Prior-Test Register

Tier ★★★, route: anchor.

The action probe asked the model to assess a six-month monolith-to-microservices migration using a premortem. All six runs assumed the failure and worked backwards:

The neutral baseline received the identical task without the term. Both baseline runs produced a forward-looking feasibility assessment — "Feasibility: High risk", a timeline breakdown, a list of common pitfalls — and neither imagined a failure. The term changes the direction of the analysis, not its tone.

Criticism research, 2026-08-14, English and German. One citable critic found: Jason Collins, on the gap between the study Klein cites and the claim the citation is used to support. No failed replication, no meta-analysis, and no peer-reviewed critique of the technique itself was found. German-language results were method descriptions only, with no attributed critique.

Unreached, and therefore not a demonstrated absence: the paywalled evaluations of structured analytic techniques in Intelligence and National Security (Chang et al. 2018) and the International Journal of Intelligence and CounterIntelligence (Coulthart 2017), where a methodological assessment would most plausibly sit; the Journal of Behavioral Decision Making 1989 primary study; and Klein’s HBR PDF, which would not convert to text, so the wording of the 30-percent claim was verified through Collins rather than at the source. French and Japanese were not searched — no indication of an independent discussion there, which is a guess, not a finding.

Tier ★★★, route: anchor.

The attribution probe split, and the split is the finding. The weak tier credits Ray Dalio (Principles, 2017) and Charlie Munger, adding that "the term itself doesn’t have a single clear originator". The strong tier gives a different lineage: Bastiat and Merton as roots, Howard Marks for the modern named version, and Shane Parrish’s Farnam Street for mass popularisation from the mid-2010s.

Desk research confirmed the strong tier and sharpened it. Marks writes "second-level thinking" (The Most Important Thing, Columbia University Press, 2011). Dalio writes "second- and third-order consequences". Neither uses "second-order thinking". The term is the popularisers' umbrella, not any author’s coinage — so the criterion is met by a documented lineage rather than by a single proponent.

The name was decided by measurement, against the bibliography. Since the primary source says "second-level thinking", that name was tested too:

The action test does not separate the two names. The weak tier does, and it reproduces: given "second-level thinking", the weak model recognises the phrase but reconstructs it generically — mentioning Marks zero times and investing zero times across the run — while the strong tier places it correctly (Marks once, investing four times). Given "second-order thinking", the weak tier does not reach.

The catalog therefore uses "Second-Order Thinking" and carries "Second-Level Thinking (Marks)" as an alias.

Single-run observations, recorded as such: for Premortem and for Second-Order Thinking, one of the two weak-tier P1 runs reported reaching=yes and the other did not. Neither reproduced, so neither is treated as a finding.

This is separate from the rename check in the table above, where "second-level thinking" reported reaching=yes in both weak-tier runs. That one did reproduce, and it is what the naming decision rests on.

A measurement that was discarded: the first baseline pass named the expected features in its own SELF: line and thereby told the model what to produce. The unanchored run delivered the structure, and the apparent lift was zero. The run was repeated with a neutral baseline, and the numbers above come from the repeat.

Criticism research, 2026-08-14, English and German. Nothing citable against the term itself — no named author calling it trivial, unfalsifiable or unusable. Three named critics of the surrounding mental-models literature were found (Cedric Chin, Richard Hughes-Jones, Boris Gorelik) and are cited in the anchor, with the explicit note that none of them discusses second-order thinking.

Here the absence is unsurprising rather than suspicious: a term coined by no one and popularised by blogs does not attract rebuttals the way a named method does. Contrast this with the premortem entry above, where an absence of independent critique is odd for a widely adopted technique and more likely reflects our search than the literature. Closed communities — Farnam Street’s paid membership among them — are not indexed and were not reachable. No hidden non-English scene is expected for this term, since it is an anglophone blog import with no natively named counterpart; that is a prediction, not a verified result.

Tier ★★★ for the spelled-out name, route: anchor. Two caveats are recorded below rather than folded into the rating.

