The writing process — a human columnist's, ours today, and the one proposed built 2026-09-17 as recipe r25 · r33 since 09-19 · IN THE MORNING BREW 09-19 (the console's "the persona drives" switch, on by default)

Your hunch, after the outline questions: putting code in charge of designing the outline is the wrong thing, and the writing steps themselves deserve suspicion. This page sets our steps beside a human columnist's, says what the comparison shows, and proposes a process where the persona drives and the code only checks.

the answer in one paragraphThe hunch is right. A columnist's point comes first and drives everything after it: what to look up, what goes in the outline, which facts stay. Ours runs the other way round: we gather first, the same way for every story; code picks an outline by its proportions; a model deals facts out to fill the cards; the writer sees one card at a time; and nobody ever rereads the piece. The 44% background we measured in the big papers is a result of having something to say and going to find what backs it. We made it a target and let code steer by it. Proposed: the persona reacts, asks its own questions, states its point, outlines from the point, writes the piece whole and rereads it once. Code and readers move to where a newsroom also keeps checkers: facts, repeats, names from nowhere.

The words used here. The point (or claim): the one thing the piece says, in a sentence. The outline: the order of the piece, written before the body. A card: one line of the outline. The mix: how a piece's words split between the news, background (what came before or beside the news), the writer's own life, and the writer's view; the big papers run 14 · 44 · 8 · 35. A gauge is a number we read after the fact; a target is a number something is steered by. The reread: the writer reading its own finished piece as a reader would and reworking it. A checker: someone who does not write or design the piece and only tests it against the material.

the owner's rulings · 2026-09-17 · built the same day · the reading page1. The 09-11 ruling is loosened: the writer may reread and edit its own piece. 2. Test whether the personas' questions converge; if they do and are boring, fold them into the canned list. 3. The owner reads only the gifts piece; more may be written and read by the readers.

The questions test (twelve personas, 70 questions, one reader): they differ. 58 of 70 could only come from a close reading of that story, and the trade shows: Feddo asks which statute allows an outright ban, Ive asks only for measurements, Darwin asks whether the age is tied to the specimen or the rock. Plato's were the weakest, all headline-level. Half ask about the event itself and half about what lies behind it. What they almost never ask: the same thing elsewhere (0), the critics (1), earlier coverage (2), precedent (3); those four canned searches now run beside the persona's own. Four kinds they kept asking that the eight lacked are added to the canned list as a fallback: the document itself, the rule it rests on, the next date, what the other side said back.

six commissions, r25 against r23, read blind	reader keeps	teaches more	holds together: old → new	restating ¶	fact check ok · altered · unsupported
Kurnit · the gifts (two readers)	new, new	old, old	stitched → holds	6 → 4	34 · 1 · 6
Brown · dengue	new	new	stitched → holds	4 → 2	37 · 4 · 4
Didion · the edited video	new	old	stitched → holds	5 → 4	27 · 2 · 4
Plato · Hungary	new	new	stitched → holds	6 → 6	39 · 5 · 10
Bourdain · the air-traffic glitch	old	old	holds → stitched	3 → 2	40 · 2 · 5
Feddo · the Canada ban	old	new	holds → stitched	4 → 7	34 · 4 · 1

The reading. Readers kept the new piece in 4 of 6 (the fuller-notes recipe had managed 3 of 6 against its own predecessor, part-by-part 1 of 4), and it holds together in 4 of 6 against the old pieces' 2. On the gifts piece both readers gave the same reason, "one idea, and every paragraph serves it", and both said the old piece knows more: the writer dropped what did not serve its point, which is a columnist's choice with a columnist's cost. The reread works as a reread: unprompted it caught the point made twice, a person never introduced, a memory that stopped the piece, a claim the reporting never made. Two things did not work. The in-pipeline checker (a DeepSeek call) flagged mostly fine paraphrase, missed the invented envelope, and its corrections pushed the writer into copying the source word for word; a reader with the material beside it found the real slips. And background runs 6 to 22% against the big papers' 44%, because half of what the personas ask is about the event; a gauge reading, and "teaches more" split three to three. Every version still reads as a machine: the defects are where they were, and with no code re-roll the counted ones are back (18 on the gifts piece). About 8¢ a piece, 5¢ of it search fees.

