/declauding

The register

Every entry: the tell (surface pattern to scan for), why it is a tic, the fix, and a real before/after. Entries are grouped by mechanism because the lexical bans always leak — new phrasings arrive faster than a phrase list grows. Scan at the sentence level, not the phrase level.

1. Staged reveal (the root)

A sentence is staged when it withholds, contrasts, or significance-tags to manufacture a beat. It is stated when it lands content in subject-verb-object. Most entries below are special cases of this one.

Generative test, per sentence: stating or staging?

2. Negation-first reveal

Tell: says what something is NOT before saying what it is — "It's not that X, it's that Y", "The problem wasn't A. It was B.", "not because P — because Q", "X, not Y" as a closer, "doesn't just X, it Y's".

Why: the reader never proposed the wrong answer. You supplied it so you could knock it down, which converts a plain fact into a small drama in which you are right.

Fix: state what it is. If the contrast is genuinely informative, keep it in one clause.

was: It is not a wrong answer. It is a non-answer. Counting it as a failure is the conservative choice, and it is what these numbers do. now: It is a non-answer rather than a wrong answer, and these numbers count it as a failure, which is the conservative choice.

Earned when the reader was actually holding the wrong answer — the previous paragraph proposed it, or it is the obvious default in the field. Then the negation does work.

3. Significance designation

Tell: "the real X", "the actual question", "the part that matters", "the part that transfers", "the useful question", "the detail that makes the point", "the most interesting number", "the one thing nobody said", "here's the thing", "and that's the interesting part", "and it has a name here". Also the staged variants of the same gesture: "here's the twist / the catch / the kicker / the rub", "that's the part", "the part that makes me trust the rest", "my favourite part of".

Why: it ranks your own sentences for the reader instead of writing a sentence worth ranking first. The reader decides what is interesting.

Fix: delete the designation and state the content. If the content cannot carry itself, the problem is the content.

was: Nobody had checked whether it could still think. It is the leg that answers the actual question. now: Earlier testing measured its size and not its accuracy, so this run is the first accuracy number for it.

Watch for reuse. Emphasis is a budget. The same designation twice in one piece ("the part that matters" / "the part that transfers") is no longer emphasis, and reuse is the easiest instance to catch. declaude_lint.py groups its own hits and reports a reuse block for any construction used twice, so this one costs nothing to check.

4. Abstraction agency

Tell: a non-agent in the subject slot doing something — "the table shows it", "the median hides it", "the data tells us", "the failure modes live in different places", "the metric lies", "quantization is a concern", "truncation cuts compute", "the tool is about to discover", "the joke does the work an argument would have to".

Why: nothing in the sentence has hands. Nominalized verbs and inanimate objects get the subject slot, and the actual actor disappears.

Fix: put a real agent in subject position — a person, a named artifact, a concrete thing — and let the verb be a verb.

was: Median hides it: the middle of every distribution is similar. now: The middle of every distribution is similar.

was: Truncation cuts compute natively. now: A shorter vector is a shorter dot product.

5. Deferred noun

Tell: a placeholder where the noun belongs — "One thing is not flat", "The sixth is not", "There was just one problem", "That is the variable...", "What X means" — when the actual noun and number were available.

Why: points at the content instead of writing it, forcing the reader down a line to collect what could have been in the subject slot.

Fix: name it, with its number.

was: Five of those six rows are one cluster. The sixth is not. now: Five legs pass 77 to 81 of the 92 and exhaust 6 to 9. fit-17g passes 62 and exhausts 21.

6. Structural-metaphor locator

Tell: "the seam where...", "the hinge", "the joint", "the fault line", "the leg that...", "the place where it breaks down", "X is doing the work that Y would".

Why: makes an ordinary observation sound like architecture you uncovered. Usually stacks with abstraction agency.

Fix: name the plain mechanism.

was: It's the seam where the laugh is doing the work an argument would have to. now: The piece never engages the bull case; the joke covers the gap.

7. Coy and thesis-shaped headers

Tell: a header that states a verdict ("The default was wrong", "The one gap that does clear the bar") or hides its contents ("What 'exhausted' means", "What the data actually shows", "Why it matters"). Also any comma in a header — an appositive, participle or relative clause ("The obvious follow-up, measured"; "The catch, which is not small").

Why: headers are navigation. A header that could top three different sections is a thesis, not a label. No real person puts sentence clauses in a headline.

Fix: name the content as a flat noun phrase. Or, if the definition is the point, make the header the definition.

was: What "exhausted" means → now: Exhausted cases was: The one gap that does clear the bar → now: The 2.876-bit gap

Test: read the header alone, as a table-of-contents entry. Does it tell you what is in the section?

