Changelog

Building a Second Brain (BASB) (Tiago Forte) — #892: the method that holds the catalog’s two Forte anchors together. Asked what to do with scattered notes before a talk, or with 120 Kindle highlights, the full name produced the CODE cycle in 12 of 12 runs, Progressive Summarization in 12 of 12 and Intermediate Packets in 11 of 12, on every tier. Tier ★★★.

CODE is an alias, not a second entry. "CODE Method (Capture, Organize, Distill, Express)" resolves to Forte on every tier, but every model explains it through BASB, and as an instruction it is a subset: on Haiku it gave the four steps without Progressive Summarization, PARA or Intermediate Packets.

"Second Brain" alone is a name trap. Every tier knows it is Forte’s, but "Set up a second brain." produced CODE only on Opus; Sonnet built PARA folders without CODE. The catalog name is "Building a Second Brain"; the bare term is a tag. "BASB" works on all tiers.

PARA does not cover BASB. "Use the PARA method." produced the four buckets in 12 of 12 runs and rarely went further: CODE once and Progressive Summarization twice, all on Opus; layered highlighting twice; Intermediate Packets never. P.A.R.A. Method now points to BASB as the method it belongs to and drops "Second Brain Organization System" from its aliases.

The Second Brain Maintenance contract (category Knowledge Management, EN + DE) — 894 — keeps one second brain in the current working folder, after Building a Second Brain (BASB): the P.A.R.A. Method is the folder structure, Karpathy’s LLM Wiki the engine inside Resources. The agent never reads or writes a second brain outside that folder, so privacy comes from where the owner starts the session. Where the methods disagree, the contract decides: capture only what connects to a Project, an Area or one of the 12 Favorite Problems, and ask otherwise; file by actionability, topic pages only under Resources; never overwrite the owner’s wording; gather an Archipelago of Ideas before drafting; move, never delete; end each session with a Hemingway Bridge.

Why one brain. The first draft split a personal brain from each project’s documentation and guarded the boundary with rules. Review kept finding gaps at that boundary, so the owner replaced the rules with a folder: whichever folder the session starts in is the brain.

Tested before merge. Against the three anchors named in one sentence, the contract kept Haiku, Sonnet and Opus 5.5 out of a second notes folder the prompt mentioned in 6 of 6 runs (anchors: 0 of 6), left off-goal material out with a question in 6 of 6 (anchors: 0 of 6), and ended 18 of 18 runs with a Hemingway Bridge (anchors and baseline: none). The owner’s own paragraph stayed word for word in every condition.

A contract may now list other contracts under requires. Every export path ships them too, required contracts first and each once: the per-card copy and the multi-select copy and download on the contracts page, and the semantic-contracts-installer skill (new scripts/resolve-contracts.py). The card says "Includes: …", and contracts.txt, llms.txt and the contract pages add a "Requires" line. An unknown id or a cycle fails the build and the tests.

No contract uses it yet. It was built for the two-brain draft of Second Brain Maintenance, which needed LLM-Wiki Documentation; the one-brain contract does not. The mechanism and its tests, which run on synthetic catalogs, stay for the next contract that builds on another.

All 271 anchors carry a :definition: in English and German: two or three sentences that say what the whole concept is, who it comes from, and what it is for, using only facts from the anchor’s own file. Before #886, 225 anchors showed the first Core Concepts entry instead, which often described one part of the concept, not the concept.

The derived fallback is gone. A unit test fails for any anchor file without a definition, and CONTRIBUTING and the template list the attribute as required.

The meta description of each anchor page now comes from the definition too: it ends after the last full sentence within 160 characters, or at a word boundary with an ellipsis when the first sentence is longer.

Carbon Design System (IBM) — #874: a team-settings page imports @carbon/react and builds a DataTable in six of six runs.

Fluent 2 Design System (Microsoft) — #875: the same page uses @fluentui/react-components, makeStyles and Fluent tokens in six of six runs.

Primer (GitHub Design System) (GitHub) — #880: @primer/react with Octicons in six of six runs; the entry warns that the models still write the removed sx prop.

Shopify Polaris (Shopify) — #881: a Polaris Page, IndexTable and Badge in six of six runs — all in the deprecated React library, so the entry says to ask for Polaris web components.

Atlassian Design System (Atlassian) — #882: Atlaskit, DynamicTable and Lozenge in six of six runs.

Adobe Spectrum (Adobe) — #883: @adobe/react-spectrum with TableView in six of six runs; the bare "Spectrum" means physics.

Salesforce Lightning Design System (SLDS) (Salesforce) — #884: slds- markup in six of six runs, none with Lightning Web Components.

shadcn/ui (shadcn) — #876: a sign-up form imports from @/components/ui/ and names the npx shadcn CLI in six of six runs.

Bootstrap 5 (Mark Otto, Jacob Thornton) — #877: the same form uses form-control, is-invalid and invalid-feedback in six of six runs.

Ant Design (Ant Group) — #878: Form.Item with rules, Input.Password and onFinish in six of six runs.

U.S. Web Design System (USWDS) (18F, U.S. Digital Service) — #879: a parking-permit service gets the government banner, a step indicator and usa- forms in six of six runs; both Opus baselines already used usa- classes without the name.

WAI-ARIA Authoring Practices Guide (APG) (W3C Web Accessibility Initiative) — #866: names of UI controls fix what a control is called; the APG fixes how it behaves. An account menu gains Home, End and type-ahead in six of six runs (baselines: four and one of six); a site navigation uses the disclosure pattern in six of six runs and never menu roles. The entry explains how to prompt with it. Tier ★★★.

Bento Grid (community trend, often attributed to Apple) — #857: a features section moves from a uniform three-by-two card grid (five of six baselines) to mixed tile sizes on one grid (six of six). Say "bento grid": the bare "bento" resolves to food. Tier ★★★.

Glassmorphism (Michał Malewicz, 2020) — #858: pricing cards become translucent with backdrop-filter blur in six of six runs, none of six baselines. Current Status points to Apple’s Liquid Glass. Tier ★★★.

Bauhaus (Walter Gropius, Hannes Meyer, Ludwig Mies van der Rohe) — #833: the bakery page drops photography and script for circle, square, triangle, Futura and the primary colours in six of six runs.

Dieter Rams' 10 Principles of Good Design (Dieter Rams) — #834: a bedside clock becomes repairable or recyclable in six of six runs; no baseline mentioned repair.

Nielsen’s 10 Usability Heuristics (Jakob Nielsen, Rolf Molich) — #835: a form review maps every defect to at least five named heuristics, against none in the baselines.

Gestalt Principles (Max Wertheimer, Wolfgang Köhler, Kurt Koffka) — #836: the same review explains the defects through proximity and at least two further laws; the name keeps "Principles" because the bare "Gestalt" reaches psychology and therapy.

Laws of UX (Jon Yablonski) — #837: the review cites Fitts’s and Hick’s Law in six of six runs.

8-Point Grid (Bryn Jackson, Google Material Design) — #838: the share of pixel values on multiples of eight rises from 58 to 88 per cent.

Modular Scale (Tim Brown, Robert Bringhurst) — #839: heading sizes come from one ratio in a custom property in six of six runs.

The Elements of Typographic Style (Bringhurst) (Robert Bringhurst) — #840: article CSS gains old-style figures and indented paragraphs in six of six runs.

Every Layout (Heydon Pickering, Andy Bell) — #841: a card list loses its media query and gains a Stack in six of six runs.

Atomic Design (Brad Frost) — #842: a dashboard library is sorted into atoms, molecules, organisms, templates and pages in six of six runs.

Material Design 3 (Google) — #843: a settings screen uses surface roles, on-surface colours and tonal colour.

Apple Human Interface Guidelines (Apple) — #844: the same screen gets 44-point targets in six of six runs, large titles and Dynamic Type in five.

GOV.UK Design System (Government Digital Service) — #845: a parking-permit service gets one question per page in six of six runs, a pattern no baseline used.

Design Tokens (W3C DTCG) (Jina Anne, Jon Levine, W3C Design Tokens Community Group) — #846: values come back with $value and $type in six of six runs.

Refactoring UI (Adam Wathan, Steve Schoger) — #847: spacing replaces dividers on Sonnet and Opus; on Haiku the nudge is softer. Tier ★★☆.

WCAG 2.2 AA (W3C Web Accessibility Initiative) — #848: the form review cites at least ten success criteria, among them the 2.2 additions 2.5.8 and 2.4.11, in six of six runs.

Mobile First (Luke Wroblewski) — #849: the card-list CSS switches from a max-width to min-width media queries in six of six runs.

Neo-Brutalism (community trend; defined by Nielsen Norman Group, 2025) — #850: the strongest action effect of the run — black borders and hard offset shadows in six of six runs, none in the baselines. The term has no single originator; the maintainers decided that a reliable effect is enough and waived the Attributable criterion for this entry. Use the prefix: the bare "brutalism" resolves to architecture. Tier ★★★.

Swiss Style (International Typographic Style) (Ernst Keller, Josef Müller-Brockmann, Armin Hofmann, Emil Ruder) — design-principles + communication-presentation: a modular grid set before anything is placed, sans-serif type in few sizes, flush-left ragged-right text, asymmetric composition and objective imagery. Proposed by @shuhne in 821 — thank you. Asked for a jazz poster without the name, the baselines reached for illustrated silhouettes and condensed display faces; with "Swiss Style" every run on Haiku, Sonnet and Opus 5.5 set the poster on a modular grid in Helvetica or Akzidenz-Grotesk. The bare name resolves to graphic design on all three tiers. Tier ★★★. The criticism section carries the post-1968 charge of a cold, dogmatic style and the opposition led by the Basel teacher Wolfgang Weingart. Full detail in the prior-test register, run 2026-10-25.

Plain Language (ISO 24495) (ISO/TC 37, International Plain Language Federation) — communication-presentation + documentation: plain language defined by its outcome — readers can find what they need, understand it and use it — with the four principles relevant, findable, understandable and usable. Proposed by @mrsimpson in 819 — thank you. Asked to rewrite an official notice, the baselines and the runs told to use "plain language" returned one block of shorter prose; the runs told to follow ISO 24495-1 added headings and reader actions on all three Claude 5.5 tiers, and Sonnet and Opus walked through the four principles by name. Registered under the standard’s number because the bare term only shortens sentences. Tier ★★☆. The criticism section carries Yonathan Arbel’s big-data challenge to the movement’s readability claims. Full detail in the prior-test register, run 2026-10-25.