the check · 2026-09-18 · the owner: "use Gemini 3.8 lite to have a test"There is no Gemini 3.8 lite; the lite ids Google serves are 3.1 and 3.5, so 3.5 lite and 3.8 flash were tried, beside DeepSeek Flash, each against what a Claude reader with the material beside it had found (7 slips in the gifts piece, 15 in the Hungary piece). Asked in one call to "find everything wrong", all three under-read: Gemini 3.8 returned 0–2 flags, 3.5 lite returned four that were all wrong (it flagged facts the material plainly held). Narrowed to two stages, list every checkable statement first, then verdict them a dozen at a time: DeepSeek Flash found 2 of 7 and 5 of 15 with no false flags, in ten seconds and for about a tenth of a cent; 3.5 lite found 0 and 2 with false alarms; 3.8 flash found 0 and 4 but took two minutes and failed to return JSON three times. What was missed by all three is the subtle kind: the invented addressee, the box the word was "typed into", the government's "typeface", one of the two envelopes. Decision taken: the two-stage check on DeepSeek Flash replaces the one-call check in recipe r25 (the recipe's checker field can name another model). Open: the cheap models read at about a third of a Claude reader's recall; a Claude Haiku call for the check would cost about a cent a piece and is the owner's to approve.

1 · Side by side

	a human columnist	Ink today (recipe r24)
1 · the start	Reads the news and has a reaction: something bothers or interests them.	The editor hands over the news and a question. The persona has reacted to nothing.
2 · what they bring	The news reminds them of things they know and have lived through.	Nothing is recalled. The profile sits in the prompt.
3 · research	Driven by their own question: is that true, has it happened before. What they find can change their mind.	Runs before the writer has any idea. The same eight searches for every story; a desk picks the twelve "best" facts.
4 · the point	They work out the one thing they want to say, often by talking or drafting.	Never formed as a step. A claim just appears in the outline or the text.
5 · the outline	Follows the argument: what the reader needs first, where the piece turns. Many skip it at 900 words.	Six sampled; code picks by the mix of the cards' labels; code re-rolls any card with a "not" or a dash.
6 · choosing material	Keeps what serves the point and drops the rest, good facts included.	A model deals every fact to a card, to fill the cards.
7 · drafting	Writes the whole piece, in one voice, with the whole piece in their head. Facts are woven into the argument.	One card at a time, seeing only that card's material. Four versions; code picks one.
8 · revision	Rereads as a reader; cuts, reorders, removes what repeats. Most of the quality comes here.	None. We select between versions and never rework one.
9 · the editor	Asks "what's your point?", trims, pushes back.	Nobody does this.
10 · fact check, copy desk	A separate person checks every name, number and quotation.	Partly covered: code catches names and numbers from nowhere; a reader checks facts, as a measurement only.

2 · Who drives

The same three processes as chains. The colour says who decides at each step.

the writera desk (a model that is not the writer)code decidesa checker (code or a reader; tests, never designs)

a human columnist

reacts to the news→recalls what it reminds them of→looks things up, by their own question→finds the point→outlines from the point→writes it whole→rereads, cuts, reorders→editor · fact check

Ink today

the reporting→eight fixed searches, twelve facts kept→six outlines→code picks one by the mix, re-rolls cards→page notes→facts dealt to the cards→four versions of each part→code picks each part→readers measure

proposed

the reporting→the reaction→its own questions→the archive answers them→the point→the outline, from the point→written whole→reread once→fact check · repeats · names from nowhere

In today's chain the writer holds two of nine steps and code holds two of the decisions that matter most: which outline, and which paragraph. In the proposed chain the writer holds six of nine, the desk fetches what the writer asked for, and code holds no decision about the piece.

