–2021 median
paragraphs restating the claim (the reader)	9 → 6	5 → 5
the reader's marks, all labels	49 → 42	38 → 49
the reader's "source dump" marks	1 → 2	1 → 2
names per 100 words	6.7 → 4.6	3.1 → 3.7
quotations per 1,000 words	14.6 → 0	6.2 → 4.2
hollow verbs %	37 → 29	40 → 27
the reader's verdict	keep before · machine both	keep after · machine both

The honest reading. The skill did its own job: both pieces now carry a dated fact the source page does not, and on the gifts piece the restatement fell by a third because there was a second thing to say; the hollow-verb share fell on both. It did not lower the tells, and on the dengue piece it raised them, the gavels doubling. And both rewritten pieces lost the one lived scene the earlier versions had — the father's zipper advertisement, the lake — which is what the reader kept the gifts "before" for. The archive displaced the memory rather than joining it. That is the case for skill ② being next and not optional: the archive supplies what came before in the world, the life store what came before for the persona, and a piece needs both or it trades one for the other. Also seen: names and quotations went down, because the archive items are rules and figures, not people — the witness (skill ③) is a separate gap, as the density read said. Cost: about 2¢ of search fees and 5¢ of writing a piece at peak; 109 seconds for the two.

second run · 2026-09-17 · the owner's read of the first · the workbench · the dossierThe owner read the first run and said four things: the archive is still a tiny portion of the piece and most of it is the writer's own; the reporting comes back paraphrased, sometimes word for word; the defects persisted; and the beats themselves carry the habits. Two orders: make the archive research more extensive and aim at the majors' proportion, and make the outline clean.

The target was measured first, because we did not have it. Two readers labelled every sentence of five news-pegged columns from the bank by where its material comes from. By words: news 14% · background 44% · the writer's life 8% · the writer's view 35%. Our two first-run pieces: 24 · 11 · 14 · 52. Background = anything from before or beside the event that a writer looks up. View = argument carrying no outside material.

Built. The wide archive (archive(wide=True)): eight questions instead of four, ten pages instead of four, each page distilled alone, a pool of twelve instead of three kept. The plan (ink_brew.plan(), recipe r22): its own call, six outlines sampled as index cards, every card read by the code (ink_lint.line_faults: any "not", any dash, any stage direction fails a card), the outline picked for the mix nearest the majors', each faulty card re-rolled alone; the writer gets the cards and only the archive items they use. The dossier (archive_dossier(), recipe r23): the notes of every page the plan cites, up to 150 words a page, because twelve one-line items are three hundred words and cannot fill 44% of a piece.

per piece (the readers)	gifts · r21 → r22 → r23	dengue · r21 → r22 → r23	the majors
background, % of words	12 → 24 → 18	10 → 16 → 9	44
the writer's view, % of words	45 → 55 → 38	58 → 50 → 47	35
the news, % of words	36 → 15 → 28	12 → 21 → 31	14
faulty lines in the outline	6 of 7 → 3 of 8	5 of 6 → 0 of 7	—
the reader's defect marks	42 → 54 → 56	49 → 58 → 56	—
paragraphs restating the claim	6 → 6 → 6	5 → 5 → 5	1
the reader keeps reading	r21 over r22 · r23 over r21	r21 over both	—

The honest reading. (1) A clean outline did not give a clean piece. The cards went from mostly faulty to nearly clean and the marks rose; the habits live in the writer's sentences, not in its plan. (2) The writer turns whatever it is handed into about half view. A wider archive moved the background share a little; only the dossier on the gifts piece (800 words of notes from six pages) brought view down to the majors' level, and that is the one piece a reader preferred, for the right reason: five paragraphs in a row each add something new. (3) A plan of separate cards makes a stitched piece: a person introduced twice, the same history told in two places that disagree, the news never stated because it was one card of seven. (4) Two pieces are not a measurement: the dengue dossier failed to fetch (one page, 80 words), so r23 there is a second roll of r22, and the background share moved by as much as the recipe change had moved it. Proposed next, before skill ②: the owner's Algorithm B, now that there is a plan to drive it: write paragraph by paragraph, each paragraph shown only its own card's material and the accepted text before it, so the mix is true by construction and nothing can be introduced twice; and keep the page text from the first fetch so the dossier never depends on a second one.