The LLM-Wiki Documentation contract (category Documentation, EN + DE) applies Karpathy’s LLM Wiki to project documentation and names only where it departs from the pattern: the wiki’s pages are the arc42 chapters, ADRs and glossary in AsciiDoc, built by docToolchain (Docs-as-Code); every statement cites its source, and an unsupported one becomes an issue for the role that can answer it; changes reach the documentation only through pull-request review. The contract stays this short on purpose: a prior test the same day (prior-test register, run 2026-10-15) showed that the anchor alone makes Haiku, Sonnet and Opus 5.5 build the raw layer, index, log, lint pass and filing back of good answers.

Karpathy’s LLM Wiki (Andrej Karpathy) — knowledge-management + documentation: the LLM compiles immutable raw sources into interlinked markdown pages, keeps index.md and log.md, files good answers back and lints the whole wiki for contradictions and gaps. Published in April 2026 and already dense on all three Claude 5.5 tiers: on a support-assistant task every baseline built chunked retrieval with a vector index, every anchored run built the wiki, index, log and lint pass and filed good answers back. Registered under the qualified name because the weak tier hedges on the bare "LLM Wiki" and drops parts of the pattern. Tier ★★☆. The criticism section carries Michael Tuszynski’s report of what broke after six months at three hundred pages.

Twin Peaks Model according to Nuseibeh (Bashar Nuseibeh) — requirements-engineering + software-architecture: requirements and architecture refined concurrently, with equal status, instead of frozen requirements handed to design. Registered with Nuseibeh’s name because "Twin Peaks" alone is the TV series. Tier ★★☆.

Domain Storytelling according to Hofer and Schwentner (Stefan Hofer, Henning Schwentner) — requirements-engineering + software-architecture: actors, work objects and numbered activity arrows, one concrete story per diagram. Event Storming and Domain-Driven Design now point at it. Tier ★★★.

req42 gains its goal notation PAM (Purpose, Advantage, Metric) and its top-down presentation from vision and goals to a backlog of epics, features and stories, each cited to the template and the req42 tips.

Defensive Programming according to McConnell (Steve McConnell) — design-principles: validate untrusted input once at a barricade, assert inside, and choose robustness or correctness per system. Registered with McConnell’s name because the weak tier attributes the practice to Bertrand Meyer, and Design by Contract argues against defensive double-checking. The barricade reconciles the two, and both entries now point at each other. Tier ★★☆, #797.

Finite State Machine (George H. Mealy, Edward F. Moore, David Harel) — design-principles: a finite set of named states and a transition table in which every cell is a decision. The model behind the GoF State pattern, which now names it as such; the pattern is one way to code it. Tier ★★★, #798.

MISRA C (MISRA) — testing-quality: the safe subset of C for critical systems, with mandatory, required and advisory guidelines and documented deviations. The first coding standard for embedded safety in the catalog, next to IEC 61508 SIL Levels. The entry describes themes only and quotes no rule text, because the texts belong to MISRA. Tier ★★★, #799.

BlinkWithoutDelay (David A. Mellis, Arduino) — design-principles: never block the loop; one timestamp per task and now - last >= interval, which survives the millis() wrap. The first platform idiom in the catalog, registered under the sketch name because that is the string that activates. Tier ★★★, #800.

The Writing Style contract now ends with a stumble test (EN + DE): eight checks for the sentences an expert reader stops at. The trigger was a program chair’s verdict on an AI-written abstract — "rhetorically polished, strangely empty" — and the observation that experts stumble over unearned claims, not over jargon. The checks cover unresolved contrasts, announced instead of shown differences, references the reader has to hunt for, absolute promises, terms that are named but not used, half-introduced terms and empty rhetorical figures. The last check hands the text to a fresh context without background, because the author never stumbles over their own sentences.

It also gains a rule for English text: write it as English, not translated from German — no German sentence patterns with English words, no literal idioms; contractions where a native speaker would use them, full forms for formal instructions.

Refactoring Catalog according to Fowler (Martin Fowler, Kent Beck, William Opdyke) — design-principles + development-workflow: named, behaviour-preserving transformations with written mechanics. Code Smells says why; the catalog says how. Tier ★★☆, because the bare verb "refactor" covers any code change — the qualified form is the anchor. #775.

CQS (Command-Query Separation) (Bertrand Meyer) — design-principles: a method either answers a question or changes something, never both. A rule for every single method, which is what separates it from CQRS. Tier ★★★, #776.

Railway Oriented Programming (Scott Wlaschin) — design-principles: two-track composition where a failure short-circuits the rest. Names the style, not a library. Wlaschin wrote the counterweight himself in 2019. Tier ★★★, #780.

Principle of Least Privilege (Jerome H. Saltzer, Michael D. Schroeder) — design-principles + testing-quality: the 1975 principle, measured here on the artifacts nobody thinks to scope. Tier ★★★, #784.

Arrange-Act-Assert (AAA) (Bill Wake, Gerard Meszaros) — testing-quality: three blocks, one action per test. Registered under the written-out name, because the bare acronym anchors Authentication, Authorization and Accounting. Tier ★★★, #778.

Test Data Builder (Nat Pryce, Steve Freeman) — testing-quality: safe defaults and fluent overrides, so a changed constructor breaks the builder instead of two hundred tests. The densest prior in the batch. Tier ★★★, #779.

OWASP ASVS (OWASP Foundation) — testing-quality: numbered, verifiable requirements at three assurance levels, the testable counterpart to OWASP Top 10. Tier ★★☆; name the level and the version, because v5.0 renumbered everything. #783.

4+1 View Model according to Kruchten (Philippe Kruchten) — software-architecture + documentation: four views, each written for a named stakeholder group, plus scenarios as the cross-check. The stakeholder mapping is what arc42 and C4 do not carry. Tier ★★★, #777.

Enterprise Integration Patterns (Gregor Hohpe, Bobby Woolf) — software-architecture: 65 named messaging patterns with an icon notation. Event-Driven Architecture is the style; this is the vocabulary. Tier ★★★, #786.

Bulkhead Pattern (Michael T. Nygard) — software-architecture: cap what one dependency may consume. Circuit Breaker stops the calls; the bulkhead limits the damage while they still run. Tier ★★★, #781.

Richardson Maturity Model (Leonard Richardson, Martin Fowler) — software-architecture + design-principles: a design target, not a score. Tier ★★★, #782.

Defense in Depth (US National Security Agency, James Reason) — software-architecture + testing-quality: independent layers, each assumed to fail. Tier ★★☆, and the entry states the independence requirement explicitly because the name does not carry it. #785.

Keep a Changelog (Olivier Lacan) — documentation + development-workflow: a curated release history with six change types and an [Unreleased] section, deliberately not generated from the commit log. Tier ★★★, #787.

Connextra User Story Format (Connextra team, Mike Cohn) — requirements-engineering: the three-part template that forces a story to name who wants it and why. INVEST judges quality, this names the form. Tier ★★★ in its qualified form only. #788.

DMADV (Motorola, General Electric) — problem-solving + statistical-methods: Define, Measure, Analyze, Design, Verify, the design-side sibling of DMAIC for when the process does not exist yet. Tier ★★★, #754.

PDCA (Walter A. Shewhart, W. Edwards Deming) — problem-solving + development-workflow: state a hypothesis before acting and standardise the gain after, which is what separates it from generic iteration. Tier ★★★, #755.

FMEA (US Armed Forces MIL-P-1629, NASA, AIAG, VDA) — testing-quality + problem-solving: per-failure-mode rows with severity, occurrence and detection ratings, and derived actions instead of a prose risk list. Tier ★★★, #756.

SIPOC (Total Quality Management practice, no single originator) — problem-solving + requirements-engineering: Suppliers, Inputs, Process, Outputs, Customers, drawn to settle where a process starts and stops before anyone argues about details. Tier ★★★, #757.

Ishikawa Diagram (Kaoru Ishikawa) — problem-solving: parallel cause categories where 5 Whys follows a single chain. Tier ★★★, #758.

Value Stream Mapping (Mike Rother, John Shook, Learning to See, LEI 1999) — problem-solving + development-workflow: separates process time from waiting time, then demands a future state. Tier ★★★, #759.

Poka-Yoke (Shigeo Shingo) — design-principles + testing-quality: make the wrong input impossible rather than validating and training against it. Tier ★★★, #760.

QFD (Yoji Akao, Shigeru Mizuno) — requirements-engineering + strategic-planning: a method for deploying customer needs through design and production. Its best-known tool, the House of Quality, carries a correlation roof showing where two technical characteristics fight each other — the view neither Pugh Matrix nor Kano Model has. Tier ★★★, #761.

Jidoka (Sakichi Toyoda, Taiichi Ohno) — design-principles + testing-quality: halt on the anomaly instead of passing the defect downstream. One of the two TPS pillars. Tier ★★★, #762.

A3 Problem Solving (John Shook, Managing to Learn) — problem-solving + documentation: the whole analysis on one sheet. The constraint is the method. Tier ★★★, #763.

Toyota Kata (Mike Rother, 2009) — problem-solving + strategic-planning: a measured current condition and a dated target condition before any experiment, plus the coaching routine that makes it stick. Tier ★★★, #764.

Andon (Taiichi Ohno, Toyota) — development-workflow + testing-quality: stop authority granted to whoever sees the problem, with a defined response. Tier ★★★, #765.

Genchi Genbutsu (Taiichi Ohno, Toyota) — problem-solving + knowledge-management: go and see. Named, it makes a model ask for the actual trace before it theorises. Tier ★★★, #766.

Muda, Mura, Muri (Taiichi Ohno, James P. Womack) — problem-solving + strategic-planning: registered as the triple, because naming only waste hides its cause. Tier ★★★, #767.

Kaizen (Masaaki Imai, 1986) — problem-solving + development-workflow: standardise the gain, then improve from the new standard, with kaikaku as the deliberate contrast. Tier ★★★, #768.

Boy Scout Rule (Robert C. Martin, Clean Code, Prentice Hall 2008, chapter 1) — design-principles + development-workflow: every time you open a file for some other reason, leave it slightly better than you found it, bounded to the code the task already touched. Tier ★★★, proposed by @PhilippJust in #742.

