An AI writing catalog should keep its history

A pattern catalog should become more useful as it gets older. An entry that rarely helps with today's drafts may still matter to an editor reviewing a 2023 article or a researcher tracing how generated prose changed. Delete the entry, and the next reviewer loses the observation, its examples, and the reasons people accepted or challenged it.

The synthesis content-quality catalog, which supplies methodology to Slopcheck, follows a compounding-archive principle: retire patterns without erasing their history. New observations accumulate alongside corrections and counterexamples. I want the catalog to preserve what we learn, including what we learn from being wrong.

Revision prepared September 28, 2026. The catalog counts below come from its June 2026 source; the numerical-calibration correction comes from the maintainer's August 2026 record.

Keep the earlier observation available

The archive distinguishes Active, Declining, Historical, and Deprecated entries. These labels record a pattern's status in the catalog. An entry can move out of current use while retaining its identifier, dated examples, and the explanation for the change. A superseded entry can point to its more precise replacement.

The historical catalog retains examples such as the “As an AI language model” preamble, Bard's “as a large language model trained by Google” stem, early Claude disclaimers, refusal templates, and emoji bullets. These are useful places to begin a historical inquiry. Their presence in a text does not establish which product produced it, and an era label does not establish that a particular training update removed it.

A newsroom revisiting older articles needs the observations and sources available at the time, alongside later corrections. A scholar comparing language across model generations needs dated examples rather than a catalog that silently replaces its past. A regulator examining an older disclosure dispute needs a trail back to the underlying records. Keeping the archive makes those investigations possible.

Separate the date a passage was written from the date it was reviewed. If the production date is unknown, record that. If a model or product version is known from a retained source, include it as context. Inferring a version from punctuation or vocabulary would turn the observation into an unsupported provenance claim: a claim about how the text was produced.

The archive can also preserve the history of editorial judgment. Why did a reviewer once recommend an edit? What counterexample changed that recommendation? Which earlier articles were reviewed under the old rule? Those questions remain useful even when the original rule turns out to be wrong.

Give the catalog an architecture

The version 4.0 catalog organized the work into three sections and a shared research layer. Its June 2026 source contains 108 named family-pattern entries, 17 substance tests, 76 general criteria, and 86 combinations.

A1, model-family patterns, organizes observations associated with Anthropic Claude, OpenAI GPT, Google Gemini, Meta Llama, xAI Grok, DeepSeek, Mistral, and Qwen. The practical research question is how output differs across known producers, versions, tasks, and interfaces. Preserve those conditions with the examples. They allow someone else to investigate an observation without guessing a producer from a word or a punctuation mark.

A2, substance and depth, asks what the writing gives its reader. Does a sentence carry a claim, evidence, a useful distinction, an opinion, or a needed transition? Does an apparent insight say anything specific about its subject? The deletion and any-company tests belong here, along with tests for generic insights, unsupported profundity, and conclusions that never reach a conclusion. For a quick pass, mark the sentences on which the argument depends. Three consequential claims per hundred words is a useful prompt for dense explanatory prose, not a quota for every kind of writing.

A3, editorial criteria, groups problems by the work an editor needs to do: language and tone; style and structure; technical formatting; citations and sourcing; context; hyperbole; confidentiality; behavior and tone; frame and audience; and social register. Stable identifiers such as A3-SS-001 let a discussion point to an exact entry. A citation problem and a repetitive sentence need different investigations even when both make the article worse.

The B layer connects entries to proposed causes, combinations, and calibration research. Its twelve causal codes distinguish hypotheses about reward shaping, training-data sources, safety tuning, refusal behavior, helpfulness, system prompts, tokenization or architecture, and product wrappers. Keeping these explanations separate is useful: a phrase introduced by a chat interface is a different problem from a recurring weakness in an article. A causal label is a hypothesis until evidence isolates the cause.