8. Suspense construction

Tell: a paragraph ending on a colon that introduces the culprit on the next line; a sentence that withholds the noun to force a break; "1, 2, and then 3 — the killer"; "Then the part that almost killed it."; "But here's where it gets interesting."

Why: suspense is a fiction technique. In technical writing it costs the reader time to buy the writer a beat.

Fix: state the finding when you have it.

9. Drama line break

Tell: a single sentence alone on its own line as a gravitas beat — "It isn't.", "That is not a result.", "And then everything broke."

Why: whitespace is being used as a drum roll.

Fix: put the sentence back in its paragraph. Earned for a real pivot: new actor, category shift, time jump.

10. Rhetorical question plus fragment answer

Tell: "So how does it fail?" / "Mostly by not finishing." — asking the reader something in order to answer it yourself.

Why: two beats where none are needed, and it stages the writer as a guide walking the reader through a discovery.

Fix: one declarative. "fit-17g fails mostly by not finishing."

Two in a row is the same tic amplified. Do I know how it works? Where it breaks? Which corners it cut? — the second and third are fragments riding on the first, and none of them is asked of anyone. The linter reports runs of consecutive questions separately from the single standalone one, because a run does not need the fragment answer to be the tic. An interview transcript and a FAQ are the earned cases.

11. Straw-man knockdown

Tell: quoting a wrong reading attributed to the reader, then rejecting it — "It thinks twice as long" is the obvious reading of that, and it is wrong.

Why: entry 2's negation-first shape with an invented interlocutor supplying the negated half. The correction is the same length without them. (This sentence used to read "the negation-first shape wearing a costume", which is entry 37.)

Fix: give the correction directly. "That median does not mean it thinks twice as long. On the questions it finishes..."

12. Aphoristic closer

Tell: a final sentence that compresses the paragraph into a contrast, moral, or quotable line — "It is the kind of number that looks like evidence and is not", "The bridge was the move, not novel math", "Verifier > judge by a wide margin", "the failure that looks like diligence", "X is the load-bearing part".

Why: it restates what the paragraph already established, in a shape built to be quoted. It also grades your own argument.

Fix: delete it. If the paragraph needs it, the paragraph is not working. Where a landing is genuinely wanted, land on a fact: "Default retries are back to 3."

13. Self-grading

Tell: "earned, not asserted", "this isn't just X, it's genuinely Y", "to be clear, this is rigorous because", "That is what the data shows", "That distinction changes what you would predict", "not a relabel", "worth noting", "a distinction worth keeping separate".

Why: narrates the epistemic status of your own reasoning instead of reasoning. If the point is earned, the reader can see it.

Fix: make the point and stop.

14. Performed humility

Tell: "Their phrasing for it is better than mine", "This might be a small thing, but", "I'm not sure this is worth writing up, but", "Probably nobody cares, but", "classic me".

Why: a bow. It asks for credit for modesty and weakens the piece.

Fix: if it is worth publishing, publish it straight. Attribution without ranking: "from a conversation with X".

15. Staccato fragment cadence

Tell: three or more fragments in series for rhythm — Six legs. One GPU, one server build, one sampler, one question set. — or parallel fragment pairs, Most questions it handles at a normal length. A subset runs away and hits the wall.

Why: rhythm engineering. The content is a list; the drumbeat is decoration.

Fix: one sentence. "Six legs, all on the same GPU with the same server build, sampler and question set."

16. Em-dash gotcha

Tell: a clause after an em-dash that delivers the punch, especially at the end of a paragraph.

Why: the dash stages the beat. Em-dashes for genuine inline asides are fine and human; the tic is the dash-as-drum-roll.

Fix: if the clause is the point, make it the sentence. Also: count em-dashes. More than roughly one per 150 words reads as machine-written regardless of what each one is doing. The linter reports the density and, separately, each dash whose clause runs to the end of its sentence, which is the drum-roll shape rather than the aside.

Demotion test: replace the dashes with commas. An aside survives it. A beat does not, because the beat was the punctuation.

17. Throat-clearing and process narration

Tell: "I want to talk about", "In this post, I'll cover", "Let me explain", "First, some background", "Before I get into it", "Let me consult my memories", "Storing this before I answer", "First I'll X, then Y".

Why: preamble wearing a competence badge. The reader sees the result, not the procedure.

Fix: start where you would start with no preamble budget.

18. RTFM as revelation

Tell: "It turns out that", bare sentence-initial "Turns out", "I finally discovered", "Hidden in the API", "Buried in the docs" — followed by standard documented behaviour.

Fix: if the finding is that you missed something obvious, say that.