Dale Carnegie Principles (Dale Carnegie, How to Win Friends and Influence People, 1936; revised 1981) — communication-presentation + dialogue-interaction: open with honest appreciation, raise the problem second, point out mistakes indirectly and let the other person keep their standing. Tier ★★★, proposed by @xamurej in #725.

CUPID Properties (Dan North, dannorth.net, 10 February 2022) — design-principles + software-architecture: Composable, Unix philosophy, Predictable, Idiomatic, Domain-based, offered against SOLID as qualities you move toward rather than rules you comply with. Tier ★★☆, proposed by @simonthum in #737.

Crisis and Emergency Risk Communication (CERC) (Barbara Reynolds and Matthew W. Seeger, Journal of Health Communication 10(1), 2005; manual by the US CDC) — communication-presentation + problem-solving: Be First, Be Right, Be Credible, Express Empathy, Promote Action, Show Respect, for the case where you have to speak before the facts are in. Tier ★★☆, :tier: 1, proposed by @rweisleder in #741.

"Uncle Bob" was dropped from the Boy Scout Rule. The proposal named it "Boyscout Rule Uncle Bob". The bare name already returns software engineering on both tiers, so the qualifier repairs nothing and the attribution belongs in :proponents:, where a reader looks for it.

CERC had to be spelled out. Asked about the bare acronym, the strong tier listed four expansions — none of them the framework — and closed with "If you don’t specify a context, the Indian electricity regulator is the most likely meaning." The proposal’s "(CDC)" qualifier did not survive into the model’s own default reading. This is also the run’s sharpest inversion: the stronger model is the more ambiguous one, because it knows more expansions.

The quick-nav now opens with a ✨ chip that filters the catalog down to the twelve most recently added anchors. The category chips beside it jump; this one filters, so it is a toggle button with a pressed state and a dashed border rather than a link — a jump link is never "on".

Twelve newest, not "added in the last 30 days". A time window empties out between release bursts, and a control that filters to nothing is worse than no control at all. The consequence is deliberate and visible: the list reaches back further than the NEW badge, so an anchor can appear in it without carrying the badge. The badge is a claim about age, the chip is a claim about order.

Tufte Style (Edward R. Tufte, The Visual Display of Quantitative Information, Graphics Press 1983; 2nd ed. 2001) — communication-presentation + design-principles: maximise the data-ink ratio, remove chartjunk, keep the lie factor near 1, compare with small multiples on shared axes, set sparklines inside the sentence, label directly instead of with a legend. The point that decides whether the term was understood: density is the goal, not emptiness — remove what carries no information, then spend the reclaimed space on more information. Tier ★★★, proposed by @follperson in #712.

The anchored runs reach for constructs generic minimalism does not contain. Asked for a one-page executive report, the anchored strong tier specifies a range-frame axis "drawn only from the series minimum to its maximum" and a dot-dash plot "with marginal rug ticks on both axes … so the distribution of the data doubles as the axis". Both are Tufte’s own inventions, and neither was named in the prompt. The sharpest single contrast is one element decided in opposite directions: anchored, the summary table carries "a signed value only, no arrows, no traffic lights"; the neutral baseline puts "arrow glyphs — ▲ 3.1% MoM" in exactly that cell.

The first action probe measured nothing, and that is recorded as a defect in our method rather than smoothed away. It asked for a dense CI dashboard — twelve teams, ninety days, one screen — so density was a requirement of the task, not a contribution of the anchor, and the strong tier’s baseline out-scored the anchored run. The rule this produces: an action task must not pre-load the term’s characteristic move, or a null result cannot distinguish "the model absorbed this long ago" from "I asked for it in the prompt". The probe was repeated on a task whose default is decorative, and there it separates on every tier.

A second correction to the battery. The weak tier reported reaching=yes in both runs — normally the clearest sign of a thin prior — while naming the lie factor, moiré patterns, small multiples and sparklines. Testing three name forms showed the flag attaches to the eponym, not to any suffix. reaching is a self-report and has to be graded against the content, not read off the SELF: line.

The Criticism section is the best-supplied in the catalog so far, and it keeps three targets apart that are routinely conflated: Bateman et al. (CHI 2010) and Inbar/Tractinsky/Meyer (ECCE 2007) test the principles — embellished charts were recalled significantly better; Robison, Roger Boisjoly, Hoeker and Young (2002) dispute the Challenger analysis, and Boisjoly is one of the engineers Tufte wrote about; Don Norman argues with the PowerPoint essay. Stephen Few, expected to be the in-house critic, turns out to defend Tufte against Bateman and is cited as the counter-voice.

Anchors carry a :tested-on: attribute and the modal renders it at the foot of the entry, linked to the prior-test register. The tier stays out of the interface deliberately: a tier is a judgement, a test result is a fact.

Every tested model is named, including the ones where the anchor did nothing. Listing only the successes would read as "the others were not tried", which is a weaker and different claim than "tried, no effect". Simplified Technical English is the case in point: it is marked as tested on all three tiers, and as having no measurable effect on the weakest one, where its rewrite was not distinguishable from an unanchored prompt.

Anchors without the attribute show nothing at all. Absence is a statement about our records, not about the term.

Simplified Technical English (ASD-STE100) (ASD, maintained by the STEMG; first issued 1986 as AECMA Simplified English, Issue 9 from January 2025) — documentation: 53 writing rules and a dictionary of roughly 900 approved words, each admitted in one meaning and one part of speech. Tier ★★☆, not ★★★: the anchored rewrite is clearly distinguishable from an unanchored one on frontier models, but not on weak ones, where a neutral baseline produced work of comparable quality. Our own cross-model rule prescribes ★★ for exactly that.

The literature reached the same conclusion thirty years ago. Holmback, Shubert and Spyridakis — the authors of the Boeing-funded efficacy studies — ask in CLAW '96 "whether SE is any more beneficial than good quality, professional technical writing that does not conform to the SE standard", and report no significant time saving and unclear translatability results. The Criticism section carries that, along with the adoption cost and the fact that ASD certifies no tool as fully compliant.

Rejected: Zinsser’s Four Principles (706) — see Rejected Proposals. Attributable to William Zinsser verbatim, and still not an anchor. A three-armed action test rewrote one paragraph three ways: naming Zinsser, naming only the four words without him, and neutrally. All three produced the same rewrite; the name added commentary about the principles rather than better prose. A prediction that the book’s own wording ("four articles of faith") would measure better was tested and refuted — it measured worse. No contract was added either: Writing Style already names Strunk & White and Wolf Schneider and states eight concrete rules, which four adjectives do not improve on.

Locality of Behaviour (Carson Gross, htmx essay, 29 May 2020) — design-principles: "The behaviour of a unit of code should be as obvious as possible by looking only at that unit of code." Optimise for read-time cost over write-time cost; behaviour at a distance is the failure mode, co-location the remedy. Tier ★★★, proposed by @simonthum in #701.

It produced the clearest lift measured so far, because the term changes the verdict and not just the vocabulary. Asked to review a design that registers every click handler in one central JavaScript file, both neutral baseline runs returned a balanced review that listed Separation of Concerns among the strengths and never mentioned locality. With the term, the first heading is "Behaviour at distance". The baseline praises exactly the position the essay attacks.

Two findings are recorded rather than smoothed away. The abbreviation LoB returns "Line of Business" on every tier, so the catalog name is spelled out. And the author uses two spellings — British in the essay, American in Hypermedia Systems — so both were tested; the British form measured stronger on weak models, which is also the essay’s own spelling. Details in the prior-test register.

The Criticism section carries three named critics, including one who turns the principle against htmx itself, and records that no worked rebuttal from the established frontend side could be found — an absence in our search, not a demonstrated absence.

Premortem (Gary Klein, Harvard Business Review 85(9), September 2007) — problem-solving + strategic-planning: state as fact that the plan has already failed, then have each participant write down why, independently, and collect the reasons round-robin. The failure is a premise, not a possibility, and that is the whole technique. Tier ★★★. The Criticism section reports the one citable critic found: Jason Collins shows that the widely repeated "30 percent" figure does not say what it is used to say — the 1989 study behind it measured the number of reasons generated, and "did not assess the quality of the reasons". The claim that the technique improves correct identification by 30 percent originates in the 2007 restatement, not in the study.

Second-Order Thinking (Howard Marks, Ray Dalio; popularised by Shane Parrish) — problem-solving + strategic-planning: follow a consequence into the next consequence instead of listing more effects at the same level. Tier ★★★. The entry is named against its own bibliography: Marks writes "second-level thinking", Dalio writes "second- and third-order consequences", and neither uses the term this anchor carries. Both names were tested, and the measurement decided it — see below.

Anchor Prior Test — the skill that has been in the plugin all along now has a page. It runs a candidate through a fixed probe battery across model tiers in a clean room, compares against a neutral baseline, and returns a criteria matrix, a tier, and a route to anchor, contract or rejection. Publishing it makes the catalog falsifiable: anyone can rerun the method and disagree with a tier.

Prior-Test Register — the evidence behind each :prior-test: date. Resolved model identifiers, procedure version, run counts, matrices and verbatim quotes. The date lives in the anchor, the evidence lives here, because model aliases move: "tested on haiku" says nothing a year later, and even resolved identifiers are only partly pinned.

New :prior-test: attribute on anchor files. The 192 existing anchors do not carry it, which is a statement about our records rather than about those terms; they will be retrofitted gradually, starting where the claim is most contestable.

Documentation Verification — the SDD framework benchmark ran five spec-driven approaches against the same challenge and sampled three documented claims per run. Fifteen out of fifteen had drifted from the code. Every framework in the field had produced traceability tooling, and every one of them only ran forward: it enforced that new rules get documented, never that existing statements still hold. The periodic prompt "does the specification or architecture documentation need updating?" is the reason. It asks what we learned, so an agent that learned nothing new answers "no" truthfully and a claim that was already wrong survives every review. The new contract reverses the direction into a checkable question — confirm each structural claim (module names, file paths, table and column names, enum values, start commands) against the code, report how many were checked and how many drifted — and forbids the tempting repair: never change the code to match the document, because the document states an intention and the code states reality. Anchors: Docs-as-Code, Mental Model according to Naur, Fagan Inspection.