3 · What the comparison says

The point comes first, and ours never comes at all

For the columnist the point decides what to look up, what goes in the outline and which facts stay. We gather first, plan by proportion, and never form the point. That matches what the readers keep finding: every version of the gifts piece restates one thin claim, the form is honest and the gesture is the message, between three and seven times. A piece with nothing more to say says it again.

The proportion is a result, and we made it a target

The big papers come out at 44% background because the writer had something to say and went to find what backs it. We measured that correctly and then used it wrongly: the outline is picked by how close its cards' labels sit to 14 · 44 · 8 · 35. Three things went wrong at once. The labels are the persona's own and are often false (a "life" card with nothing of his life in it; a "view" card that was really news). Cards are not words: three archive cards in seven still gave 24% background and 49% view. And a badly ordered outline can win, because the pick never looks at order. It stays a good gauge. It was never a steering wheel.

Code belongs where a newsroom keeps checkers

Row 10 is the one row where a human newsroom also uses someone who is not the writer. That is where our code has earned its place: it caught the invented "Linda Painter… In 1993" as names and a number that were never in the room, and the fact checker found the envelope nobody reported. Designing the piece, which outline and which paragraph, is the writer's job in every newsroom there is.

We fixed the stitching by taking the whole piece away from the writer

Writing part by part did what it was built for: both blind readers said both attempts hold together, and restatement fell from 5–7 paragraphs to 3–4. But a columnist avoids stitching differently, by holding the whole piece in mind and then rereading it. We have no reread. The 09-11 ruling against AI editing was aimed at machine passes that polish style and teach the piece to please the meter; it also removed row 8, which is where a human writer gets most of their quality.

4 · The proposed process, step by step

	step	who	what happens	on the gifts piece (an illustration)
1	the reporting	desk	As today: the research desk reads the story and writes the notebook of facts, numbers, people and quotations.	The four aides, the sums, the form's four words, Painter, Ronquillo, the White House's answer.
2	the reaction	the writer	The persona reads the notebook and says, in a few plain lines: what struck it, what it reminds it of from its own life and work, and what it does not yet know and wants to.	"A label on a form. I spent forty years on labels nobody reads. Was this ever done before? Is a gift from a boss a gift in law?"
3	its own questions	the writer	From the reaction, three to six questions it wants answered before it will take a position. These become the searches. The eight fixed searches retire as the driver and stay as a fallback when the persona asks too few.	"Has a president given staff cash before?" · "What does the gift rule say about a superior giving to a subordinate?" · "What did ethics officials say this week?"
4	the archive answers	desk	The machinery already built: search, read the pages once and keep the text, write the notes. Returned to the persona as answers to its questions, each with its source, said once.	No known precedent; tokens before. The rule bars gifts going up, says nothing of gifts coming down. § 209. Fox: "troublesome".
5	the point	the writer	After reading the answers, one or two sentences: what it now holds, and what changed its mind if anything did. Two tests from the claim step: the point must not already be in the source story (code can test that), and it must say how something works, not only which side it is on (a reader tests that).	Before: "the form is honest and the gesture is the message." After, if the answers do their work: "the gift rules were written to stop money flowing up, so nothing on the books meets money flowing down."
6	the outline	the writer	Its own, from the point, in its own order. One outline, not six. No pick by code. The mix of the finished piece is read afterwards as a gauge.	The news → the rule that points the wrong way → no precedent → the zipper ad and what a footnote fixes → what the form cannot say.
7	written whole	the writer	One draft with the whole piece and all the answers in view, as recipes up to r23 did.	—
8	the reread	the writer	The persona reads its own piece once, as a reader, against four questions and no style rules: what is said twice, what is used before it is introduced, where does a paragraph not follow from the one before, what would a reader skip. Then it writes the piece again, whole. This needs your ruling; see §6.	The r23 piece introduced Painter and Fox twice and told "a third" and "a quarter" of a salary. A reread is where a person catches that.
9	the checks	checkers	Code: names and numbers found nowhere in what the writer was given; sentences that repeat. A reader with the material beside it: every name, figure and quotation (skill ④). A flagged statement goes back to the writer with the material, to correct or cut. The mix, the restatement count and the defect marks are read and reported, never steered by.	The envelope, "a check", "out of his own pocket", "two lawyers" where one is a professor.