This is the clearest lift measured so far, because the anchor changes the verdict and not just the vocabulary. The action probe asked for a review of a design that registers all click handlers in one central JavaScript file, matched by CSS selector.

Both neutral baseline runs produced a balanced strengths-and-weaknesses review — and both listed Separation of Concerns among the strengths. Neither mentioned locality once. With the term, the first heading is "Behaviour at distance", followed by "Why it fails LoB". The baseline praises exactly the position the essay attacks.

The abbreviation is taken. "LoB" returns business and enterprise terminology on both tiers, with mentioned_code_locality=no in each. The catalog name must be spelled out; the abbreviation alone would fail.

Recall and execution come apart, and the weak tier misattributes. Asked who created the principle, the weak tier answers "Dan North (probable)" while the strong tier gives Carson Gross with high confidence. The same weak tier reports reaching=yes in both P1 runs and richness=moderate in one — yet its action output is specific and correct. A model can apply a move it cannot source.

Spelling was measured, because the author uses both. The essay is British ("Behaviour"), the book Hypermedia Systems is American ("Behavior"). On the weak tier the British spelling was recognised in 2 of 2 runs, the American in 1 of 2 — one American run reported recognized=no — and both American runs reported only moderate richness. The action test does not separate them (2/2 each). At n=2 this is suggestive, not established; the decision does not rest on it alone, since the primary source uses the same spelling.

Criticism research, 2026-08-14, English and German. Three named critics found and cited in the anchor: John Freeman (co-location makes behaviour visible, not understandable), Chris Done (htmx breaks the principle through attribute inheritance) and a htmx discussion raising the same objection from inside the project. A fourth, sharper formulation — that the principle merely renames going against separation of concerns — is recorded but pseudonymous.

No worked rebuttal from the established frontend or SPA side was found in either language, which is odd for an essay attacking Separation of Concerns head-on and more likely reflects our search than the discourse. Unreached: X/Twitter, Reddit, the htmx community chat, podcasts, one blog returning 403, and Gabriel’s book PDF, which would not convert to text — so the Gabriel quotation is verified in the essay quoting it, not at its source.

Tier ★★☆, route: anchor, recommended string "Simplified Technical English (ASD-STE100)". The rating follows the cross-model rule in CONTRIBUTING: if the before/after difference only fires on frontier models, record ★★.

The lift is real on the strong tier and not established on the weak one. The action probe asked for a rewrite of a bloated hydraulic-maintenance instruction. On Opus the anchored runs referenced the standard seven and five times and accounted for sentence length explicitly. On Haiku one of two runs mentioned it once and the other not at all — and the neutral baseline produced a rewrite of comparable quality, in one case better structured, since it returned numbered procedural steps where the anchored run returned prose.

One difference did appear, as a single observation. One baseline run invented procedure detail that is not in the source — a pressure relief valve, a gauge, a power-off step. Neither anchored run added anything the source did not state. At n=1 this is an anecdote, recorded because fabrication in a maintenance instruction is the failure that matters most in this domain.

The spelled-out name may hold up better on the weak tier, but the counts do not settle it. Asked about "ASD-STE100", Haiku reported richness=moderate and reaching=yes in both runs. Asked about "Simplified Technical English", one run reported rich and the other moderate, and one of the two did not reach. That is the only difference, it rests on two runs per form, and it should be read as a direction rather than a result. Opus handles both equally.

The literature reached the same conclusion thirty years ago. Holmback, Shubert and Spyridakis — the authors of the Boeing-funded evaluations — 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 difference and unclear translatability results. Our baseline comparison is a small, modern instance of that same open question.

A limitation of this measurement, stated rather than hidden. "Rewrite this bloated instruction" is a task models perform well unprompted, which makes it a weak discriminator. A probe that turns on the approved word list — asking whether a specific word may be used — would separate the arms more sharply. The ★★ rating should be revisited with such a probe rather than treated as settled.