Not one of the five benchmark runs had a test guarding the weighting of a calculation, at 78, 90, 108, 146 and 165 tests alike. The cause is structural rather than careless: tests get derived from the rules, and a rule such as "all shares are equal" makes weighting invisible, so a half share counted as a full one produces a plausible number and stays green everywhere. Implement Next now asks for invariant tests alongside example-based ones — name the property connecting a calculation’s inputs to its output, test it over a range of values, and ask which of those tests would still fail if a change altered the units, the granularity or the cardinality of an input. The contract’s "check if docs need updating" step was also sharpened into the verification direction described above. Added Property-Based Testing to its anchor set.

Step 8’s Feedback Loop to Specification now carries one prompt asking both questions at once — what implementation revealed, and whether each structural claim still holds — rather than only the forward-looking one. A new Invariant Tests subsection sits alongside it, and Step 12 now closes by asking both documentation questions again against the code that exists after the bug-fix loop, the point where it pays most: fixes change behaviour, and changed behaviour is what makes a correct structural claim stale.

The /de/contributing page was built from website/public/CONTRIBUTING.de.adoc, but every contributor edit since #603 went into docs/CONTRIBUTING.de.adoc — the file that sits next to its English sibling and looks maintained. Nothing rendered it. German readers therefore never saw the artifact taxonomy (#585), the Viability Test (#640), the ADR-008 topic-neutral gate, the issue-title convention (#564), or the developer setup, PR review policy and Criticism/Current-Status rules added in #695. render-docs.js now renders docs/CONTRIBUTING.de.adoc, and the stale copy under website/public/ is deleted, so there is one German source again. Also fixed five relative link: targets in both Contributing files (ADR-007, ADR-008, docs/roles/) that resolved against the served route and 404’d — they now point at absolute GitHub URLs, which work on the website and in GitHub’s rendered view alike.

DMAIC (Bill Smith, Motorola 1986; popularised by GE under Jack Welch from 1995) — problem-solving + statistical-methods: the five gated phases Define, Measure, Analyze, Improve, Control. The anchor’s utility is structural, not descriptive: naming it forces a baseline before any improvement is proposed, separates validated root causes from solutioning, and adds a Control phase that most improvement approaches simply lack. It closes a hole in the catalog — SPC, Control Chart and Nelson Rules all referenced DMAIC as the cycle they serve, without it existing as an anchor. Tier ★★★, confirmed on a weak local model (qwen3.6-27b named all five phases unprompted), bilingual, with a fetch-verified Criticism section (Juran: "there is nothing new there"; the 3M/McNerney case where it stifled research; >60% project failure rate per the WSJ; Wheeler on the 1.5-sigma shift) and a Current Status noting ISO 13053-1:2011 and Lean Six Sigma. Proposed by @Advisior in #696.

Residuality Theory (Barry M. O’Reilly) — software-architecture: start from a deliberately inadequate naive architecture, apply stressors (technical, regulatory, competitive, organisational), identify the residue — what survives — and redesign so failure happens in a controlled way; contagion analysis then surfaces the coupling that otherwise only appears in production. The anchor inverts the design question from "what do we extract first?" to "what survives every stressor?", which is why it is not a reframing of Strangler Fig. Tier ★★☆ — the qualified form ("apply Residuality Theory: run a stressor analysis") is needed, because weaker models dilute the specific vocabulary into generic complexity-science terms. Bilingual, with a fetch-verified Criticism section (the author’s own "case studies is marketing rather than science"; a learning curve he compares to learning OOP; single-originator evidence base) and a Current Status listing the three Procedia Computer Science papers (2020–2022) and the 2024 Springer chapter. Proposed by @dennissterzenbach in #693, reviewed by @JensGrote.

A parity audit found 121 metadata drifts between anchors' English and German files: :tier: was missing from ~90 German files, the GoF-pattern and SOLID German files carried tiers inconsistent with their English source (EN 3 vs DE 1/2), and several narrative anchors (three-act-structure, hero’s-journey, save-the-cat, …) had :related: / :tags: only in English; problem-space-nvc.de was also missing a category and a role. Synced all identifier/number metadata (:categories:, :roles:, :tier:, :related:, :tags:) from the authoritative English files to their German counterparts across 100 files. Two :proponents: fields keep their localized German phrasing, by design.

Repointed pre-existing broken cross-references to their real anchor IDs: kiss-principle/yagni → dry (was dry-principle), yagni → ssot-principle (was spot-principle), effective-go → effective-java (was clean-code). All :related: targets across the catalog now resolve to existing anchors.

Added curated :related: cross-links to 70 anchors (EN + DE) that lacked them — 2–5 genuinely related anchors each, using category and tag overlap as candidates and hand-curating to real conceptual connections (the SOLID and GoF-pattern clusters, the personality models, the TDD schools, the SPC trio, and so on). 187/188 anchors now carry :related: (the meta-anchor what-qualifies-as-a-semantic-anchor is intentionally standalone). This closes the #668 metadata audit: Phase A (:proponents:), Phase B (:tags:), Phase C (:related:) — the catalog’s required and recommended metadata is now complete.

Completes :tags: for the whole catalog — 100% of anchors now carry search keywords. Added tags to 39 anchors that lacked them (EN + DE), derived from each concept’s name, aliases, key sub-concepts and proponents (e.g. arc42 → arc42, architecture documentation, Starke, Hruschka). Follows the Phase A :proponents: backfill; Phase C (:related:) is the remaining follow-up.

DISC Model (William Moulton Marston, 1928) — communication-presentation: four behavioural styles (Dominance, Influence, Steadiness, Conscientiousness) as a lightweight communication-style shorthand. Like MBTI, DISC passes the ADR-008 Utility Gate — naming it reliably imposes the four-quadrant frame on output — so it is admitted, at tier ★☆☆ with an :advisory: flag (#624) and a fetch-verified Criticism section (classified as pseudoscience; no demonstrated predictive validity; the apparent accuracy is partly the Barnum effect; DISC data are better explained as Big Five combinations). Bilingual, cross-linked with MBTI, Big Five, and HEXACO. Proposed by @rdmueller in #671. This completes the personality-model cluster under ADR-008.

Applying ADR-008 to the existing MBTI anchor, without grandfathering. MBTI passes the gate — naming it reliably imposes the 16-type / four-dichotomy structure on output, which is genuine output-structure change, not mere recognition — so it stays. But its criticism did not meet the #603 convention: it now carries a proper fetch-verified == Criticism section (Pittenger — 39–76% receive a different type on a five-week retest; widely regarded as pseudoscience) and a == Current Status, plus an :advisory: flag (#624) and :related: links to the empirically validated Big Five and HEXACO. The gate keeps its teeth: MBTI clears it on utility while being held to the same criticism standard as new anchors.

HEXACO (Kibeom Lee & Michael C. Ashton) — communication-presentation: the six-factor personality model — Big Five plus Honesty-Humility, the strongest trait predictor of unethical/exploitative behaviour and the Dark Triad. Admitted under ADR-008 because that sixth factor delivers distinct utility the Big Five cannot (the redundancy test, not a synonym). Tier ★★☆ (fires with the qualified form "assess along HEXACO, especially Honesty-Humility"), bilingual, with a fetch-verified Criticism section (De Raad et al. — the sixth factor does not always replicate cross-culturally) and a Current Status referencing the Big Five. Proposed by @raifdmueller in #674.

Big Five / OCEAN (Lewis R. Goldberg; Paul T. Costa & Robert R. McCrae) — communication-presentation: five continuous trait dimensions (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism), the empirically dominant trait model and the scientific benchmark against MBTI/DISC. Added as the first anchor to clear the new Utility Gate: tier ★★☆ (fires with the qualified form "profile along the OCEAN dimensions"), bilingual, with a fetch-verified Criticism section (Gurven et al. 2013 — the structure failed to replicate among the Tsimane; the evidence base is heavily WEIRD) and a Current Status pointing to the HEXACO challenger. Proposed by @raifdmueller in #673.

The catalog held one personality anchor (MBTI) while the proposal template rejected "personality/assessment frameworks" — a contradiction that recurred with each new proposal (DISC #671, Big Five/OCEAN #673, HEXACO #674). ADR-008 resolves it with a topic-neutral Utility Gate: no subject area is banned or blessed; a framework is admitted only if naming it demonstrably changes the LLM’s output structure (the existing Viability Test), and carries a Criticism section. MBTI is re-subjected to the same test rather than grandfathered. CONTRIBUTING (EN + DE) and the proposal template are updated to state the topic-neutral rule; the arc42 Chapter 9 ADR index lists ADR-008.

:proponents: is a required attribute, yet a catalog-wide audit found it missing in many anchors. This completes :proponents: for the whole catalog — now 100% of anchors carry the required field. 22 anchors were newly attributed to a clear, canonical originator (EN + DE) — e.g. Clean Architecture (Robert C. Martin), Domain-Driven Design (Eric Evans), Hexagonal Architecture (Alistair Cockburn), Wardley Mapping (Simon Wardley), Pyramid Principle (Barbara Minto) — and 11 German files were brought in sync with their English source. Two anchors gained an institutional/project attribution (devils-advocate → the Roman Catholic Church’s advocatus diaboli; the meta-anchor what-qualifies-as-a-semantic-anchor → the project itself). For established terms of art with no single originator (sota, ssot-principle), CONTRIBUTING now defines a Community sentinel — a missing proponent is treated as a signal to recheck whether the term is truly a semantic anchor. Names only; no invented dates. Phases B (:tags:) and C (:related:) follow.

The og:image tag now also carries the name="image" attribute (<meta name="image" property="og:image" …>), a form LinkedIn’s scraper reads more reliably. The image, filenames and other tags were already correct (see 2026-07-02); this only hardens how strict scrapers pick it up.

Sharing the site on LinkedIn, X, Slack and the like showed no preview — the og:image and twitter:image meta tags pointed at og-image.png / twitter-card.png, which never existed (404). Both files are now shipped: a 1200×630 branded card (dark theme, anchor mark, "Shared vocabulary for precise communication with LLMs"). No HTML change was needed; the tags already referenced the correct names.

Deep Modules (John Ousterhout, A Philosophy of Software Design) — design-principles: a deep module hides a powerful implementation behind a simple interface. Proposed by @rdmueller in #660.

Test Seam (Michael Feathers, Working Effectively with Legacy Code) — testing-quality: a place to alter behaviour without editing there, to get legacy code under test. Proposed by @rdmueller in #661.