5 · What carries over, what steps back

carries over	steps back from steering
The archive's machinery: searching, reading ten pages, keeping the page text from the first read, the page notes.	The eight fixed searches as the driver of research (kept as a fallback).
Saying each fact once: the de-duplication of the notes before the writer sees them.	The deal as an allocation of facts to cards.
The names-and-numbers guard and the repeat guard, as checks on the finished piece.	Code picking the outline by the mix; code re-rolling cards for a "not" or a dash.
The readers: blind pairs, the defect marks, the material mix, the fact checker. The workbench page.	Code picking each paragraph from four. Kept on the shelf: it is the only thing so far that removed the stitching, and the test below says whether the reread can replace it.
The mix 14 · 44 · 8 · 35, as a gauge on the finished piece.	The same mix as a target.

what this gives upPart-by-part writing is the one change this week that both blind readers agreed on: the piece holds together. Writing whole brings the risk of stitching back, and the reread is the only thing standing against it. If the reread does not hold the piece together, the honest fallback is part by part driven by the writer's own outline, with code checking and not choosing.

6 · For you to decide

1 · the rereadYour 09-11 ruling says no AI editing. It was made against machine passes that polished style and pleased the meter. A writer rereading its own piece for what repeats and what does not follow is a different act, but it is still a model rewriting a model's text. Three ways to allow it: (a) a full reread and rewrite, as in step 8; (b) a reread that may only cut and reorder, never add a sentence; (c) none, and keep part-by-part as the guard against stitching. I would try (a) and (b) on the gifts piece and let the blind readers say.

2 · the questionsDo the persona's own questions replace the eight fixed searches, or run beside them? I would let them replace, with the eight as a fallback when the persona asks fewer than three. Beside them doubles the material and brings back the desk choosing what matters.

3 · the testThe gifts piece only, as you asked. The new process against the piece you liked (r23, written whole with the fuller notes) and against part-by-part attempt two: the same two blind readers, the fact checker, the mix as a gauge. About 3 to 4¢ a piece: three short calls for the reaction, the questions and the point; the archive as now; one whole draft; one reread.

7 · The risks

A model asked for a reaction will perform one. The same caution as on the claim step: asked for insight, it produces the look of insight. The guard is that the reaction must end in questions, and the questions must be ones the notebook cannot already answer; code can test that.
The point may be the editor's. The commission's question already leans toward a thesis (the upstream study found it pre-writes one). If the point after research reads like the commission's question turned into a statement, the step has done nothing. A reader can test that.
Generic questions. "Has this happened before?" is one of the eight fixed searches. If the persona's questions are the same eight in its own words, nothing was gained but nothing was lost.
The defects do not move. Nothing here touches the habits in the sentences. That stays the separate matter you named it.

for InkLet the persona drive and let the code check. The week's measurements stay as gauges: the mix, the restatement count, the marks, the fact check. The week's machinery stays as the desk: the archive, the notes, the guards. What changes is the order of the steps and who holds the pen at each one.

Proposed and built 2026-09-17 (recipe r25; ink_brew.reaction · research · the_point · own_outline · reread · fact_flags · correct; exam/ink_loop.write_driven) · nothing deployed · the evidence: part by part · the dossier · the workbench · the roadmap this revises: the writer as an agent · the gate design