The distinction between an article and a complete chat response also belongs in the design. A published artifact may omit an assistant's greeting and closing offer of help. A full transcript includes them. Record whether an example comes from the opener, body, closer, a mid-body insertion, or several zones; otherwise a reviewer can compare unlike material without realizing it.

The intellectual work behind the tests

The substance section draws on a longer inquiry into language that sounds meaningful without doing the work it appears to do. Harry Frankfurt's On Bullshit distinguishes indifference to truth from lying. Hicks, Humphries, and Slater's 2024 paper applies that philosophical account to large language models. That question about fluent output and truth informs the catalog's substance checks.

Pennycook and colleagues' 2015 work studied how people receive apparently profound but vacuous statements. Padmakumar and He's 2024 study found reduced diversity in the argumentative essays co-written with InstructGPT in their experiment, but not in the GPT-3 condition. Sourati and colleagues' review brings together research on homogenization of expression and thought.

These contributions help explain why an editor should examine specificity, meaning, and distinctive voice alongside fluency. They do not turn a preferred prose style into a scientific test. I want the catalog to help writers express something worth reading, including an ambitious argument or an unusual perspective that a formulaic rewrite would flatten.

Start with the passage and its evidence

Consider an illustrative source note: “The preview shows a stale title. The cause has not been established.” Those are the only facts available for the following edit.

The draft says:

This issue requires careful consideration. The preview shows a stale title. The cause has not been established. This highlights the importance of taking a thoughtful approach.

The repair is:

The preview shows a stale title. The cause has not been established.

The removed sentences neither describe the problem nor help the reader choose an action. The edit preserves both the observation and the uncertainty. A catalog entry could call this an empty announcement, but the label is less useful than the explanation and the exact deletion.

The exception belongs in the same entry. An opening sentence can orient readers before unfamiliar evidence; a transition can explain why one result leads to the next question. Deleting those sentences merely because they contain no new fact can make the argument harder to follow. Read the passage with and without the sentence before deciding.

A second illustrative source note says: “The parser passed one test input. No other inputs were tested.”

The draft says:

The parser is now reliable.

A supported revision says:

The parser passed the tested input. Its behavior on other inputs has not been checked.

This edit narrows the claim to the evidence. Adding an error rate, a larger test suite, or a story about a previous failure would create new facts. Those additions need their own sources. Replacing vague prose with invented specificity makes the passage less trustworthy.

For an illustrative business sentence such as “The company is committed to excellence,” try substituting another company's name. If the paragraph promises to explain this company's achievement, the sentence has not done that work; ask for the practice or result the writer means to describe. As a stated aspiration, it can belong in a values statement. Definitions and standard instructions may also be useful precisely because they apply across organizations. The test has to respect the job the passage is doing.

The same discipline applies to common structural advice. A repeated conclusion may waste space in a short essay and help readers navigate a long reference document. Bold labels may make a repetitive argument look organized, yet be exactly what a checklist needs. A contrast is useful when the distinction changes a decision. The editor has to explain the effect in this passage.

Specify an entry someone else can challenge

The original design specified a fourteen-field template. It remains a useful maintenance contract, with the old numerical fields interpreted according to the calibration correction:

FieldWhat the record needs
idA stable reference to the entry, such as A3-SS-001
nameA human-readable label
descriptionAn operational definition, the editorial problem, and a proposed repair
eraActive, Declining, Historical, or Deprecated, with a reason for the status
years_activeDated observations and the bounds of the period examined
familiesKnown producer context for the examples; leave unknown provenance unknown
zoneWhether the pattern appears in the opener, body, closer, or another defined part of the response
signal_strength_when_presentAn ordinal research score and its evidence status
base_rate_by_familyAn occurrence estimate tied to its sample, model, task, interface, and date, or explicitly marked unvalidated
causal_attributionA proposed explanation and the evidence needed to distinguish it from alternatives
examplesExact passages, their sources and dates, permission to retain them, and the observed problem
non_examplesSimilar-looking passages where the advice would be wrong or harmful
combined_signalsRelated combinations to investigate, with their evidence status
notesEdge cases, language and genre considerations, review decisions, and supersession history

For a proposed model-behavior entry, ask the contributor for at least three sourced examples and the conditions under which they were obtained. Three examples make an observation inspectable; they do not establish its frequency. An invented example can teach an editorial principle, as the examples above do, but it must remain labeled illustrative.