Progressive Disclosure (Jack Carroll; Jakob Nielsen / NN/g) — communication-presentation: show the common few first, reveal the rest on demand. Proposed by @rdmueller in #662.

Zone of Proximal Development (Lev Vygotsky) — communication-presentation: the gap between solo and guided performance, and scaffolding within it. Proposed by @rdmueller in #663.

Sender-Receiver Discrepancy — communication-presentation: the cause-agnostic umbrella over communication-gap models (meaning is reconstructed by the receiver). Proposed by @rehsack in #639.

50/72 Rule (Tim Pope) — development-workflow: 50-char subject, 72-char body wrap for readable Git commits. Proposed by @steffakasid in #620.

A new presentation-planning contract (category Communication, EN + DE) bundles known anchors into a plan-a-talk workflow: Socratic Method to elicit the brief, Curse of Knowledge for audience empathy, Pyramid Principle + MECE for top-down structure, Three-Act Structure / Story Circle for the narrative arc, 4MAT for why-first, and BLUF for the opening. It produces a plan (core message, audience, arc, per-section takeaways) rather than slides, and works one step at a time. It deliberately excludes delivery/stagecraft — no established anchor covers it yet. Proposed by @raifdmueller.

Anchors added in the last 30 days now show a green New badge on their card, so returning visitors can spot fresh entries at a glance. The "added" date is derived from each anchor file’s first git commit at build time (no manual upkeep) and stored as addedAt; the badge is computed at render time against the current date, so it expires on its own. Bilingual ("New" / "Neu"). The deploy workflow now checks out full history (fetch-depth: 0) so the dates are available in CI.

A GEO audit found the site ranks #1-3 in traditional search yet gets zero citations from ChatGPT, Perplexity, Gemini and Copilot for "What is Semantic Anchors?". The fix: every anchor’s pre-rendered page now carries a concise, standalone direct-answer block that AI crawlers (which mostly do not run JS) can quote verbatim, and the home page gains a crawlable "What is Semantic Anchors?" answer section (EN + DE). A shared resolver prefers a curated :definition: attribute and otherwise derives one from the first Core Concept, so the answer block, the modal lead-in, and the DefinedTerm JSON-LD description all agree. Curated definitions seeded for 11 high-value anchors (MECE, BLUF, arc42, Clean Architecture, TDD London School, Gherkin, SOLID, Cynefin, Socratic Method, Chain of Thought, DDD); the rest use the derived fallback until curated. Proposed by @raifdmueller in #580.

The semantic-anchor-onboarding skill installed a list of anchors the model mostly already knows. It is renamed and repositioned as semantic-contracts-installer: its default mode now installs Semantic Contracts — the compositions and rendered templates the model cannot guess (e.g. "Specification" = Cockburn use cases + Gherkin + EARS) — reading a bundled data/contracts.json snapshot (refreshable via scripts/refresh-contracts.sh). The former anchor-block flow stays as an --anchors-only sub-mode. The old skill name remains for one release as a deprecation alias that forwards to the new one and idempotently overwrites blocks it previously installed (same markers). The plugin is bumped to v0.6.0.

SessionStart hook fixed (#515): the installer no longer installs a Claude hook for CLAUDE.md (Claude Code loads it natively, so the hook double-loaded the block), and the hook is registered with matcher startup only — no more noisy re-injection on every resume.

Some terms activate reliably yet prime a framing their own field cautions against (e.g. the Eisenhower Matrix and the "mere-urgency effect"). Two new handling paths: a hard rejection category Counter-consensus framing in rejected-proposals.adoc (which requires a cited consensus source from the framework’s own field), and — for terms worth keeping — an optional :advisory: anchor attribute that renders as a visible amber badge on the card and a banner in the modal, with the detail and source staying in the anchor’s Criticism / Current Status section (single source of truth). The Eisenhower Matrix is the seed: it gains an :advisory: flag and a == Current Status section. Proposed by @rehsack in #624.

Contribution guidance now tests not just whether a model recognizes a term but whether naming it changes the output. A new "Viability Test" section in CONTRIBUTING.adoc / CONTRIBUTING.de.adoc adds a before/after test, a substitution check (silent substitution / confabulation on weaker models), and a cross-model check; propose-anchor.yml gains an optional viability-test field. A term that passes the four quality criteria but fails viability belongs in a Contract, not an anchor — the destination the taxonomy (#585) defines. Proposed by @JensGrote in #640.

The "Not in training data" rejection category was sharpened and renamed Insufficient training-data density, distinguishing recognition from activation and pointing thin-prior terms at a Semantic Contract.

The Vitest unit suites (the website app’s 100+ component/util tests and the scripts metadata-extraction test) ran nowhere in CI — only lint, E2E, Lighthouse and the security audit did. A drifted assertion had silently gone red on main as a result. A new Unit Tests job in test.yml runs both suites on every push and PR; scripts gains Vitest and a test script so its test is executable in CI.

The architecture-documentation contract had grown to seven paragraphs — most of it procedure, not vocabulary. It now shrinks to one paragraph of shared vocabulary (arc42; C4 via PlantUML; Nygard ADRs with a 3-point Pugh matrix; six-part Quality Attribute Scenarios). The procedure — scaffolding the arc42 with-help template, the five cross-section traceability rules, the Chapter 11 Risks-vs-Technical-Debt structure, the ADR↔risk-ID wiring, the Chapter 1.2-vs-10 quality-goal marking — moves into a new on-demand arc42-documentation skill (SKILL.md + a chapter-by-chapter checklist), shipped in the plugin (bumped to v0.5.0) with its own website page. This is the Anchor/Contract/Skill taxonomy (ADR-007, #585) applied to the project’s own artifacts: a contract that drifted into a skill, split back. crosscutting-concepts is already terse and stays as-is.

The three artifact types the catalog ships now have an explicit, written definition instead of an implicitly-reconstructed one. New ADR-007 (Nygard + Pugh, +63 vs. keeping it implicit) defines each type by function, linguistic shape and loading behaviour, plus a discrimination test; the working definitions are mirrored into CONTRIBUTING.adoc and CONTRIBUTING.de.adoc so contributors meet them where they look. ADR drafted by @JensGrote in #585.

Thin-prior rule: a name not dense in training data is not an anchor — if the missing piece is meaning it becomes a Contract (which supplies its own meaning), if it is procedure it becomes a Skill. This reconciles the one point where #585’s draft and the viability proposal (#640) disagreed, and matches the project’s "weak prior → contract" position. The taxonomy unblocks the contract/skill split (#529) and the onboarding-skill pivot (#521).

The arc42 Chapter 9 ADR index was completed (it had also been missing ADR-006).

The deploy build broke at the metadata step because the req42 anchor’s title equalled its filename id, which the validator treats as "no real title set". Gave it a descriptive title (req42 Requirements Framework), in the same style as arc42 Architecture Documentation.

npm audit fix cleared the three high-severity advisories that were failing the CI security audit (npm audit --audit-level=high): Vite 7.3.2 → 7.3.6 (server.fs.deny bypass on Windows alternate paths, plus the bundled launch-editor NTLMv2 hash disclosure) and ws (uninitialized memory disclosure, memory-exhaustion DoS). Pure lockfile bumps within the declared ^7.3.1 range — build and unit tests stay green.

The four remaining moderate advisories (js-yaml, uuid) are dev-only transitive dependencies of @lhci/cli (Lighthouse CI) whose only fix is a breaking downgrade to @lhci/cli@0.6.1; they sit below the CI gate and are left in place rather than break the Lighthouse step.

Circuit Breaker — wraps a protected call in a state machine (Closed/Open/Half-Open) that trips to fail-fast once failures cross a threshold, preventing resource exhaustion and cascading failure (Michael Nygard, Release It!, 2007; Martin Fowler; proposed by @rehsack in #630)

Strangler Fig — incrementally replaces a legacy system by routing functionality to new code around its edges until the old system can be retired, avoiding a big-bang rewrite (Martin Fowler, 2004; proposed by @rehsack in #631)

Team Topologies — organises delivery around four team types (Stream-aligned, Enabling, Complicated-Subsystem, Platform) and three interaction modes for fast flow (Matthew Skelton & Manuel Pais, 2019; proposed by @rehsack in #632)

Twelve-Factor App — twelve practices for portable, scalable SaaS (explicit dependencies, config in the environment, stateless processes, disposability) (Adam Wiggins, Heroku, 2011; proposed by @mc2259 in #622)

Separation of Concerns — study and change each distinct aspect of a system in isolation, so one can be reasoned about without interference from the others (Edsger W. Dijkstra, 1974; proposed by @rehsack in #627)

Unix Philosophy — make each program do one thing well and compose small sharp tools via pipes (Doug McIlroy, 1978; Eric S. Raymond, The Art of Unix Programming, 2003; proposed by @rehsack in #628)

Design by Contract — specify routines with preconditions (caller’s obligation), postconditions (supplier’s guarantee) and invariants, assigning blame precisely when a contract is violated (Bertrand Meyer, Eiffel, 1986; proposed by @rehsack in #633)

OKR (Objectives and Key Results) — couples a qualitative Objective with a few measurable Key Results to focus and align goals (Andy Grove at Intel; popularised by John Doerr, Measure What Matters, 2018; proposed by @rehsack in #635)

Fermi Estimation — decompose an unknown quantity into estimable sub-quantities and reason in powers of ten to land the right order of magnitude (Enrico Fermi; proposed by Ralf D. Müller in #608)

Eisenhower Matrix — plots tasks on urgency × importance to separate the two and prioritise accordingly; admitted deliberately with a fetch-verified Criticism section (mere-urgency effect; Covey’s own caution) (Dwight D. Eisenhower; popularised by Stephen Covey; proposed by @rehsack in #623)

Laddering (Interview Technique) — reconstructs the means-end chain attribute → consequence → value through repeated "why is that important to you?" probing (George Kelly; Hinkle; Reynolds & Gutman; proposed by @rehsack in #634)

Consent vs. Consensus — distinguishes decision by absence of a reasoned objection ("good enough for now, safe enough to try") from decision by everyone’s active agreement (Gerard Endenburg, Sociocracy; Brian Robertson, Holacracy; proposed by @fdesoye in #626)

req42 — the pragmatic, agile requirements framework structured as twelve "drawers" mirroring arc42; the requirements companion to arc42 (Peter Hruschka & Markus Meuten; proposed by @rehsack in #625)