19. Generic-developer vocabulary

Tell: footgun, shot itself in the foot, almost killed it, fell apart, rabbit hole, yak shaving, belt-and-suspenders, load-bearing (as metaphor), moving the needle, first-class citizen, under the hood, magic, just works, sane defaults, batteries included, zero config, small enough to fit in your head.

Why: the writer reached for the easy phrase instead of the accurate one. The accurate phrase is usually shorter.

20. Slop vocabulary

Tell: delve, tapestry, testament to, navigate the complexities, in today's fast-paced, it's important to note, landscape (figurative), realm, robust, seamless, leverage (as verb), utilize, crucial, pivotal, myriad, plethora, elevate, unlock, harness, embark, dive deep, at the end of the day, that said.

Why: these do not signal a tic so much as an absence of choice. Replace with the specific word.

21. Sanitized quotes

Tell: reported speech where the verb is "expressed", "indicated", "voiced concern about".

Fix: if they said "WTF", write "WTF".

22. Time-scale inflation

Tell: vague durations on recent work — "a month ago", "for a long time", "all year", "recently", "yesterday" — when the real timeline is hours or days.

Fix: check the timestamp, or drop the time framing. The pull toward narrative time is strong and needs an empirical counter.

23. Editorializing modifiers

Tell: adjectives that pre-load the verdict the piece is supposed to reach — "aggressive", "surprising", "impressive", "brutal", "collapse" for a 15-point drop, "exactly" where the contrast already carries it.

Fix: let the number be the adjective.

Encyclopedic and chatbot patterns

Entries 1 to 23 are staging mechanisms: a sentence built so the reader feels a finding arrive. Entries 24 to 36 are a different family. They come from the Wikipedia AI Cleanup project's Signs of AI writing, by way of the humanizer skill, and they show up in the flatter registers — encyclopedic summary, product copy, README boilerplate, pasted chat transcripts — where the writing is not staging anything, just running on defaults.

Several of these are surface patterns rather than mechanisms, which is a real limitation: a phrase list misses the next paraphrase, and entries 30 and 31 in particular are lists. They stay because they are cheap to check and they fire on text nobody would otherwise re-read.

24. Copula avoidance

Tell: "serves as", "stands as", "represents", "marks", "boasts", "features", "offers" where "is" or "has" would do.

Why: the elaborate verb adds a beat of ceremony and usually a claim the source does not make. "Serves as" implies purpose; "is" states a fact.

Fix: use the copula.

was: Gallery 825 serves as LAAA's exhibition space. The gallery features four separate spaces and boasts over 3,000 square feet. now: Gallery 825 is LAAA's exhibition space. It has four rooms totalling 3,000 square feet.

25. Participle tail

Tell: a clause tacked onto the end of a sentence with a present participle — "highlighting", "underscoring", "emphasizing", "reflecting", "symbolizing", "showcasing", "ensuring", "fostering", "contributing to", "encompassing".

Why: the tail asserts significance the sentence has not earned, and it is almost always unsourced. Two or three stacked in one sentence is the strongest single tell in this family.

Fix: cut the tail. If the claim inside it is real and sourced, make it its own sentence.

was: The temple's palette resonates with the region's natural beauty, symbolizing Texas bluebonnets and the Gulf of Mexico, reflecting the community's deep connection to the land. now: The temple is painted blue, green and gold, colours chosen to evoke Texas bluebonnets and the Gulf of Mexico.

Earned when the clause carries a sourced claim. Then it is the shape that is wrong, not the content: promote it to its own sentence rather than cutting it.

26. Forced triad

Tell: three parallel items where the content has two or five — "keynote sessions, panel discussions, and networking opportunities", "innovation, inspiration, and industry insights".

Why: three reads as complete, so the model pads to three or truncates to three. The count is chosen by rhythm rather than by the content.

Fix: list what there is. Two items is a normal number.

was: The event features keynote sessions, panel discussions, and networking opportunities. now: The event has talks and panels, with time between them to talk to people.

Earned when there are genuinely three things, or a superlative or ranking rides on the phrasing.

Three shapes, one tic. The comma list ("keynote sessions, panel discussions, and networking opportunities") is the obvious one. The anaphora form carries the parallelism in the opening words instead: It was a message. It was a permission slip. It was an out. across sentences, or something to steer toward, something to be measured against, something to fear inside one. The echo form keys on neither the commas nor the opening, but on a skeleton reused whole: A shopping cart is an object in the system. A chat room is an object in the system. Two sentences sharing four or more words in sequence are a template being filled, and a reader who has read the first has read the second. The linter catches all three, and the document-level opening repetition count is the same tic spread thin enough that no single passage looks wrong. Entry 27 read in the right direction is the guard: repetition that is the point stays.