the check on six · 2026-09-17 · blindThe owner judged the gifts piece with the dossier to have "more meat", and the defects to be a separate matter from the content. Before building on one piece, the same recipe (r23) was run on six commissions, the gifts piece a second time on the same material. One fix first: the archive now keeps each page's text from its first read, so the notes never wait on a second one. Each new piece was read against the old one (r20, no archive) by a reader who was not told which was which.

commission	notes, words	background %	view %	paragraphs that add · restate	teaches more	keeps reading	holds together
Bourdain · the air-traffic glitch	584	7 → 12	39 → 22	4 → 6 · 5 → 3	new	new	holds
Plato · Hungary expels ten Russians	352	0 → 27	60 → 44	4 → 6 · 7 → 5	new	new	stitched
Brown · dengue	201	4 → 9	56 → 60	4 → 5 · 4 → 4	new	new	stitched
Didion · the edited video	339	11 → 44	43 → 30	5 → 7 · 2 → 5	new	old	stitched
Feddo · the Canada ban	476	6 → 27	49 → 26	8 → 6 · 5 → 4	new	old	stitched
Kurnit · the gifts (second roll)	799	0 → 16	59 → 38	5 → 5 · 6 → 4	new	old	stitched

The reading. The content lever is real: the new piece teaches the reader more in six of six, background rises in all six (a mean of 5% to 22%), view falls in five of six, restatement falls in four. The owner's second point holds too: the defects are a separate matter — the reader called both versions machine-written in five of six, and the marks on the new pieces (33 to 61) are where they always were. What now decides whether a reader stays is neither: it is that five of the six new pieces read stitched, and every one carries more slips than the old (a person introduced twice with the same quotation, "a third of a salary" in one paragraph and "a quarter" in another, two paragraphs that are two drafts of the same paragraph, an invented envelope). The gifts piece, rolled again, lost the reader for exactly that reason. One cause is visible in the material: six pages that all tell the same story give six sets of notes that say the same thing, and the writer uses each. Next, in this order: the notes merged into one dossier with nothing said twice; the piece written paragraph by paragraph, each paragraph shown only its own card's material and the accepted text before it (the owner's Algorithm B); then the check of every figure and name against the notes (skill ④). The owner's reading sample from here on is the gifts piece alone.

the deal + part by part · built 2026-09-17 · the gifts piece only · the reading pageThe owner's go for steps 1 and 2, on the gifts piece alone. The deal (ink_brew.plan_material): one call deals the reporting, the archive items and the page notes out to the plan's cards, every fact, figure, quotation and person to exactly one card, said once; the code drops any line that repeats another or carries a number the material never had. Part by part (ink_brew.write_by_parts, recipe r24, the owner's sequential algorithm of 09-14): one card at a time, the writer seeing the outline, the accepted text and only that card's material; four candidates a part; the code keeps the one that repeats nothing already said, carries no name or number from nowhere, and has the fewest counted defects. About 1¢ and twenty seconds a piece.

	r23, written whole	r24, attempt one	r24, attempt two
holds together (two blind readers)	stitched	holds	holds
paragraphs restating the claim	5–7	3–4	3–4
paragraphs that add something new	10	9	7
the reader's defect marks	56	41	45
fact check: supported · altered · unsupported	not run	28 · 5 · 8	21 · 5 · 5
news · background · life · view, % of words	28 · 18 · 16 · 38	48 · 20 · 2 · 30	19 · 24 · 8 · 49
blind readers who kept it over r23	—	0 of 2	1 of 2

The reading. The stitching is solved by construction: material dealt once cannot be introduced twice, and restatement fell with it. What now decides the read is the outline. Attempt one drew an outline with no line from Kurnit's life and came out a chain of quoted experts ("the I appears once and vanishes"); attempt two drew one with his father's zipper ad, which a reader said "does real work", and used less of the archive. A bug found on the way: a card labelled "view" that named Painter was dealt nothing, and the writer invented "a woman named Linda Painter… In 1993 the office wrote"; now a card that names a person, a figure or a law gets its material whatever its label, and the ruler counts every capitalised name and every number found nowhere in what the writer was given. Next: the outline's pick requires a life line and prefers the outline that uses more of the archive; then skill ④, the check — a reader with the material beside it still finds five altered and five to eight unsupported statements a piece (an envelope nobody reported, "a check", "out of his own pocket", "two lawyers" where one is a professor). Defects stay a separate matter.