Curse of Knowledge — once you know something you cannot reconstruct not knowing it, so experts overestimate shared context (Camerer, Loewenstein & Weber, 1989; popularised by Chip & Dan Heath; proposed by @rehsack in #638)

Four-Sides Model (Schulz von Thun) — every message carries four sides at once — fact, self-revelation, relationship, appeal — and is heard with four matching "ears" (Friedemann Schulz von Thun, 1981; proposed by @rehsack in #636)

Dreyfus Model of Skill Acquisition — five stages from Novice to Expert describing how rule-following gives way to intuitive, context-driven mastery (Stuart & Hubert Dreyfus, 1980; proposed by @rehsack in #637)

Effective Java — Joshua Bloch’s catalogue of ~90 concrete best-practice "Items" for idiomatic, robust Java (Joshua Bloch, 3rd ed. 2018; proposed by @schtiefl in #618)

ADDIE Model — the instructional-design process of Analysis, Design, Development, Implementation, Evaluation (Florida State University origin; codified by Robert Maribe Branch, 2009; proposed by @davidlohner in #621)

The Explaining and Teaching contract now adds a single clause (EN + DE): for a long or multi-session explanation, persist the running understanding-checklist as a file so it survives context loss and can be resumed. An empirical A/B run — our contract versus the source prompt it was distilled from, both driven headless in a clean room with no CLAUDE.md leakage — showed the source prompt’s one genuine advantage was the durable checklist file it wrote to disk; everything else (honoring opt-out, one step per turn, not announcing the method) our contract already did better. This closes that gap without sacrificing the scale-to-the-question behaviour.

Search results were invisible: typing in the search box filtered the catalog grid, but the hero section above it pushed the results below the fold. An active search now collapses the hero with a smooth 300ms animation (it slides back the moment the query is cleared), so results appear directly under the header. Honors prefers-reduced-motion.

Search controls only where they work: the header search box, role filter, and anchor counter used to appear on every page while being dead outside the home catalog (no grid to filter, the counter stuck at "0 / 0"). The whole control row is now shown only on the home route.

Every semantic contract now has its own pre-rendered page at /contract/<id>, with a German variant at /de/contract/<id> — own title, description, canonical URL, social-share metadata, and hreflang pairs, exactly like the anchor pages from #597. Sharing a single contract (e.g. Specification) now produces the right preview instead of the generic contracts-page card.

The cards on /contracts link to their detail pages; in the SPA, a /contract/<id> URL opens the contracts page scrolled to the highlighted card with the real contract title in the tab.

All 38 contract pages joined the sitemap, and contracts are now DefinedTerm entities in the site’s structured data alongside the anchors — each citable by search engines and retrieval-grounded AI under its own URL.

Per-context output files (#531) — Phase 1 now writes QUESTION_TREE-<context>.adoc and OPEN_QUESTIONS-<context>.adoc instead of fixed filenames, so the documented "one bounded context at a time" workflow no longer silently overwrites earlier runs; Phase 2 reads the context-specific tree, and ADRs get a context prefix (adrs/<context>-adr-NNN-*.adoc) so two contexts cannot clobber each other’s decision records.

Fixed Q3.2 constraints sub-level (#532) — the Architecture Constraints node now always emits Q3.2.1 technical, Q3.2.2 organizational/process, and Q3.2.3 conventional constraints, mirroring arc42 Chapter 2’s constraint kinds; for this branch the prompt explicitly admits non-code evidence (CLAUDE.md/AGENTS.md, CONTRIBUTING, CI workflows, linter configs), closing the gap where process and convention constraints surfaced untraceably in Phase 2 output.

The remaining 14 versioned/living anchors now carry a Current Status section (EN + DE), all facts fetch-verified: arc42 (v9.0, July 2025, Section-10 split), C4 (levels stable, site rewritten 09/2024), Diátaxis (core stable, site grown), GitHub Flow (Chacon’s 2011 continuous-deployment original vs GitHub’s deploy-free current docs), PEAA (catalog unchanged since 2003, eaaDev frozen since 2006), BEM (definition stable, practice moved to utility-first — State of CSS no longer tracks methodologies), IEC 61508 (Ed 2 still current; Ed 3 forecast ~10/2026), LINDDUN (now a GO/PRO/MAESTRO family; no "MAID"), SRE (books canonical; platform-engineering convergence), SOTA (Papers with Code — the canonical tracker — no longer exists), Chain of Thought (reasoning models internalise it; OpenAI advises against "think step by step"), LLM-Evaluations (MMLU already superseded), Fagan Inspection and CRC Cards (definitions intact, practice moved on).

The five thin-prior anchors (GoM, IOSP, Meaningful Human Control, todo.txt-flavoured Markdown, LASR) get verified canonical sources plus explicit supply-the-definition advice.

LASR rewritten: verification revealed the previous entry was a silent substitution — it described a fabricated "lightweight architecture documentation template" with an invented acronym expansion (and the AgentSkill catalog carried a second, different invented expansion). The real LASR is the Lightweight Approach for Software Reviews (Toth & Zörner, lasr-reviews.org, "Reviewing Software Systems", Leanpub 2023), a streamlined ATAM alternative. The corrected anchor documents its own failure case as evidence for the thin-prior warning.

The About page now states the lexicon positioning: inclusion means precision, not endorsement.

The remaining twelve contested anchors from the catalog triage now carry a Criticism section (EN + DE), every claim with a named critic and a fetch-verified linked source: 4MAT (learning-styles evidence — Pashler et al. 2008, Kirschner 2017), AIDA (hierarchy-of-effects critique — Heath & Feldwick 2008, Vakratsas & Ambler 1999), Clean Architecture (indirection tax — Bogard, Comartin; alternative: Vertical Slice Architecture), CQRS (author-acknowledged overuse — Fowler, Dahan, Young), DRY (the authors' own "we did a poor job of explaining" concession plus Sandi Metz), Freytag’s Pyramid (Jane Alison), GoF Design Patterns (Norvig’s 16-of-23, Paul Graham’s human compiler), Hero’s Journey (folklorists Dundes and Toelken), Jobs To Be Done (the Christensen-vs-Ulwick two-schools problem), Kotter’s 8 Steps (no empirical validation of the whole model — Appelbaum et al. 2012; Kotter’s own Accelerate revision), MVP (term dilution — NN/g; Kniberg’s skateboard corrective), and Save the Cat (Suderman’s "Save the Movie!").

With batch 3, all 23 Criticism candidates and the 6 highest-drift Current Status candidates from the #603 triage are done — 28 anchors carry sourced sections.

The six anchors whose referent has editions or successors now carry a Current Status section (EN + DE) naming what is current, which edition an LLM’s training-data prior most plausibly serves, and how to disambiguate in prompts — every claim fetch-verified by research agents: OWASP Top 10 (2025 edition is current; bare category IDs are ambiguous — A03 means Injection in 2021 but Software Supply Chain Failures in 2025), ISO/IEC 25010 (2023 revision: nine characteristics, Safety added; prior serves the withdrawn 2011 model), MADR (4.0.0; prior reproduces the 2.x-era template), Effective Go (deliberately frozen by the Go team since 2009, predates generics and modules), Bloom’s Taxonomy (1956 nouns vs 2001 verbs — priors blend both), and Cockburn Use Cases (Use-Case 2.0 is an anchor of its own; the official Use Case 3.0 ebook exists since 2024 but its prior is thin — supply the definition in the prompt).

New contribution convention: every new anchor proposal now includes a criticism / edition-drift research step; findings land in Criticism / Current Status sections with named, fetch-verified linked sources. Documented in CONTRIBUTING (EN + DE) and the anchor template.

Fixed the Wolf Schneider criticism source from batch 1: it now links the concrete Sprachlog post in which Anatol Stefanowitsch calls Schneider the "oberster Sprachnörgler der deutschsprachigen Journaille" (Sprachbrocken 49/2012), instead of a category archive.

The catalog is a lexicon, not a list of endorsements — and a lexicon marks its contested entries. Following a full triage of all 161 anchors (#603), the ten anchors with the most prominent documented critique now carry a Criticism section (EN + DE) citing named, linked sources and the alternatives the discourse proposes: GoF Singleton (Gamma 2009, Hevery 2008), Testing Pyramid (Testing Trophy, Spotify honeycomb), SWOT (Hill & Westbrook 1997), Plain English / Strunk & White (Pullum 2009), Gutes Deutsch / Wolf Schneider (descriptive-linguistics critique), Postel’s Law (Allman 2011, RFC 9413), Red/Green TDD ("Is TDD Dead?" debate, Coplien), CAP Theorem (Brewer 2012, Kleppmann 2015), Five Whys (Card 2017, Allspaw 2014), and SOLID (North’s CUPID, Ousterhout). The MBTI anchor pioneered this format.

Per-page Open Graph / Twitter cards — every pre-rendered page now carries its own og:url, og:title, og:description and twitter:* tags (previously all subpages shipped the home-page values, so sharing an anchor link rendered the generic home preview). German pages get og:locale=de_DE.

Accurate home description — the hard-coded "110+ semantic anchors" claim (real count: 161) is replaced by counts computed from the data files at build time; the static shell texts are now count-free so they cannot go stale.

Normalized snippet lengths — meta descriptions are capped at ~160 characters on a word boundary; very short anchor extracts get the anchor title prefixed.

Every /anchor/<id> URL was an HTTP 404, served only through the JavaScript SPA fallback — while the sitemap and the DefinedTerm structured data advertised all 161 of them. Each anchor is now a real pre-rendered page (own title, description, canonical URL, content expanded), plus a German /de/anchor/<id> variant for the 155 translated anchors, with hreflang pairs. The static home catalog now links to these pages directly.

Static navigation — the HTML shell shipped empty header/footer placeholders and not a single <a href> link, so without JavaScript the site was a dead end for crawlers, LLM fetchers, and no-JS visitors. The shell now carries a real header (logo + nav) and footer with plain clean-URL links; with JavaScript the SPA hydrates on top exactly as before.

Static home catalog — the full category → anchor catalog (161 anchors) is pre-rendered into the home page, each anchor linking to its stable section in the Full Reference (/all-anchors#<anchor-id>).

Static German pages — the home page and every doc page with a German translation are pre-rendered under /de/… with honest hreflang pairs (the previous ?lang=de alternate served identical English HTML). A /de URL switches the SPA to German for JS users.