27. Elegant variation

Tell: the same referent renamed on every mention — protagonist, main character, central figure, hero; the model, the system, the network, the architecture.

Why: repetition penalties push the next token away from the word already used. Human technical writers repeat the noun, because a second name for a thing reads as a second thing.

Fix: pick the noun and keep it. Repetition is clarity here, not a defect.

was: The protagonist faces many challenges. The main character must overcome obstacles. The central figure eventually triumphs. now: The protagonist faces many challenges and eventually wins.

Earned when the second term names a genuinely different thing. Check that the referents really are identical before collapsing them.

28. False range

Tell: "from X to Y" where X and Y are not endpoints of any scale — "from the Big Bang to dark matter", "from startups to enterprises", "from onboarding to retention".

Why: the construction promises a spectrum and delivers two examples, so it implies coverage the sentence does not have.

Fix: list the items, or name the actual range with its actual endpoints.

was: Our journey has taken us from the singularity of the Big Bang to the grand cosmic web, from the birth of stars to the dance of dark matter. now: The book covers the Big Bang, star formation and current theories about dark matter.

Earned when the endpoints are real: "from 2 to 8 bits", "from onboarding to offboarding". A range with a scale under it is a range.

29. Inline-header list

Tell: bullets that open with a bolded label and a colon, where the label restates the first words of the item — - **Performance:** Performance has been improved.

Why: it is an outline rendered as prose. The labels carry no information the sentences do not, and the shape survives from the model's plan into the output.

Fix: write the paragraph, or cut the labels and keep the bullets, or keep the labels only where they are a real index (a term being defined, an option name).

was: - Performance: Performance has been enhanced through optimized algorithms. now: Load times dropped after the index moved off the hot path.

Earned when the label is a real index rather than a restatement: a term being defined, an option or flag name, a case name in a table of cases.

30. Typographic tells

Tell: bold scattered mid-sentence on phrases that are not terms; Title Case On Every Word Of A Heading; emoji as bullet or heading decoration; curly quotes in a plain-text or code context.

Why: none of these is wrong on its own, and each has an innocent source (Word, a CMS, a house style). They matter as a cluster, and they matter most in documents where the surrounding convention is clearly different.

Fix: bold for terms on first use and nothing else; sentence case in headings; no emoji unless the document already uses them; straight quotes in anything a program will read.

Not a tell on its own. Curly quotes are the default in most editors, and one bolded term is normal writing.

31. Chatbot residue

Tell: the conversational wrapper left in the pasted text — "Great question", "Certainly", "You're absolutely right", "I hope this helps", "Let me know if you'd like", "Would you like me to", "Want me to give examples", "Here is an overview of".

Why: it is correspondence, not content. It reaches the document because someone pasted a reply rather than an artifact.

Fix: delete. There is no rewrite; the sentence has no content.

32. Filler and hedge stacking

Tell: phrases that expand without adding — "in order to", "due to the fact that", "at this point in time", "in the event that", "has the ability to", "it is important to note that"; and hedges in series — "could potentially possibly", "might arguably suggest".

Why: each is a token-cheap way to sound careful. Stacked hedges do not make a claim more careful, only harder to hold to.

Fix: "to", "because", "now", "if", "can". One hedge carries all the uncertainty three of them do.

was: It could potentially possibly be argued that the policy might have some effect on outcomes. now: The policy may affect outcomes.

Do not overcorrect. A hedge with content ("on the two runs that finished") is a caveat and stays. See the Overcorrection section of SKILL.md.

33. Speculative gap-filling

Tell: "while specific details are limited", "based on available information", "not publicly available", "maintains a low profile", "keeps personal details private", "likely grew up", "it is believed that", "as of my last update".

Why: two tells with one cause. The model cannot find a source, writes a sentence about not finding one, then fills the gap with the stock guess. The guess is unsourced and the meta-sentence is about the model, not the subject.

Fix: say what is not known, or cut the sentence. Never dress the guess as fact.

was: Information about her early life is not publicly available, suggesting she maintains a low profile. She likely grew up in a middle-class household. now: Her early life is not documented in the sources used here.

34. Diff-anchored documentation

Tell: documentation or a code comment that narrates a change rather than describing the thing — "this function was added to replace", "we now use", "the previous approach was", "this has been updated to".

Why: the document is being written from the diff that produced it, so it reads coherently only to someone who knows what the last commit did. Six months later nobody does.

Fix: describe the current state. The change belongs in the commit message or the changelog.

Earned when the document is version-scoped by design — a changelog, release notes, a migration guide, an upgrade path. Narrating the change is the job there.

was: This function was added to replace the previous approach of iterating through all items, which caused quadratic performance. now: Looks up items through a hash map, so cost is constant per item rather than quadratic in the list length.

