The writing loop — write like a person first, then put the persona back plan 09-10 · checkpoint 1 09-11 · §9

Owner (09-10), after reading the 09-09 and 09-10 editions:「humans don't speak or write this way」— seven tells named (negation titles · flat「X is Y. X isn't Z」sentences · the house vocabulary · metaphors that don't work · obscure, contrived language · streets that ask and cities that answer · numbers that sound awkward). The plan asked for: a study of our writing against good human writers, quantitative and granular; then a loop that takes the persona out and writes like a human writer until a critic agent rates it 8/10 against the majors' 9; then the persona back, with the quality held. V4.1 Flash makes the loop cheap; the cap is $6 of DeepSeek. This page is the plan; nothing on it is built except the first-cut count in §1.

1 · The study — where Ink leaves the human page

Revised 09-10 after the owner's comments: the first draft verified a hand list (my regexes for the seven tells). The owner's point — find the structures and the words, don't verify mine — turns §1a into disproportion finders over the whole vocabulary and over sentence shapes, and §1b into open discovery before any catalogue. The numbers below are the second cut: 88 English pieces, every public edition 09-04 → 09-10 (79.5k words), against the 267-piece majors bank (338k words).

1a · By code — disproportion finders, not hand lists

The instrument is the log-odds ratio with an informative prior (the「fightin' words」method): every word, every two-word phrase, every sentence shape, and every title word is scored by how far its rate in Ink sits from its rate in the majors, z-scored so a rare word cannot win by luck. Nothing is named in advance. What comes out is not a list of words; it is a register, and it reads the same from four directions.

Reproducible: scratchpad/dispro.py (words · bigrams · shapes · openers · titles · number contexts → dispro.json); to become lib/ink_tells.py. The 09-09/09-10-only first cut stays on the ledger for comparison: sentence length matched (17.5 vs 17.9 words); contractions half; questions a third; exclamations none; dashes a fifth; metaphor nouns 2.5 vs 0.9–1.6 per 1k; disclaimers 0.45 vs 0.15 a piece; paragraphs 109 vs 50–69 words. The owner's box notes on 09-09/09-10 sit beside the numbers: Wolfram「ice says, ice wins」; Kurnit's retirement disclaimer「again and again」; Newton「I will not amuse the world with conjectures」; Lu Xun's zh original「城市问、街问、办公室开口」; Ang Lee「plainer, more approachable」; Rams reviewing a phone he never held (a slate rule, parked in §8).

1b · By reading — discovery first, then the catalogue

Counts say how much; they cannot say which metaphor works or why a sentence sounds like a machine, and they only find what they were told to look for. So the reading runs in two passes. Pass one is open: a critic agent (a Claude subagent, not the writer's family) reads 40 matched pairs — each Ink piece beside a major of the same class and beat — with one instruction: list every way the two differ that a reader would feel, in your own words, with the sentence that shows it. No dimensions are given. The notes from 40 pairs are clustered; the clusters that recur become the dimensions. Pass two files the catalogue under those dimensions: the human sentence, the Ink sentence, one line on why one works and the other doesn't. Two worked examples of the kind of entry it files:

Are nine dimensions the major signals? No — nine was my seed list, and the owner's question is the right one. Pass one exists so the list is found. From the code above and from reading the 09-09/09-10 pieces, these are the dimensions I expect pass one to add to the nine (metaphor · sentence shape · vocabulary · numbers · things that speak · title · opening · ending · hedge):

Outputs of the study (one session, no DeepSeek spend): docs/ink-writing-gap.html = the record — the disproportion tables in full, pass one's clusters, and the catalogue, for the owner's read; lib/ink/critic.md = the critic's anchors (per dimension, three「this works」and three「this doesn't」); lib/ink/shelf-sentences.json = the writer's example sheet, whole human paragraphs keyed by dimension, so the brief can show instead of instruct. The NYT question: not needed to start — the plain band already has 106 pieces — but it is blogger-heavy (Graham · Urban · Housel · the Conversation); one subscribed week of NYT Opinion guest essays (~40 pieces, harvested in the owner's logged-in Chrome as the New Yorker was) would give the plain band the newsroom columnist it lacks. Worth ~$1; do it in week 1 if convenient.

2 · The instruments — a frozen dev set and a tells audit

• The dev set: 12 commissions frozen from the 09-09 / 09-10 box brews — the peg, the notebook, the format, the question — copied from the box once; 6 plain (Essay · Notes · Review), 4 literary (Dispatch · Letter), 2 light (Shouts · Kicker). The same 12 every loop, so the only thing that differs between two loops is the one change under test. Notebooks are paid once; a loop pays only for writing, editing and judging.
• The tells audit (lib/ink_tells.py, code): the disproportion finders of §1a run every loop over the 12 pieces against the bank — words · phrases · sentence shapes · openers · titles · numbers — so the list of what is over-used is recomputed, never hand-kept (the answer to whack-a-mole: when honest is cured and candid appears, the finder sees it the same night). Each finding carries the majors' rate as its allowance; a piece is「within」when nothing in it sits over the majors' p75 at z ≥ 3. Three tells need a reader to count honestly and use the extraction pattern proven for 隐喻率 — a Flash call that must quote the expression and give the plain replacement, never a score: things that speak · metaphors that don't work · a title that turns on something not happening. The numbers tell is code: a figure not in the notebook or the peg is unsourced; ten · hundred · times as colour are counted apart.
• The ledger page (docs/ink-writing-loop-ledger.html, generated like the brew log): one row a loop — the change, the tells table, the win rate, the calibrated score, the critic's three notes, dollars to date.

