The gate design — where code judges, where a reader judges, and what nobody scores the defect track BUILT 09-19 (reader marks → the writer mends, r31→r33) · in the morning brew since 09-19 · the seam repair retired
The gate as this week's evidence now supports it. The code x-ray overlapped 15% of your marks on pairs 7–12; six model readers, given a taxonomy written from those marks, overlapped 97%. That is not a verdict against code. It says our defect list is mostly of a kind code cannot define, and we built it as if it were. This page draws the line, and sets the four rules that make a reader safe to use. The writer side of the redesign is its own page: the writer as an agent.
the answer in one paragraphCode owns the closed class — every defect with a finite definition: the antithesis forms, the gavel, the over-use list, dash pairs, numbers absent from the reporting, length, repetition — because it is free, identical across months, and has no favourite model. A reader owns the open class — strained metaphor, lecturing, wise filler, incoherence, restatement, an invented scene, a persona apologising for its era — because those need reading. Between candidate drafts, code filters and a reader picks the survivors. On whether a piece ships, a reader decides with code hard-fails underneath. Nobody produces a score. The reader returns marks with verbatim quotes, which are checkable and which you overrule by tapping, as you did 29 times. Code's second job is to verify the reader: every quote real, every count reported the same way each week.
1 · Two classes of defect
Short answer: if the defect can be written down as a pattern that two people would count the same way, it is closed and code owns it. If it needs a reader who knows the world, it is open and a reader owns it.
Closed class — "not X but Y" in its seven forms. A paragraph whose last sentence opens on That is. A phrase from the over-use list above the bank's rate. Two dashes within 220 characters. A number in the piece that is not in the reporting. The word count. Content words repeated. Open class — whether a figure makes the idea harder or easier. Whether this paragraph is the third paragraph said again. Whether the piece is going anywhere. Whether the person in the anecdote existed. Whether the writer is talking down to the reader. Whether a nineteenth-century persona should be apologising for not knowing about plate tectonics.
| defect | class | who | how it is known |
|---|
| antithesis, all seven forms incl. the split "That's not X. That's Y." | closed | code | pattern; the split form is a known gap, owed |
| the gavel — a paragraph closing on That is / It is, contracted or not | closed | code | pattern; the contracted form is a known gap, owed |
| over-used phrases | closed | code | the profile's list, regenerated from data, never typed |
| dash pairs, rule of three by shape, numbers not in the reporting, length | closed | code | pattern or arithmetic |
| repetition — content-word variety, compressibility | closed | code | arithmetic on the whole piece; the one code measure that found pair 12 |
| restatement — the claim said again in new words | open | reader | code's "novelty" measured new words and called them new claims; retired |
| strained metaphor · wise filler · lecturing · double negation as rhetoric | open | reader | the taxonomy from your marks |
| incoherence, the stiff turn · the invented scene · the era disclaimer · the source dump | open | reader | the taxonomy from your marks |
| things that speak or know ("the ice says") | open, with a closed proxy | reader; code as a first pass | a prototype regex found it at 4.8× the bank's rate with half its hits noise; parked |
2 · Who does which job
| the job | who does it | why that one |
|---|
| Track drift across months — is the model changing under us; did a change help | code, always | One ruler in September that you had in August. A reader changes with its version; the over-use profile does not. The re-baseline rule lives here. |
| Closed-class defects | code, alone | Free, instant, runs 28 times a piece, identical on both sides of every comparison. |
| Open-class defects | a reader, alone | 1,441 marks against the code's 490; 97% agreement with you against 15%. |
| Choose between candidate drafts | code filters, a reader picks the survivors | Ranking is where a biased judge does the most damage, so code kills the clearly worse. But the code ceiling is "cleanest", not "best"; the last choice needs reading. |
| Ship or don't ship | a reader's verdict, code hard-fails underneath | A shipping decision is a reading decision. Dropping a good piece is cheap at 16 written for 12 needed. |
| Verify the reader | code | Every quote exists as text on the page; the same span is not marked twice; counts reported identically each week. |
| A score out of ten | nobody | Your 09-07 ruling. A score is unanchored, drifts between versions, and cannot be argued with. |
3 · Four rules that make the reader trustworthy
- Marks, never scores. The reader returns a quote, a label, and one line of reason. 1,441 marks, every quote verified as real text, none unplaceable — and you disagreed with 29 of them by tapping. A number could not have been disputed that way, and a wrong number teaches nothing; a wrong mark teaches exactly one line of the taxonomy.
- Never the writer's own maker. DeepSeek writes, so Claude or Gemini reads. A model prefers its own family's prose by about a third when it judges its own output.
- The taxonomy is the interface, and it is yours. The reader is only as good as the label list, and the list came from your marks. The loop: you mark a sample; the list is rewritten; the reader applies it at scale; you correct a sample again. The reader never invents a category.
- Code verifies and counts; it never judges the open class. Its job on the reader's output is checking, not second-guessing.
what your 29 corrections already taught the taxonomyThe reader over-fires on two things: comic escalation (Then… Then… Then…) read as rhythm by rule, and an ordinary first-person move (So here's what I would do) read as lecturing. Both get an exception written into the label before the next run. This is the loop working, and it is the whole argument for marks over scores.