35. Subjectless fragment

Tell: a claim with the actor removed — "No configuration file needed", "The results are preserved automatically", "Changes applied on save".

Why: the reader cannot tell who does the thing, which matters most in exactly the documents where these appear: who preserves the results, and can I rely on it? Related to abstraction agency (entry 4), which puts the wrong subject in the slot rather than none.

Fix: name the actor.

was: No configuration file needed. The results are preserved automatically. now: You do not need a config file. The runner writes results to out/ when the job exits.

Earned when the register is genuinely clipped throughout — release notes, a feature table, a CLI help string.

36. Predicate-position hyphenation

Tell: a compound modifier keeping its hyphen after the noun — "the report is high-quality", "the team is cross-functional", "the pipeline is end-to-end".

Why: the model hyphenates the pair uniformly wherever it appears. Human writers hyphenate attributively and mostly drop it in the predicate.

Fix: keep the hyphen before the noun, drop it after.

was: The team is cross-functional and the report is data-driven. now: The team is cross functional and the report is data driven.

Minor. One instance proves nothing; a document that hyphenates every pair in both positions is a cluster member.

37. Dressed metaphor

Back in the staging family, not the encyclopedic one, and numbered last because it arrived last.

Tell: a figure of speech standing in for a mechanism you could have named — "wearing the costume of", "dressed up as", "in a new hat", "is X in Y's clothing", plus visceral or lurid imagery on a mundane observation. Entry 6 is the special case where the figure is locational ("the seam where"); this is the general one.

Why: the figure feels vivid, so it reads as insight, and it usually replaces the plain description rather than adding to it. The accurate phrase is almost always shorter and less pleased with itself.

Fix: describe the mechanism.

was: That is information loss wearing the costume of a style fix. now: The edit removes content while looking like it removed only style.

was: every time we pushed on a mechanism the hard part squirted out. now: each mechanism we examined had the same difficulty underneath it.

Earned when the metaphor is the clearest available description and no plain noun fits. Delete the ones doing significance work, not the ones doing work.

No regex reaches this one. declaude_lint.py cannot see it, and a phrase list for it would be three entries long and obsolete by the next draft. It is caught on the sentence pass or not at all. Both specimens above shipped past a clean lint report.

38. Announce-then-deliver

Tell: a counted noun-phrase label with a clipped qualifier, standing before the content it names — "One factual note, not fixed:", "Two things I did not do.", "One judgment call left alone.", "The exception, worth restating:", "and worth saying why it works:".

Why: the label carries nothing the content does not. It buys the writer a beat to compose in and costs the reader a clause before the subject arrives. A person either writes the label the way they would say it — "One note I did not fix" — or, more often, just states the thing.

Distinct from entry 5, where a placeholder sits in the subject slot; here a whole announcement sits in front of the sentence and the real subject is fine. Distinct from entry 17, which narrates the writer's procedure rather than labelling the content. Entry 29 is the bulleted cousin: same restating label, rendered as a bold lead-in instead of a standalone phrase.

Fix: delete the announcement and start with the content. Where a label is genuinely needed, write it as speech, not as a caption.

was: One factual note, not fixed: X and Y both assert what people generally do. Nothing in the repo supports either. now: Two sentences assert what people generally do, and nothing in the repo supports either: X and Y.

Earned when the count is real navigation over a list the reader will scan — "Three shapes, and the fix differs for each:" ahead of an actual three-item list. The test is whether deleting the label loses anything.

Watch the reuse. Measured nine times in one 40 KB document, including twice as a section header, in a piece whose own subject was bad headers. A writer who reaches for this once reaches for it throughout; the linter's reuse block is the cheapest way to see it.

39. Welded epigram

Tell: a factual clause joined by and or so to a second clause that restates it as a general truth — ", so a child never carries a grant its parent lacks", ", and the failure it prevents is silent", ", and nothing errors". The second clause typically contains no concrete noun and would read as true of any system, not this one.

Why: the first clause says what happens; the second says what it means, in a shape built to be quoted. It also loses precision, because a maxim has to drop the conditions that made the mechanism specific. A child never carries a grant its parent lacks is vaguer than the rule it paraphrases, and a reader cannot act on it.

Fix: keep the clause that says what happens. Delete the one that says what it means.

was: Entries the caller does not itself hold are dropped, so a child never carries a grant its parent lacks. now: List a tool you do not have yourself and it is dropped.

was: The rest of the surface depends on this parameter, and the failure it prevents is silent. now: Leave it out and nothing errors — the child just comes up bare.