2026-09-17 · the process itself is under reviewAfter the outline questions the owner's hunch: code should not be designing the outline, and the writing steps deserve suspicion. The comparison with a human columnist's process, and a proposed order where the persona drives and the code only checks, is on the writing process. Proposal; nothing built.

② The life store — a retrievable life, not a prompt block

Today a persona's life is ten anchors inside the prompt, dealt at random. A person has thousands of episodes and recalls the one that fits. The skill: write each persona's life once, properly — fifty to a hundred episodes, places, years, people, jobs, failures — and keep it on disk, not in the prompt. At writing time, retrieve by relevance to today's story and pass only what clears a threshold. The prompt gets smaller and the life gets bigger, which is the opposite of today. For real figures the corpus is the store and already exists. For invented personas this is also where a wider licence to imagine belongs: an invented life is characterisation, not dishonesty, provided it is written once and stored, so it is consistent across every piece rather than improvised fresh each morning.

discussed · 09-16 · retrieval, with three rules plain retrieval lacksYes, this is retrieval in shape: the life is a store on disk and the writer queries it when it has a story. No fixed cap on how many episodes come back — retrieve by relevance. But three rules, each answering a failure already seen. Keep the take small even when the store is large: top-k with scores, show the writer only what clears a threshold, let it use one or none; a piece that used three episodes is a piece that lost its shape, and we already record which anchors were used, so this is measurable. Zero must be a legal result: today the hand is always dealt and the writer forces an anecdote in; with a threshold, most mornings return nothing good enough, and that is fine — the Heigo passage worked because it genuinely rhymed. Query on the situation, not the keywords: a painting theft retrieves "a thing made to be seen, hidden away", not "painting"; so the episodes are embedded and written as episodes (place, year, people, what happened), never as traits, because a trait cannot be retrieved against a situation. Two consequences: for real figures the store already exists — their own writing, retrieved by passage, so the persona quotes itself with a source instead of echoing the profile's summary of itself; and the pitch should use the same query, so a persona pitches the stories where its store returns a strong match.

expected effectThe reader's "invented scene" and "echo" marks fall to near zero, because the memory used is a real stored one and the profile is no longer quoted back. Anchors used per piece is at most one, and the share of pieces using none is reported, not hidden. Every claim carries a vantage: the claim step (§3) can name where the persona stands. The prompt shrinks by the profile's anecdote block.

built · 2026-09-18 · recipe r26 · the reading pageBuilt as ink_brew.life_store (the persona's life written once from its profile and corpus as episodes and positions, on disk under the Ink dir; Kurnit 57 + 36, Brown 136 + 39, Plato 77 + 27), recall (a Flash reader scores every episode 0–10 for rhyming with the news and what struck the persona, floor 7, at most three, duplicates of one position folded, zero legal), life_text (the block the persona reads, or the line that nothing fits), life_used (a sentence of the piece sharing a third of its words with the episode), and writer_profile(drop_memories=True). About a cent a piece.

six commissions, r26 against r25, blind	recalled · used	reader keeps	teaches more	echo marks r25 → r26	incoherence r25 → r26
Kurnit · the gifts (two readers)	3 · 1	old / new	new / new	2 → 1	3 → 0
Brown · dengue	3 · 1	new	new	4 → 3	5 → 2
Feddo · the Canada ban	3 · 0	new	old	3 → 1	5 → 0
Didion · the edited video	3 · 1	old	new	1 → 0	4 → 0
Bourdain · the air-traffic glitch	1 · 1	old	new	1 → 1	2 → 1
Plato · Hungary	0 · 0	old	new	2 → 0	2 → 2