A reviewer can then challenge the entry: do the examples meet its definition? Does the counterexample expose an overbroad rule? Was the source attributed correctly? Does a claimed rate have the denominator it needs? Accept the contribution, request changes, or reject it with an explanation. A completed template starts that review; it does not automate acceptance.

Keep frequency separate from diagnostic value

A common pattern and a revealing pattern are different things. The original two-axis design separated how often a pattern appeared from how much it was supposed to tell a reviewer.

The maintainer's August 2026 correction says the inherited false-positive estimates were not produced from a preserved labeled corpus. It treats the old signal-strength values as ordinal research scores and the base-rate values as unvalidated estimates unless their measurement conditions are supplied. The earlier below-0.5% and below-1% claims therefore cannot serve as measured performance.

The combinations remain available for investigation. B2-COMBO-003 brings together em-dash density, bolded list leads, and uniform paragraph length; B2-COMBO-001 combines saturated vocabulary, exhausted metaphors, and a section-ending summary. An editor can inspect what those features do to a passage. Several clues may share one explanation, such as house style, so their co-occurrence does not by itself produce an authorship probability.

Research on changes in writing can inform a catalog without determining an individual article's origin. In a study published in 2025, Kobak and colleagues examined vocabulary changes in biomedical abstracts and estimated that at least 13.5% of 2024 abstracts in their analysis had been processed with large language models. That is a population estimate from a particular body of writing. It does not mean that 13.5% of abstracts contained one identifying word, or that a particular writer's use of that word establishes AI use.

Liang and colleagues' 2023 study found that the detectors they tested misclassified writing by non-native English speakers. It is a reason to scrutinize an authorship claim and its consequences. It does not supply a false-positive rate for a different catalog or establish that adding another stylistic marker makes a judgment safe.

An editorial system should protect writers from that error. A style rule that penalizes their language background would make the review worse. The catalog's negative-marker entry, B2-COMBO-010, preserves this concern about uniform paragraphs, restricted vocabulary, and heavy transitions. Keep such counterexamples prominent and judge the safeguard against the errors it prevents on relevant human writing.

Let the learning accumulate

The public repository gives contributors a place to inspect entries, propose corrections, and retain earlier versions. The content-quality skill declares CC0-1.0 licensing; Slopcheck's software has an MIT license. Keeping the methodology reusable allows a newsroom, researcher, or tool builder to adapt it and challenge it. I want writers to have access to this editorial help through a browser, a coding agent, or their own tools.

Slopcheck assembles catalog material into its review prompts. Its source includes a web interface, command-line tool, browser extension, email handler, and configuration material for GPT and Claude surfaces. Readers can inspect those adapters and choose an integration with the setup and provider requirements their workflow needs. The catalog is also packaged in the Agent Skills format, whose instructions and references can be loaded by supporting clients.

For the request path and implementation, see the companion Slopcheck architecture article. The writing companion develops the editorial purpose. The catalog itself gives other tool builders the definitions, examples, history, and research questions needed to develop their own approach.

Keep the original passage beside the accepted edit. Record disagreements and cases where the advice failed. When a later review changes the recommendation, preserve the earlier reasoning with its date. An editor returning to that passage can then see which facts were available, which judgment changed, and why.

I want each new editor to inherit the earlier editor's reasoning, challenge it, and leave a better record. A catalog that remembers its mistakes can improve through use instead of asking every generation of writers to start over.