Distinct from entry 12, which is a whole sentence at the end of a paragraph. This one hides inside a single sentence, mid-paragraph, and survives a pass that is only checking closers.

40. Spec-ese

Tell: the register of a standards document rather than of a person explaining something. Demonstrative subject where "it" would do ("That child holds no repository"); reflexive emphatics ("does not itself hold"); latinate state verbs for a program doing nothing ("idles", "holds", "carries", "awaits"); participle tails on those verbs ("awaiting input"); passive with the actor displaced into a subordinate clause.

Why: it reads as translated from a schema. The formality signals precision without adding any — every specimen below says less than its plain rewrite, because the plain rewrite names who does the thing.

Fix: use "it". Use the verb you would say aloud. Name the reader as the actor where the reader is the actor.

was: That child holds no repository and waits for input. now: It comes up with no repo and does nothing until you send it a message.

was: Omit it and the child idles awaiting input. now: Leave it out and the session just sits there.

Earned when the term is the system's own — a queue whose documented states are waiting and held, a flag literally named AWAIT. Then it is a quotation, not a register choice.

41. Nominalized header

Tell: a header whose head noun is an abstraction — "GitHub authorship in a sourced child", "Spawn availability", "Channels between siblings", "Detecting a parked child". Shapes to scan for: <-ship/-tion/-ment/-ance/-ity noun> in|of|for <thing>, and a bare gerund standing in for the sentence the section actually makes.

Why: it names a topic area instead of the content, so it fails entry 7's table-of-contents test while passing entry 7's regex. A reader scanning "Spawn availability" learns that the section is about spawning, which they knew.

Fix: name the concrete thing, or ask the reader's question.

was: GitHub authorship in a sourced child → now: Who the delegate posts as was: Spawn availability → now: When create_session fails

This is the standard overcorrection from entry 7. Diagnosed 2026-08-23: a draft's four verdict-shaped section headers ("A sourced child signs GitHub writes as the token owner", "There is no reply channel between siblings") were all rewritten into nominalizations in a single pass, and the linter then reported the document clean. Fleeing a verdict into an abstraction is not a fix. The target is a concrete label, and both failures miss it in opposite directions.

42. Contents-list standfirst

Tell: a subtitle or opening line built as <noun phrase>, plus <noun phrase> — "The create_session parameter surface, plus the behavior the schema leaves out". Variants: "…, and what it means for X", "…, and the N things you need to know". The second half is usually a coy reduced relative that claims value without naming any ("the behavior the schema leaves out", "the part the docs skip").

Why: it is a table of contents wearing a sentence, and the second half is a significance tag (entry 3) in disguise — it tells the reader the withheld material is worth having rather than saying what it is.

Fix: say what the page contains, or drop the subtitle. A reference does not need one.

was: The create_session parameter surface, plus the behavior the schema leaves out. now: Every parameter, and four behaviors that only show up in practice.

Flat-certainty patterns

Entries 43 to 47 are the register this skill produces, once the staging is gone. They come from a count rather than from a draft: the cluster of GitHub pull request descriptions that went from 0.70% of early 2025 to 39.5% of August 2026, whose highest-lift words are load-bearing, plainly, quietly, refusal, re-derived, asserted, nobody, genuinely, outright, byte-identical. references/corpus.md carries the method, the numbers and the limits.

Nothing here is staging a reveal. These sentences are trying to sound settled — adjudicated, verified, exhaustively negated — and the same test applies one register over: am I stating the finding, or performing having settled it?

Every family below is also a family a careful writer uses correctly, which is why each entry has an earned column. Do not treat any of these words as a blocklist; references/corpus.md explains what the measurement does and does not support.

43. Flat-certainty adverb

Tell: a bare adverb doing the work of the evidence — "plainly", "quietly", "outright", "merely", "genuinely", "deliberately", "provably", "empirically", "demonstrably", "structurally", "legitimately", "precisely", "honestly", "vacuously", "adversarially", "silently", "loudly", "routinely", "identically".

Why: plainly and quietly are the second and third highest-lift words in the corpus cluster. The adverb asserts the reader's reaction — that this is obvious, that it happened without noise, that the check was real — where the sentence has not shown it. "Provably correct" without the proof is "correct" with a costume on. It is entry 23 in the flat register: the modifier pre-loads the verdict.

Fix: delete it, or replace it with what makes it true.

was: The retry is provably safe. now: The retry is idempotent: the handler keys on the request id.

was: The flag was quietly dropped in 4.2. now: The flag was dropped in 4.2. The changelog does not mention it.

Earned: the adverb names a real contrast the reader can check — "silently" against a version that logs, "identically" against a version that differs. One adverb carrying a comparison is information; three in a paragraph is a register.