The legacy #/route hash links in the JS header/footer were replaced with the same clean URLs.

Structured data — added a standalone Organization entity and a DefinedTermSet with a DefinedTerm for every anchor (name, canonical URL, and a definition where available), generated at build time from anchors.json. Lets search engines and retrieval-grounded AI resolve "Semantic Anchors" as a distinct entity and each anchor as a defined term (#579).

Fixed: the An Anchor Delivers Only as Far as the Prior Reaches article was not pre-rendered and was therefore invisible to search engines and LLM crawlers. It is now pre-rendered like every other doc page.

Landing answer block — the home page rendered its hero (the "what are Semantic Anchors?" definition) only client-side, so crawlers and LLM fetchers saw an empty shell. The hero copy is now pre-rendered into the static home page, single-sourced from the translations so it never drifts from the live hero (#580).

Strategic Architecture Analysis — a strategic-analysis recipe that combines four lenses with the questions they answer: Wardley Mapping (how each component evolves), the Cynefin Framework (match the response to the domain), a Morphological Box (explore the option space deliberately instead of anchoring on the first design), ATAM (weigh tradeoffs against the quality goals), and the Five Whys (drill to the root cause); for build-vs-buy decisions, architecture-fitness reviews, and strategic technology-radar reviews

An Anchor Delivers Only as Far as the Prior Reaches — added Fable 5 (Anthropic’s newest model, probed on its release day) as a fourth column in the prior-mapping matrix. It holds the richest Use-Case 2.0 prior of the four models yet still hedges on Use-Case 3.0, strengthening the finding that a dense 2.0 prior does not buy a 3.0 one.

Explaining and Teaching — reworked from a structured "explain in 4MAT order" loop into a question-driven teaching dialogue: open by having the learner restate what they know, then one small step per turn (ask, or explain a little and check, then stop and wait), quiz with multiple-choice questions whose answer is withheld until the learner commits, and don’t advance until they can apply it to a fresh case. The anchors (Socratic Method, 4MAT, Naur, Feynman, Bloom’s, Definition of Done) now shape behaviour as attributions rather than named steps, with an explicit "don’t announce the method" guard so the scaffold no longer leaks into the output. Also keeps the over-fire brake (a one-line question gets a one-line answer) and learner opt-out

An Anchor Delivers Only as Far as the Prior Reaches — how a semantic anchor’s power depends on how densely the concept sits in an LLM’s training data, with a reproducible A–E experiment across Claude Haiku 4.5, Sonnet 4.6 and Opus 4.8. Prompted by @simasch's #582 on the Cockburn Use Cases anchor.

IOSP (Integration Operation Segregation Principle) — a function shall either contain logic (Operation) or call other functions (Integration), but never both; the formal refinement of SRP at function level; separates integration (coordination/sequencing) from operation (business logic/computation); eliminates mock-heavy tests by avoiding new’ing in functions that contain logic; formally checkable via IospAnalyzer (Roslyn); reduces the need for Dependency Injection (DIP) at the function level (Ralf Westphal, 2022; Stefan Lieser, 2025; proposed by @JensGrote in #556)

AIDA Model — the classic copywriting/advertising funnel: Attention → Interest → Desire → Action; a sequential hierarchy-of-effects model that makes persuasive copy do each job in turn (hook, relevance, wanting, call to action) instead of jumping to the ask; pins the canonical four-stage form and notes variants (AIDAS, AIDCA, AIDA-R); fills the persuasion/copywriting gap in the communication cluster (commonly attributed to E. St. Elmo Lewis, 1898)

TDD, Hamburg Style — Ralf Westphal’s design-led TDD recipe: close the requirements/logic gap before writing code via the ACD cycle (Analyze → Design → Code), then test at service boundaries with minimal mocking; composes Red-Green-Refactor, London School, Chicago School, and IOSP (proposed by @kuerm in #557, follows from #458)

DRY (Don’t Repeat Yourself) — "every piece of knowledge must have a single, unambiguous, authoritative representation within a system"; targets duplicated knowledge/intent rather than coincidental textual similarity, with the Rule of Three (Fowler) and the wrong-abstraction caution ("duplication is far cheaper than the wrong abstraction", Sandi Metz) as guardrails; antonym WET (Andy Hunt & Dave Thomas, The Pragmatic Programmer, 1999; proposed by @GaboCapo in #560)

Event Storming according to Alberto Brandolini — a collaborative workshop that models a domain as past-tense domain events on a timeline using a fixed colour notation (orange Event, blue Command, yellow Aggregate, lilac Policy, green Read Model, pink External System, red Hotspot) across three levels (Big Picture, Process Modeling, Design Level); surfaces bounded-context seams via pivotal events and makes assumptions explicit as hotspots; Reverse Event Storming reconstructs legacy systems (Alberto Brandolini, 2012/2013; proposed by @SidekickJohn in #558)

Bloom’s Taxonomy — the six-level hierarchy of cognitive learning objectives (Remember, Understand, Apply, Analyze, Evaluate, Create) with measurable action verbs per level; pins the Revised 2001 version and its knowledge dimension; complements the teaching cluster (4MAT, Feynman, Socratic) by grading the cognitive level a step targets — so "demonstrated understanding" can be set at Apply/Analyze rather than recall (Benjamin Bloom, 1956; Anderson & Krathwohl revision, 2001)

Explaining and Teaching — a gated teaching loop for "explain / why / how" requests: explain and verify understanding. Composes 4MAT (Why→What→How→What-If), Programming as Theory Building (the why behind the design), Socratic Method (diagnose first), Feynman Technique (explain-it-back), Bloom’s Taxonomy (understood = Apply/Analyze, not recall) and Definition of Done (a checklist demonstrated, not nodded at); scales to the size of the question

Meaningful Human Control (MHC) — the requirement that humans keep genuine, substantive control over autonomous systems making high-stakes decisions, rather than formulaic "human-in-the-loop" oversight; demands situational awareness, an identifiable accountability chain that cannot pass to machines, positive human authorization of critical actions, and timely intervention; Sharkey’s five levels (L1 human deliberates → L5 fully autonomous) grade the autonomy; rooted in the autonomous-weapons / international humanitarian law debate (Article 36, 2013) and now underpinning the EU AI Act’s human-oversight rules (proposed by @JensGrote in #543, issue #540)

Conway’s Law — organizations produce designs whose structure mirrors their communication structure; underlies team-topology and module-boundary reasoning and the Inverse Conway Maneuver (Melvin Conway, 1968; #506)

CAP Theorem — a partitioned distributed system must trade Consistency against Availability; reframed by PACELC for the no-partition case (Brewer 2000; Gilbert & Lynch 2002; #508)

Fallacies of Distributed Computing — the eight false network assumptions (reliable, zero latency, infinite bandwidth, secure, fixed topology, one admin, zero transport cost, homogeneous) as an audit checklist for distributed designs (L. P. Deutsch & J. Gosling, Sun; #507)

Law of Demeter — "only talk to your immediate friends"; the Principle of Least Knowledge against train-wreck call chains (Ian Holland & Karl Lieberherr, Northeastern, 1987; #509)

First Principles Thinking — decompose a problem to irreducible truths and reason up, instead of reasoning by analogy (roots in Aristotle/Descartes; #510)

Postel’s Law — the Robustness Principle: "be conservative in what you send, be liberal in what you accept", with the modern strictness/versioning caveat (Jon Postel, RFC 761, 1980; #512)

Goodhart’s Law — "when a measure becomes a target, it ceases to be a good measure"; a critique lens for KPIs and LLM-evaluation criteria (Charles Goodhart 1975; Strathern’s formulation; #513)

Inverted Pyramid Style — the journalistic news structure: most important information first, with a long prunable tail; distinct from BLUF (short) and the Pyramid Principle (complete argument) (proposed by @mhchem in #538)

Quality Attribute Scenario — Bass, Clements & Kazman’s six-part scenario format (SEI, Software Architecture in Practice) that turns a vague quality goal into a testable statement: Source, Stimulus, Artifact, Environment, Response, Response Measure; the Response Measure is what makes it verifiable. The format underlies arc42 Chapter 10 quality requirements and the leaves of an ATAM utility tree.

Luhmann’s System Theory — Niklas Luhmann’s sociological systems theory; autopoiesis, operational closure, structural coupling, double contingency, communication as the operation of social systems; a theory for analysing complex socio-technical systems where boundaries are contested (proposed by @JensGrote in #517)

Simon’s Constructivism — Fritz B. Simon’s introduction to systems theory and constructivism; viability vs. truth, trivial vs. non-trivial machines, second-order cybernetics, perturbation instead of instruction (proposed by @JensGrote in #517)

Systemic Consulting (Heidelberg School) — the Heidelberg School of systemic consulting after Fritz B. Simon; all-partiality, neutrality, circular questions, context clarification, intervention as irritation rather than instruction (proposed by @JensGrote in #517)

Architecture Documentation — now references the Quality Attribute Scenario anchor and requires Chapter 10 quality scenarios to be written in the six-part form with a literal Response Measure.