Against the expected effect. The echo marks fell from 13 to 6 across the six and the incoherence marks from 21 to 5; the memory that reaches a piece is now a real, dated, sourced one (Kurnit: "In 2015 I told MediaPost…"); Plato's record returned nothing and the piece was written without a memory; no piece used more than one. Anchors used per piece: four of six used one, two used none. What did not happen: readers did not keep the new piece more often (2 of 6, against 4 for r25), while saying it teaches more in 5 of 6 — the run-to-run noise of one recipe is as large as this. A finding for the profiles: withholding the memory block did not stop Kurnit recalling the zipper ad unprompted, because his Background section tells the same story; the single-source rule (a fact lives in one section) is the thing to enforce, and the store is where the memories should live. Not done: the wider licence to imagine a life for invented personas; every store here is from the record only.

two fixes · 2026-09-18The Letter brief read "a letter to a named or typed reader" and produced an invented Sam and Jonah; it now reads "to a kind of reader — a young lawyer, a parent in Tampa — never to a named individual". The single-source rule: python exam/profile_lint.py reads every persona's memory block and reports a paragraph elsewhere that retells one, in the sections the writer actually reads; 9 of 92 personas do (Zhu Jiang four times; Plato, Jack Ma, Koons and five others once, all in Background). Kurnit's Background now refers to memory 1 instead of retelling the zipper ad; the other eight are owed a hand fix.

③ The witness — someone who was there, in their own words

The measured gap was people: the bank names someone every fourteen words and quotes three people a piece; Ink names someone every thirty-one and quotes nobody. A human writer telephones someone. We cannot, but we can find what people have already said in public about this story, on the record, and attribute it. The skill: find three people who have spoken about this and what they said, each with a source.

expected effectQuotations per thousand words move from 0 toward the bank's 3. Every quotation in a piece carries a URL in the piece's record, or is cut — a closed-class check. Names per hundred words rise alongside the archive's effect. The reader's "lecturing" marks fall, because a piece with other voices in it is not a monologue.

built · 2026-09-18 · recipe r29 · the reading pageBuilt as ink_brew.witness: after the point, three searches aimed at people who were there or are affected (a passenger, a patient, a neighbour, an employee — not officials or the experts already quoted), eight pages read, every quotation pulled from a page checked by code to be word for word on it (normalised quotes and spacing), up to five handed to the writer with the page each was said on; witness_used credits a quotation only when its words are in the piece. On six: quotations verified on five commissions (a stranded newlywed at Gatwick, a 74-year-old workmate who saw the man faint beside the King, people celebrating in Budapest, a Wisconsin dairy farmer; none for dengue), four pieces used one. Read blind against r26: readers kept the new piece in 5 of 6, the week's best; the gifts piece both readers kept, with the fewest defect marks of any version (22). Weakness: the "skip officials" filter leaked once (a senator on the gifts piece). Against the expected effect: quotations now carry a page; the count per thousand words is still to be measured across a week.

④ The check — one number verified, one opponent found

Two small things with a large effect on how a piece reads. Verify the figure the argument rests on. And find the strongest case against the piece's own view, because a piece with no opposition in it reads as a press release, which is half of what "lecturing" meant in your marks.

expected effectNumbers not in the reporting stay at zero, now including the archive's numbers. Every argued piece answers a named objection, checked by the reader (the taxonomy gains "no objection answered" as a mark). The share of pieces marked "lecturing" or "double negation" by the reader halves, since a straw man is what a writer builds when it has no real opponent.

⑤ The prior — what this persona has already argued

We already record a stance for every piece and never read it back. A columnist knows what they told these readers last month and either extends it or changes their mind in public. The skill: retrieve this persona's previous positions on this beat; the commission then asks for the next step in an argument rather than the argument again. This is the cross-piece version of the restatement problem, and it is nearly free — the data is already on disk.

expected effectA persona's claim on a beat is never its previous claim restated: overlap between a new stance and the persona's stored stances is checked by code and a repeat goes back to the claim step. Across a month, a persona's pieces read as a body of work with a direction, which the reader can judge on a sample, and the same-morning shape likeness on the profile page falls, since pieces stop rhyming with each other.

built · 2026-09-18 · recipe r30Built as ink_brew.prior_add / prior_text: every finished piece's point is kept under the persona's slug (<ink dir>/life/<slug>.prior.json, the same commission written again replaces its entry), and the next commission reads the last five back — "what you have held in Ink before: build on it or turn against it, but know it" — before the reaction and beside what it looked up. Tested by writing Kurnit a second story (the Tate brothers' bail, as him): his gifts point was read back, the piece did not lean on it (a different matter; the right result), and two entries now sit in his prior. The skill only shows its worth over weeks of mornings, which the stopped brew cannot yet give it.