44. Juridical register

Tell: code, tests and processes described as a court — "refusal", "refuses", "declines", "admits", "ruling", "verdict", "precedent", "carve-out", "standing", "grounds", "obligation", "owed", "remedy", "ratified", "sanctioned", "honoured", "answerable", "cites".

Why: refusal is the fourth highest-lift word in the cluster and ruling the seventeenth. The register borrows finality from a domain that has procedures for producing it. A linter does not rule; it exits non-zero. A default is not a carve-out unless something granted it. The vocabulary makes an implementation detail sound adjudicated, and it stacks with entry 4 — the thing doing the ruling is usually an abstraction with no hands.

Fix: name the mechanism and its exit condition.

was: The guard refuses any caller without standing. now: The guard returns 403 when the token has no write scope.

was: That is the carve-out the header check honors. now: The header check skips one-word titles.

Earned: the system's own documented vocabulary. A policy engine whose API says deny denies; an RFC that says MUST is an obligation. Use the word the system uses, not the word that sounds like it.

45. Verification-provenance compound

Tell: a hyphenated compound asserting how thoroughly something was checked — "byte-identical", "bit-identical", "byte-for-byte", "byte-exact", "mutation-checked", "mutation-verified", "mutation-tested", "live-verified", "cross-checked", "root-caused", "hand-verified", "re-derived", "re-verified", "re-measured", "re-checked", "re-confirmed".

Why: eleven of the cluster's top hundred words are this shape, and it is the newest family in the register — it arrived with agents that report on their own work. The compound is a claim about method compressed to an adjective, which puts it past the place a reader would ask for the method. "Byte-identical" is checkable and often true; "mutation-checked" usually means a mutation run happened, not that this line survived one. The re- prefix is the same move in a verb: it claims a second pass without saying what the second pass did differently.

Fix: state what was run and what came back.

was: The rewrite is byte-identical and mutation-checked. now: diff reports no change. The mutation run killed 14 of 15; the survivor is in parse_header.

was: I re-derived the threshold. now: I recomputed the threshold from the 46 chunks and got 0.20 again.

Earned: the compound is the precise term and the number is beside it. "Byte-identical" against a stated diff is exact and shorter than the alternative. A re- verb earns its prefix when the first pass is on the page and the second disagreed with it.

46. Privative coinage

Tell: a thing named by what was not done to it — "ungated", "unguarded", "unwired", "uncapped", "unbuilt", "unparseable", "unverifiable", "unmeasured", "unresolvable", "unsatisfiable", "unrecognised", "vacuous", "inert".

Why: the coinage sounds like a diagnosis and carries no measurement. "The hook is unwired" and "the hook is not in settings.json" say the same thing, but the first sounds like a category the writer already knew and the second can be checked. Several of these are not words outside this register, which is the tell: the writer minted an adjective rather than describe a state.

Fix: say what is missing, and where.

was: Three of the paths are unguarded. now: Three of the paths have no length check before the index.

was: The assertion is vacuous. now: The assertion passes on an empty list, which is what the test supplies.

Earned: the term is the field's ("unsatisfiable" in SAT, "unbuilt" of a build target), or the absence is the finding and the entry names it once before the detail.

47. Exhaustive negation

Tell: an absolute quantifier standing in for the argument — "nothing in it does X", "nowhere else for it to be", "no path reaches", "never", "neither", "nobody". Often in a closer, often as the whole of the evidence.

Why: nobody, nowhere and nothing are all in the cluster's top twenty-five. A universal negative is the strongest available claim and the most expensive to establish, so it is the cheapest to assert. "Nothing else could cause it" is a claim about the search, and the search is what the reader wants. Entry 2 covers the negation that sets up a reveal and entry 3 the unfalsifiable "nobody noticed"; this is the flat form, where the negation is the finding.

Fix: state the scope you actually checked.

was: Nothing in the fit varies with time. now: The fit has one parameter per cluster and none per week.

was: There is nowhere else for the rise to be. now: The weekly shares are counted after assignment, so a rise is in the assignments.

Earned: the scope is bounded and named. "None of the 46 chunks exceeds 0.57%" is a universal over a set the reader can see. So is a negative that follows an exhaustive mechanism — "the enum has three members and the switch covers all three".

48. Performative candour

Tell: "I'll be honest", "let's be honest", "to be clear", "I won't pretend", sentence-initial "Honestly," / "Look," / "Frankly,", and the invitation that follows them — "you don't have to take my word for it".

Why: sincerity announced instead of shown. The reader had no reason to suspect the preceding sentences, so the announcement marks the next one as the true one and demotes the rest. "Don't take my word for it" is the same move aimed at evidence: it offers a check and hands over nothing to check with.