Architecture Documentation — sharpened with cross-section traceability and arc42 chapter-level rules. The contract now directs scaffolding the arc42 "with-help" template via docToolchain downloadTemplate instead of restating chapter structure; adds five cross-section traceability rules (quality goal → strategy, context ↔ building blocks, building block → runtime, Chapter 9 ADR index, building-block detail); separates Chapter 11 Risks from Technical Debt with probability/impact/priority columns; wires ADR Consequences to Chapter 11 risk IDs; and clarifies that Chapter 1.2 holds the top 3-5 quality goals while Chapter 10 may elaborate further characteristics (#502, #503, #504, #505)

Socratic Code Theory Recovery — the Q4 quality wording now states the Chapter 1.2-vs-10 relationship: Chapter 1.2 names only the top quality goals, Chapter 10 covers all eight characteristics, each marked as concretising a top goal or as derived (#505)

Socratic Code-Theory Recovery skill — Phase 1 decomposition is now adaptive instead of near-fixed: a node is a leaf only when answerable from specific file:line evidence, otherwise it decomposes further (capped at four levels below a fixed node). Tree depth tracks code density, so a large bounded context no longer collapses into one thin leaf per arc42 chapter. Directory-level evidence is no longer valid. The Phase 2 prompt wording was corrected to match the schema — claims trace back to a tree leaf as a build-time check, Q-IDs never appear in the final documents (#522, #524)

Hoshin Kanri — Yoji Akao’s Lean strategy-deployment discipline; 3-5 year True North objectives cascade into annual hoshin via the X-Matrix, with two-way catchball negotiation and monthly Bowling-Chart reviews; deliberately restricted to the few vital goals (proposed by @jaoltr in #468)

Double Diamond — UK Design Council’s design-process model (2005, expanded 2019); two divergent-convergent cycles — Discover/Define (problem space) and Develop/Deliver (solution space); "design the right thing, then design the thing right" (proposed by @jaoltr in #467)

Decisional Balance Sheet — Janis & Mann’s four-cell decision matrix (1977) extending Franklin’s "moral algebra"; surfaces utilitarian and social trade-offs, used in decision coaching and Motivational Interviewing to resolve ambivalence (proposed by @Mabla70 in #448)

Kano Model — Noriaki Kano’s two-dimensional quality model classifying features as Must-be, Performance, Attractive, Indifferent or Reverse, surveyed via paired functional/dysfunctional questions; pairs naturally with MoSCoW for backlog prioritisation (proposed in #459)

Kotter’s 8-Step Change Model — John P. Kotter’s classic transformation framework from Leading Change (HBR 1995, book 1996): urgency, guiding coalition, vision, communication, empowerment, short-term wins, consolidation, anchor in culture; the model is the inversion of the eight common errors that kill transformations (proposed in #460)

XY Problem — Communication anti-pattern from technical-support culture, coined by Mark Jason Dominus (2001) and popularised by Eric S. Raymond ("How To Ask Questions The Smart Way"); the asker requests Y when the real goal is X (proposed by @dernerl in #457)

Specification workflow restructured — Step 4 now builds the spec in layers: Persona Use Cases (Cockburn Fully Dressed, User Goal level), System Use Cases (technical interfaces: API, CLI, events, file formats), and Supplementary Specifications (Entity Model, State Machines, Interface Contracts, Validation Rules). Aligned with Martinelli’s AIUP distinction between Business and System Use Case Specifications.

Cockburn Use Cases — Alistair Cockburn’s structured textual use case format with Goal Levels (Summary, User Goal, Subfunction) and Actor-Goal List discovery technique

Socratic Code Theory Recovery — Recover a program’s "theory" (Naur 1985) from source code through recursive question refinement; composes Socratic Method, arc42, ISO 25010, Nygard ADRs, Mental Model (Naur)

Single Level of Abstraction Principle (SLAP) — Kent Beck’s Composed Method, codified as a Clean Code function-design rule by Robert C. Martin (proposed by @eirikbell in #440)

Occam’s Razor — William of Ockham’s parsimony principle applied to explanations, debugging and architecture rationale (proposed by @danielschmeiss in #439)

Code Smells — Kent Beck / Martin Fowler / Robert C. Martin catalogue of surface indications pointing to deeper design problems (proposed by @ma7tz in #435)

What Would Chuck Norris Do? (WWCND) — Tier 3 disposition-activator for direct commitment under uncertainty; originally rejected in #426, re-proposed by @cornelius in #438 with a 16-page empirical validation (3 models × 19 prompts × N=2 = 114 manually scored responses) that passed all four catalog criteria

Lehman’s Software Classification — Meir M. Lehman’s S-/P-/E-type taxonomy and Laws of Software Evolution

Red/Green TDD — Kent Beck’s classical TDD cycle (proposed by @petehodgson-tribe in #412)

4MAT — Bernice McCarthy’s learning cycle: why → what → how → what if (proposed by @rweisleder in #413)

Walking Skeleton — Alistair Cockburn’s end-to-end production-capable skeleton (proposed by @cpoepke in #420)

Tracer Bullet — Hunt & Thomas’s exploratory end-to-end slice (proposed by @cpoepke in #421)

Thin Vertical Slice — Cockburn & Cohn’s end-to-end delivery technique (proposed by @cpoepke in #422)

Spike Solution — Kent Beck’s time-boxed XP research technique (proposed by @cpoepke in #423)

Minimum Viable Product (MVP) — Eric Ries’s lean-startup hypothesis validation (proposed by @cpoepke in #424)

Cohesion Criteria (Constantine & Yourdon) — Seven levels of module cohesion from coincidental to functional

Effective Go — The Go Authors' canonical guide to idiomatic Go

Three-Act Structure — Aristotle / Syd Field / McKee

Hero’s Journey (Monomyth) — Joseph Campbell / Christopher Vogler

Save the Cat! 15-Beat Sheet — Blake Snyder

Fichtean Curve — John Gardner / Janet Burroway

Freytag’s Pyramid (Five-Act) — Gustav Freytag / John Yorke

Story Circle — Dan Harmon

Kishōtenketsu — Classical Chinese/Japanese tradition

Hemingway Bridge — Tiago Forte’s session re-entry technique (proposed by @arampp in #389)

CRC-Cards — Ward Cunningham & Kent Beck’s collaborative OO design technique (proposed by @arampp in #386)

GRASP — Craig Larman’s 9 OO responsibility-assignment patterns

Vertical Slice Architecture — Jimmy Bogard’s feature-first architecture

KISS Principle — Keep It Simple (Tier 1)

P.A.R.A. Method — Tiago Forte’s knowledge organization framework

PERT — Program Evaluation and Review Technique (Tier 2)

ISO 25010 — Software product quality model

Semantic Contracts — composable project conventions that reference existing anchors. Select and download as semantic-contracts.md for AGENTS.md/CLAUDE.md.

Anchor Evaluations — multiple-choice evaluation framework testing 63 anchors across 10 LLMs. View results.

Rubber Duck Debugging — concept is fully covered by Socratic Method; no actionable LLM instruction on its own

DRY (Don’t Repeat Yourself) — design principle, but "make it DRY" is too vague as an LLM instruction

SPOT (Single Point of Truth) — too similar to SSOT; consolidated into SSOT

Chatham House Rule — well-defined rule, but no practical LLM application

Mental Model (Naur) → Programming as Theory Building (Naur) — sharpened focus on documenting the "why" behind code (★☆☆ → ★★☆)

Problem Space NVC → Nonviolent Communication (Rosenberg) — stronger trigger, clear 4-step process (★★☆ → ★★★)

Test Double (Meszaros) → ★★★ umbrella anchor with 5 new sub-anchors: Dummy, Stub, Spy, Mock, Fake — each a precise, actionable LLM instruction (★☆☆ → ★★★)

Jobs To Be Done — upgraded to ★★★, strong for landing pages and customer analysis

SOTA — upgraded to ★★★, "Recherchiere SOTA zu [Thema]" reliably activates research mode

Review agent now evaluates anchors against the tier system

★☆☆ proposals are rejected unless a concrete prompt pattern is demonstrated

Plain English (Strunk & White) in #177

Gutes Deutsch (Wolf Schneider) in #175

Gherkin in #171

SWOT in #169

GoF Design Patterns Umbrella — 23 GoF patterns as sub-anchors with tier system (12 Tier-1, 7 Tier-2, 4 Tier-3) in #159

SOLID Principles Umbrella — 5 SOLID principles as sub-anchors (SRP, OCP, LSP, ISP, DIP, all Tier-1) in #161

Myers-Briggs Type Indicator (MBTI) — proposed by @gernotstarke in #158. Thank you!

CQRS — contributed by @Nantero1 in #137. Thank you!

GoF Design Patterns — contributed by @Nantero1 in #140. Thank you!

Test Double (Meszaros) — contributed by @Nantero1 in #142. Thank you!

Event-Driven Architecture — contributed by @Nantero1 in #143. Thank you!

BDD (Given-When-Then) — contributed by @Nantero1 in #138, merged via #152. Thank you!

YAGNI — contributed by @Nantero1 in #139, merged via #152. Thank you!

ATAM — proposed by @rdmueller in #133

LASR — proposed by @rdmueller in #135

OWASP Top 10 — proposed by @Advisior in #130. Thank you!

GoM (Grundsätze ordnungsmäßiger Modellierung) — proposed by @raifdmueller in #126

Regulated Environment — proposed by @bit-jkraushaar in #120. Thank you!

MoSCoW — proposed by @rdmueller in #115

Chatham House Rule — proposed by @rdmueller in #101

Fowler Patterns (PEAA) — proposed by @rdmueller in #99

Morphological Box — proposed by @rdmueller in #67

IEC 61508 SIL Levels — proposed by @rdmueller in #67

Split from monolithic README.adoc into individual anchor files (#38)

Socratic Method — proposed by @rdmueller in #31

BLUF — proposed by @rdmueller in #31

Rubber Duck Debugging — proposed by @rdmueller in #31

Chain of Thought — proposed by @rdmueller in #31

Devil’s Advocate — proposed by @rdmueller in #31

Five Whys — proposed by @rdmueller in #31

Feynman Technique — proposed by @rdmueller in #31

MADR — proposed by @raifdmueller in #29

TIMTOWTDI — contributed by @rehsack in #28. Thank you!

MECE Principle — proposed by @ingo-eichhorst in #26. Thank you!

Nelson Rules — proposed by @rdmueller in #20

Control Chart (Shewhart) — proposed by @rdmueller in #21

SPC — proposed by @rdmueller in #22

DRY Principle — added in #19

SPOT Principle — added in #19

SSOT Principle — added in #19

SOTA — proposed by @raifdmueller in #16

todo.txt-flavoured Markdown — proposed by @raifdmueller in #14

Pyramid Principle — proposed by @rdmueller in #12

Mutation Testing — contributed by ThomasPeuss. Thank you!

BEM Methodology — contributed by Robert Nimax. Thank you!

Conventional Commits — contributed by @rehsack in #4. Thank you!

Semantic Versioning — contributed by @rehsack in #4. Thank you!

SOLID Principles — contributed by @bit-jkraushaar. Thank you!

TDD, London School

TDD, Chicago School

Problem Space NVC

Mental Model (Naur)

EARS Requirements

arc42

ADR (Nygard)

Pugh Matrix

C4 Diagrams

Docs-as-Code

Domain-Driven Design

Hexagonal Architecture

Clean Architecture

User Story Mapping

Impact Mapping

Jobs To Be Done

Cynefin Framework

Wardley Mapping

Property-Based Testing

Testing Pyramid

Diátaxis Framework

What Qualifies as a Semantic Anchor