3 · The critic — how it judges, and how we know it judges well

Who. A Claude subagent spawned from the session (a different family from the writer; session spend, per the cost rule). DeepSeek Flash shadows the same pairs each loop; its agreement with Claude is reported, and if it holds at 80%+ over five loops Flash takes the bulk and Claude reads a sample. The writer's own family never judges alone.

How. Blind, bylines stripped. Each dev piece is set against a length- and class-matched major from the bank, in both orders; the critic returns the winner, a margin, a 0–10 on the rubric for each, and the losing piece's three worst sentences with the reason. Those sentences are the loop's gradient: the next change is written from them, not from a hunch.

The exam, before any number counts (re-run at every checkpoint):

4 · One loop

What a loop may change — the parameters

• The brief (the writer's rules and the JSON asks) — subtract first, then add, the 09-07 ruling; every added word is a word that comes back in the piece.
• The examples — how many, whole paragraphs or openings, keyed to the piece's class and to the dimension the last loop lost on.
• The process — outline first or not; a spoken draft first (say it across the table, then tidy); draft → the critic's notes → the writer revises (a revision with specific notes, which the editor's generic passes are not); the title written last from five candidates; which editor passes run, in what order, on which model; temperature and thinking.
• Not a parameter: the commissions, the notebooks, the model (a model change is an arm to run once, later, on the box where Gemini is reachable), the production gate.

The queue — candidate changes, ranked by the numbers

5 · Step 2 and step 3 — the persona out, then back

Persona out. The writer is「a staff writer at Ink」— no biography, no exemplars of its own, first person allowed. The prompt drops from ~47k characters to under 10k (the profile was 72% of it), which is the point: what is left is the brief, the notebook, the shelf, the commission. The loop runs on that until 8/10 holds twice or $4 is spent. The recipe that gets there is the house recipe.

Persona back. The same 12 commissions, the same recipe, the profile in (the writer cut, as today). Two guards, both from the critic: the rescaled score must hold at 8.0 − 0.3 with every tell within allowance, and a point of view must show — shown a pair on the same commission, one by the staff writer and one by the persona, the critic must name the persona's ≥ 70% of the time and say what gave it away (a stance, a memory, a way of putting it). If the score drops, the loop treats the profile the way it treats the brief: what in the profile brings the tells back (the 09-07 finding: the analyst's prose in a profile is read as vocabulary). Stop at 8/10 with both guards green, or at $6 total.

6 · The bill and the clock

$6 buys 25–30 loops off-peak; the plan needs 20–35. Every loop books to its own ledger leaf (app.ink_loop) so the console shows the $6 as it goes and the stop is a number, not a feeling. Heavy runs sit outside DeepSeek's peak (09–12, 14–18 CST).

7 · Checkpoints, and what I need from you now

1. After the study: you read docs/ink-writing-gap.html and ten blind pairs; the critic's exam runs against your ten. Nothing loops before this.
2. When step 2 reaches 8/10: you read the 12 staff-writer pieces and ten blind pairs. If they read like a person to you, the recipe is frozen and step 3 starts.
3. When step 3 reaches 8/10: you read the 12 persona pieces; you decide whether the recipe goes into the brew for the next morning's edition.

8 · Not in this plan

• The Chinese renditions (the owner: 「a disaster」 on 09-10) — English first, as ruled; the same tells exist in zh (Lu Xun's 城市问) and the same loop can run on zh originals after step 3.
• The slate: Rams reviewing a phone he never held → a slate rule (a Review of a product needs a hands-on source, or is written as speculation before launch). Parked as a one-line change to validate_slate, outside this loop.
• Art, the harvest, the meeting, the production gate — untouched. The gate stays code; the critic never enters the brew.

9 · Checkpoint 1 — what sixteen loops say 2026-09-11 · for the owner's read

Built and run on 09-10/11 after the owner's go: the study (§1, the record is the writing gap), the harness (exam/ink_loop.py), the frozen dev set (12 commissions from the 09-09/09-10 box brews, each with a fixed major), sixteen loops of the staff writer, and one loop written by a Claude subagent on the same prompts. Every loop was judged blind by two Claude critics (both orders). DeepSeek spend: $1.58 of the $6. The full rows: the ledger.