⑥ The scope — the format and the length, decided after the material is in

The cheapest change on this page and possibly the most effective. Today the format and its length are fixed by the slate before anyone knows what the material is. Instead: gather, look at what is actually there, and then decide both — the shape and the size. If the topic and the material call for long form, it is long form. If they call for a scene, a letter, a list, a four-hundred-word note, that is what it becomes. The thousand-word floor is the instruction that manufactures the filler, and no prose rule can survive it. And one more answer is allowed: no piece. The biggest quality lever a newsroom has is killing the story the material does not support. The day's count becomes variable; the floor of twelve becomes a target, not a promise.

expected effectRestatements per piece fall to at most one, because a piece is never asked to be longer than its material. Word counts vary by piece and by morning, and the spread is reported. The share of commissions spiked at the scope step is reported every morning and treated as a health number, not a failure. Hollow verbs (is, has, does) move from Ink's 30% toward the bank's 23%, since a piece cut to its material has events in it rather than states.

The shape of the change. Nothing in it is a prose rule; every part of it is material, the scope decides what the piece is allowed to be, and the claim step decides what it says before a word of prose exists.

tested · 2026-09-18 · recipe r27, twice on six · NOT adoptedBuilt as ink_brew.scope: after its point the persona is asked whether there is a piece at all, which house form serves it, and how long the material carries. First cut (one number, the house bands shown in the menu): all six kept the commissioned form and asked for 950–1,250 words, the top of the band; none said "no piece". Second cut (after the outline, the bands hidden, the length built line by line — "a line with one fact needs about sixty; a filler line needs none, drop it"): all six kept the form, dropped no line, and budgeted every line at 85–210 words; totals 800–1,180, the same as before. Twelve decisions, twelve rubber stamps. The reading: asked to scope, this writer scopes generously; a model asked for a budget performs a budget the way a model asked for insight performs insight. The "no piece" answer, the one thing this skill was for, was never given. Decision: the step is not adopted (the recipe stays at r26); if Ink wants shorter pieces the length has to come from the editor's commission, which is the reverse of what this page proposed, and a "no piece" door needs a test the persona cannot flatter, such as the point's absence test (nothing in the point that the source did not already say).

3 · The claim step — insight, not a summary

Short answer: with ten times the material, the risk flips from a piece with nothing in it to a piece that is a summary of the news plus a memory. Insight is a claim that is not in the material, that the material supports, and that a reader would not have reached from the news alone. A summary fails the first test; a generic take fails the second and third. So the claim is its own step, before any prose exists, with its own tests.

Why a step and not a rule. The loop found that a clearer first ask beats a vague ask plus an editor. Asking the writer to "have an insight" inside the writing prompt is the vague ask. Asking it, before writing, to answer three plain questions is the clear one, and the piece is then written from those answers.

The claim step. After gathering, one call, plain sentences, no prose: What do I now think that none of the sources say? Which two facts in the material make me think it? What would someone who knows this field say against it? The piece is written from these, the way the interview already supplies material in the persona's own words.
The absence test — code. The claim must not appear in the sources. Overlap between the claim's words and the reporting is a closed-class check. A claim that is mostly the source's words is a summary and goes back to the step.
The support test — a reader. The two named facts must actually carry the claim. A reader can answer that; code cannot.
Mechanism, not stance. "This is bad" is a stance and costs nothing. "This happens because of X, so Y follows" is a mechanism, and every dense piece in the six had one: chart labels lead to worse care; one car space holds ten bikes; the model writes the paragraph and the bench does the rest. Ask for the mechanism or the consequence by name. A stance without one is filler with a verdict on it.
The objection must be answered in the piece. The strongest opposing view from the check skill is answered, not mentioned. Insight is what survives an objection; a piece with no objection in it reads as a lecture, which is what your marks called lecturing.
The vantage. The claim should come from where the persona stands and the reporter does not: Karpathy on what happens after the grant is funded, Kurosawa on what a hidden painting is. The life store supplies the vantage; the claim step asks what this looks like from there that it does not look like from the newsroom.
The check at the end. The count from the density read: how many paragraphs restate the claim. One statement, then the support, then the objection, then the consequence. A piece that says it three times has no second thing to say, which is the signal to shorten it or spike it — the scope step again.