4 · What moves, from the gate we run today
Short answer: the code keeps every closed-class key it has; the reader takes over the two jobs code was never able to do; and two known code gaps get fixed once your read pages are no longer live.
| today | proposed |
|---|
| the lint's pass/fail marks a piece "reads as the machine"; nothing is dropped | code hard-fails stay (antithesis, unsourced numbers); the reader's verdict decides shipping |
| the draft pick sorts on six code keys, the last of them a lint score | code sorts and discards; a reader picks among what survives, on the open class |
| the fact-count tie-break (numbers from the reporting) | retired — numbers were never the shortage; names and quotations were |
| "novelty" on the density page | retired — it rewards the habit it was meant to catch |
| the antithesis patterns miss the two-sentence form; the gavel pattern misses That's | both fixed, then the profile re-run, logged as a ruler version |
| no measure of restatement | the reader counts paragraphs restating the claim; the number the density read found separates every bank piece from every Ink piece |
and the larger pointMost of what the gate catches today is what fills a vacuum. If the writer arrives with material — the other page — the reader's gate has far less to find, and the prose rules retire on their own. The gate is the smaller half of the redesign; it is written first only because the evidence for it came in first.
The defect track — built 2026-09-19 · recipe r31 · the reading page
the owner's proposal · 09-19"Can we use a reader to highlight the defects and the defect type and have the LLM rewrite?" Built as ink_brew.defect_marks (a reader that is not the writer — Gemini 3.8 Flash — marks every span with a label from lib/ink/taxonomy.md, the taxonomy learned from the owner's own marks; every span code-checked to be verbatim, spans inside another person's quotation thrown out) and mend_defects (the writer rewrites whole with the marks and their types in view), at most two rounds. The reader test on the gifts piece against a Claude reader's 22 marks: Gemini 20 of 22 in 13 s (21 same label), DeepSeek reading itself 19 in 73 s. On six, read blind against r29: the Claude reader's marks fell 22→8, 24→7, 31→15, 37→19, 43→24, 37→28; for the first time a version was judged not machine-written while its pair was (3 of 6); the in-pipeline reader's own count shows the floor (gifts: 29 → 8 → 12 — the second rewrite makes new tells as it mends old ones). The cost: readers kept the old piece in 4 of 6 — the mended pieces are plainer and flatter ("cut it if the piece already said it" cut the letter frame and the lived scene with the tells), and two carry edit scars because the mend runs after the reread and nothing rereads the mended text. Next: the mend before the reread; "say it plainly" without "cut it"; two rounds and no more.
the two changes · 09-19 · the reading pageThe mend was moved before the reread and "cut it if the piece already said it" was removed, then the six were re-run and read blind against the mend-last versions.
The scars healed and the life came back (the gifts piece holds, keeps its letter and the Talon ad, and both readers kept it) —
and the tells came back with the reread: the Claude reader's marks across the six went 194 (no mend) → 101 (mend last) → 153 (mend, then reread); on the gifts piece 22 → 8 → 26; no version was judged not-a-machine where three had been.
The lesson: whatever rewrites the piece
last decides how it reads, and the reread is the writer thinking in front of the reader, which is where the habits live.
Next: the mend last again, followed by a seam repair that hands the writer only the reader's "incoherence" marks and asks it to repair those places and touch nothing else.
the seam repair · 09-19 · recipe r32The mend back in last place, followed by a repair that hands the writer only the reader's "incoherence" marks and asks it to touch nothing else. It touched little (96–98% of the sentences kept word for word) and it healed the seams (four of four "hold") — and even so the blind reader's marks rose on three of the four repaired pieces (the gifts piece 8 → 13; the repair wrote a fresh "not X but Y" into the very sentence it fixed), and readers kept the mend-only version in two of four. The finding of the day, three times over: every rewrite by this writer, however narrow, brings its habits with it. So the repair is selected, never trusted: the reader marks the repaired text, and it stands only if it carries no more marks than the mended text did (the 09-11 rule: select, don't edit). With that gate the gifts piece keeps its mend-last text. Current recipe: r32 — the persona drives, the life store, the point's tests, the witness, the prior, the two-stage check, the mend last, the gated seam repair. About 13¢ a piece.
the owner's read · 2026-09-19 · the rulingThe owner read the mend-then-reread gifts piece on
the page and said:
"this piece worked." Six of the blind reader's marks on it were false positives — a negative fact reported plainly ("the official did not say whether anyone else received one"), a real choice described ("looked at the size of it rather than the name of it"), a plain list of steps ("Name the donor. State the amount. Choose a category.") and an example list, both marked as transformation — and the lecturing marks miss that a Letter to young counsel is a form in which one may lecture a bit.
So the ruling overturns the ordering I had chosen: the mend runs
before the reread (the reread heals what the mend leaves, and the liveliness that comes back with it is wanted), the reader is told the piece's format so instruction in a Letter is not a tell, and the taxonomy carries the owner's four refinements. The seam repair is retired with the order that needed it.
Current recipe: r33. The marks count, by a blind reader, is a gauge with false positives in it; the owner's read is the measure.