Fix: delete the announcement and keep the sentence. Where the invitation to verify is real, link the thing.

was: Let's be honest: the benchmark does not measure what the README claims. And you don't have to take my word for it. now: The benchmark times decode only. The README claims end to end. The harness is bench/run.py, line 40.

Earned: reported speech, where a person said it. Also "to be clear" introducing a correction of a misreading the draft itself caused, once.

49. Stranded auxiliary reversal

Tell: a clause that ends on a bare auxiliary carrying the reversal — "The tool died; the data didn't.", "Reading mostly passed. Writing didn't.", "Maybe it wouldn't have."

Why: the auxiliary does the work and the verb is elided, so the clause lands as a beat rather than as a fact. Entry 2 stages a reveal by negating first; this stages one by negating last, with the content left out of the half that is supposed to carry it.

Fix: say what the second clause is claiming.

was: The tool died; the data didn't. now: The tool stopped writing checkpoints on 12 August. The rows already written are still in the bucket.

Earned: ordinary ellipsis, where the elided verb is the one just used and the contrast rides in the same breath rather than in a beat of its own — "Digression and mild informality are human. Symmetry and antithesis are not." Also where both halves are measured and the reader can see both numbers: "reads passed on all 12 shards, writes on none". The tic is the reversal given its own sentence as a drop, with the verb withheld.

50. Retroactive significance

Tell: "that's why X mattered", "this is why keeping every transcript counted", "which is why the open environment mattered".

Why: entry 3 tells the reader which sentence matters before they read it; this tells them which one did. The retroactive form is the worse of the two, because the passage being graded has already had its chance with the reader and the grade is an admission it did not take.

Fix: cut it. If the earlier passage did not carry the point, the earlier passage is the problem.

was: That's why being able to open the environment mattered. now: Opening the environment recovered 40 GB of checkpoints. Without it the run restarts at step 0.

Earned: a "which is why" that introduces a consequence the reader has not seen yet. New information, not a grade on old information.

51. Totalizing designation

Tell: "that's the whole point", "the entire business model is X", "here's the whole trick", "the only marketing I trust", "the only thing that matters".

Why: entry 3 with the scope widened to everything. "The whole point" claims the rest of the paragraph is decoration. "The only X I trust" ranks a field whose other members are never named, so there is nothing to disagree with. Neither can be checked, and both ask the reader to accept a total where a part was demonstrated.

Fix: state the part, and let it be a part.

was: That's the whole point of the format. now: The format keeps the offsets in the header, so a reader can seek without decompressing.

was: Changelogs are the only release notes I trust. now: I check release notes against the diff, which per-entry links make possible.

Earned: a real count of one over a named set — "the only one of the six runs that finished".

52. Obituary headline

Tell: "X is dead", "long live X", "the death of X", "RIP X" — usually a header or a first line.

Why: a headline built to be argued with rather than read. It states a verdict, which entry 7 already bars in headers, and it borrows a form whose function is to overstate. The body then spends its opening paragraph walking the claim back, which is the tell that the header was never the finding.

Fix: name what changed.

was: Peer code review is dead now: Three of our four teams dropped the second-approval requirement in July

Earned: the phrase belongs to something quoted or named — a post you are citing, a product whose vendor announced end of life.

Quick self-check before shipping an edit

1. Does any sentence exist to make a finding feel bigger than it is?
2. Does any sentence tell the reader which sentence matters?
3. Is anything in a subject slot that cannot act?
4. Does any header state a verdict, hide its contents, contain a comma, or name an abstraction instead of the content?
5. Is any noun deferred that was available?
6. Does the last paragraph paraphrase the subtext of the piece?
7. Did the edit remove content, change a claim, or add hedging?
8. Does the result still sound like a person with opinions?
9. Does any sentence end in a participle tail that asserts significance?
10. Does the rewrite contain a fact, name, number, date or citation that is not in the source? A fabrication is a defect even when it sounds more human than the vague original.
11. Does any sentence weld a maxim onto a fact with "and" or "so"?
12. Would you say it aloud in those words, or is it schema prose?
13. Does the rewrite drop a claim the source made? Check claim by claim, not paragraph by paragraph. Superlatives, rankings, simultaneity, scope words and the condition attached to a hedge all sit inside phrasings this register cuts, and the paragraph looks intact after they go.
14. Does an adverb, a hyphenated compound or an absolute negative carry a claim the sentence never shows? Flat certainty has tics of its own; entries 43 to 47 and references/corpus.md.
15. Does any sentence announce its own honesty, grade an earlier passage, claim a whole where a part was shown, or land a reversal on a bare auxiliary? Entries 48 to 52.