expected effectEvery shipped piece has a claim that passes the absence test by code and the support test by the reader. The reader's answer to "does this piece say something you didn't have before" — the question from the expert checklist where GPT-4 scored 44% against the New Yorker's 92% — becomes the piece-level number we report, alongside restatements per piece at one.

the cautionA model asked for insight will produce the take a thousand columnists already produced, fluently. The absence test catches copying the source; it does not catch copying the crowd. The only defence found so far is the vantage and the mechanism together: a claim tied to this persona's experience and to a named cause is much harder to make generic than a claim about what things mean.

the two tests · built 2026-09-18 · recipe r28The absence test (ink_brew.absence_test, code): the share of the point's longer words already in the reporting; calibrated on eighteen points — a real point 0.15–0.45, the news said back 0.79–1.0; the line is 0.6. A failing point is asked once more with one question ("what in it is not already in the story?"); a second failure is the honest no piece. The mechanism test (mechanism_test, a reader that is not the writer): does the point say how something works, or only which side it is on; a failure is asked once ("how does it work?") and the piece goes on either way, recorded. On six: all passed both first time (absence 0.15–0.36); the door did not open. Cheap, and they stay: this is the "no piece" test the scope step could never be, because it is applied to the point, not asked of the persona.

4 · What it would cost, and what could go wrong

Short answer: roughly double the bill, a few minutes more per morning, and three new failure modes to watch.

skill	added calls a piece	rough cost	risk
① the archive	3–4 searches + 2 reads	2–3¢	more pasted facts, not fewer — the source dump gets worse unless selection is enforced
② the life store	1 retrieval (built once per persona)	<0.5¢ after the build	a large invented life is a large surface for inconsistency; it must be written once, reviewed, and frozen
③ the witness	1–2 searches	1¢	misattributed or invented quotes — every quote must carry a URL or be cut
④ the check	1–2	1¢	none serious
⑤ the prior	0 — already on disk	0	none
⑥ the scope	0	0	short editions; the day's count becomes variable
the claim step	1	<0.5¢	the generic take — see the caution above
total	~9	5–7¢ a piece, about $2 a morning

the three real risksMore research can make the writing worse. You already caught this: half of one piece read as plagiarism from the source. Gathering more material without a rule of selection produces a longer source dump. The gathering skills must end with a step that keeps one thing and discards nine. The clock. Nine more calls a piece across sixteen pieces is real time against the 07:15 rule; the gathering has to run in the evening pass, not the dawn one. The real figures. A historical persona with a live search will find the present, and the Darwin piece already showed what happens when a persona starts negotiating with its own era. The archive skill needs a rule about what a dead writer is allowed to have read.

5 · The question underneath — answered

This redesign turns Ink into a small newsroom whose reporters happen to be characters. The persona stops generating opinion and becomes a lens on material that was gathered for it. That is a product decision, not an engineering one, and the evidence points at it hard: the densest piece in the six I read, the New Yorker's Talk of the Town, has no argument at all. It is a writer in a room for three hours, forty proper nouns, twelve quotations, no thesis. It is also the piece our brief is least capable of producing, because our brief opens by asking for a stake and a view.

the owner's answer · 09-16A newsroom, as long as the article quality is much better. So the six skills plus the claim step are the roadmap, and the prose rules retire as the material arrives. The bar is the density read's: a claim stated once, supported by facts the source page does not carry, from a vantage the reporter does not have.

For discussion · 2026-09-17 · nothing built · the other half: the gate design · evidence: the density read · information density · the model's read · what the writer reads · coverage and the pitch