What moved, and what did not

• The code tells moved as designed. Negation titles 50% → 0–17% (the title-last rule); disclaimers 0.42 → 0–0.1 a piece; metaphor nouns 3.7 → 1.0–1.7 per 1k; the short flat「X is Y」sentence 13.9% → 3–7% (the shape pass, the spoken draft); paragraphs 95 → 62–80 words; sentence length back in the human band. By the meters, the later loops are close to the bank.
• The critic did not follow the meters. It named the machine in 100% of 374 verdicts, in every loop, both orders. The raw gap to the matched majors stayed 0.1–0.7 through r0–r7 and widened to 1.0–1.8 for every loop after r8. What it names now is not vocabulary: the epigram that closes each paragraph, the thesis re-fingered every paragraph, the announced move, the invented memory dropped in to supply feeling, the gloss on every proper noun, the same two numbers three times, the closing tableau, wire copy in press-release order with a first-person frame welded on.
• Two recipe changes helped: r2, the paragraph rule (L03: gap 0.12, the best of the DeepSeek loops), and r7, Notes and Scorecard in prose instead of numbered sections (L08: the only loop where Ink's raw score beat its majors, 6.55 vs 6.45). One hurt badly and everything built on it stayed hurt: r8, which took the「you are in it」line out to stop the invented framing device, took the person out with it (L09 onward: 「no first person」,「a wire round-up」). L16 confirms it: r2 + the good changes without r8 returned to the baseline band (gap 0.54).
• Machine-on-machine passes made the writing worse while making the meters better — the revision from an editor's marks (r5), the shape pass (r9), the spoken draft (r10). This is the Goodhart case the plan warned about: every one of them is a Flash call editing Flash prose, and the critic reads the seams. The brew's own copy-editor passes are the same mechanism; the box runs three of them on every piece.
• Three harness confounds were found and fixed on the way (from r11 on): the commission's question named the persona and the staff writer bolted the persona on (「Plato hauled in as a frame」); the word「notebook」in the prompt appeared in the pieces; the notebook's NOT ESTABLISHED list became a section of disclaimers. Every word in the prompt is a word in the piece — the 09-08 finding, again.
• The model test. The Claude-written L14 on the same r12 prompts: win 50.0%, raw 6.71 vs the majors' 6.21, the machine spotted 100.0% of the time.

The honest reading

With V4.1 Flash in the writer seat, the brief and the process are worth about half a point of the critic's ten, and no combination tried makes the piece pass as a person's. A Claude subagent on the same prompts scored above its matched majors (6.71 vs 6.21, the only loop to do so clearly) — the seat is worth about a point — and was still named as the machine in every pair, for the same reasons: the aphorism at every paragraph end, the announced move, the thesis restated, the coda. So the habits the owner hears are not one model's accent; they are how a language model writes an essay when asked for one, and neither rules nor passes nor a better model removed them in this round. The tells that are counted went down; the tell that is read did not. The staff-writer step has done its job as a measurement: it says the brief is a half-point lever, the seat a one-point lever, and the paragraph-landing habit the thing still standing.

10 · The owner's blind read — 2026-09-11

Short answer: when it comes to spotting the machine, the critic and the owner agree completely — both picked it out in all ten pairs, even from ~250-word openings. When it comes to which piece is worth reading on, they disagree: the owner would rather keep reading the machine piece in 3 of 7 pairs, the critic almost never. So the critic's "spotted 100%" can be trusted; its quality verdicts can't stand in for the owner.

How it was run: ten pairs from the shape test (five from each arm, each on a different assignment; the pair whose human piece is a broken web page left out). Each piece cut to its opening, about 250 words, at a sentence break, titles kept, no bylines, order random. Two questions per pair: which would you rather keep reading, and which one a machine wrote. A blind critic answered the same two questions on the same excerpts. The page, the key and the answers: rooms-dev/_ink/loop/OWNER1/ (local only — the human pieces are copyrighted).

Pair 4: the owner's notes name the Hungary piece as the machine, and it was, but both buttons were pressed as if it were labelled A (it was B). The page, the critic's pack and the key were checked and agree, so it is counted by the notes. The owner skipped the first question on pairs 8–10.

The machine pieces the owner would rather keep reading

• Pair 2 — a Dispatch written on the "two moments side by side" shape: "well written and has a real development of memory and story" (the critic preferred the human piece, Sarah Larson in the New Yorker, and called the memory made up).
• Pair 6 — a Dispatch on the "it ends worse" shape: "reads well", except the last paragraph's "and then… and then…" (against Adam Gopnik).
• Pair 7 — the Shouts piece, against Ian Frazier; no note.

Two of the three are shaped pieces built on a memory — exactly the kind the critic marked down for "invented" scenes in the shape test. The owner and the critic value that memory in opposite directions.

What gave the machine away, in the owner's words

What the owner loved in the human pieces

"Very personal. Plain, comfortable language. Stories around himself." · "Good detail depiction — I've never seen our Ink articles illustrate the details to this degree." · "Fun to read, vivid, drawing readers into the pictures… how can we produce articles like those?" (Morgan Housel, Long-Term Money). · "The detailed story and fact writing read true."