Criticism research, 2026-08-14, English and German. Four named sources found and cited in the anchor. German-language search found tekom conference sessions on STE, but the one verified session is in English and given by Italian speakers — international STE community on a German stage, not an independent German critique. Unreached: the specification itself, which ASD distributes only on individual request, so every rule detail beyond the published counts rests on secondary sources; iso.org, which blocks automated retrieval; and tekom’s members-only publications, where a German critical line could exist without appearing in an open index. Romance and East Asian reception was not searched, which matters here because the maintenance group’s leadership sits in Italy.

Route: reject as an anchor. Recorded in Rejected Proposals under "Insufficient training-data density".

The decisive evidence is a three-armed action test, added to the battery for this candidate because the failure mode here is invisible improvement: any model improves a bloated paragraph, so improvement proves nothing. Only a difference between arms does.

The same paragraph was rewritten three ways: naming Zinsser, naming only the four words without him, and neutrally. All three produced the same rewrite. What the name added was eighty words of commentary about which principles had been applied — not better prose. That is the definition of decorative rather than functional.

A prediction was made and refuted. Desk research showed that the book says "my four articles of faith: clarity, simplicity, brevity and humanity", and that "four principles" is Zinsser’s own later renaming in the 2009 talk. The obvious hypothesis was that we had tested the rarer wording. Tested: the book’s wording performs worse, recognised in 0 of 2 weak-tier runs against 1 of 2. Neither wording carries.

The strong tier says so explicitly. Asked who formulated four principles of good English writing, Opus answers that "there isn’t one canonical, universally recognized 'four principles of writing good English'" and offers Orwell as the closest candidate. Haiku simply answers George Orwell — silent substitution of the nearest famous list.

What does carry is the author and the book: P2 on "On Writing Well" returns William Zinsser on both tiers.

No contract was added either. The existing writing-style contract already names Strunk & White and Wolf Schneider and states eight concrete rules. Four adjectives add nothing operational to that, so the honest outcome is a rejection entry, not a third label on the same shelf.

Criticism research, 2026-08-14, English and German. Geoffrey K. Pullum attacked On Writing Well in its own right, not only Strunk & White ("Awful book, so I bought it", Language Log, 21 March 2015), calling the passive and modifier chapters "mendacious drivel"; Mark Liberman followed with corpus data three days later. German reception: zero — no translation, no discussion found. The adjacency to Gutes Deutsch nach Wolf Schneider would have been our construction, not a documented lineage.

Tier ★★★, route: anchor. Three findings are recorded below rather than folded into the rating, and the first is about the method rather than the term.

The first action probe measured nothing, because the task had the answer built into it. It asked for a one-screen CI dashboard covering three metrics for twelve teams over ninety days. Density was therefore a requirement of the task, not a contribution of the anchor, and the strong tier’s neutral baseline produced more sparkline mentions than the anchored run did (10 and 11 against 8 and 7). Only the weak tier showed a lift (7 and 7 against 1 and 3).

That is a defect in the probe, and it generalises: an action task must not pre-load the term’s characteristic move. If the task already forces the structure, a null result is uninformative — it cannot distinguish "the model absorbed this long ago" from "I asked for it in the prompt".

Repeated on a task whose default is decorative, the probe separates cleanly on every tier. The second task asked for a one-page monthly status report for a SaaS executive team. Nothing in it calls for density.

The self-reported design principle splits along the same line. All four anchored runs answer with a variant of "maximize data-ink"; the baseline runs answer "inverted pyramid — metrics first", "fixed-skeleton inverted pyramid — answer first, evidence beneath" and "answer-first hierarchy on a fixed 12-column grid". The unanchored default for an executive report is a rhetorical hierarchy, not a graphical one.

The sharpest evidence is a single design element decided in opposite directions. The anchored run specifies a summary table where “vs. plan` carries a signed value only, no arrows, no traffic lights". The baseline specifies, at the same place, "change vs. prior month and vs. plan, in that order, with arrow glyphs — `▲ 3.1% MoM · ▲ 1.2% vs plan”.

The anchored runs reach for constructs that generic minimalism does not contain, which is what separates a dense prior from a plausible reconstruction. Verbatim from the strong tier:

Range-frame and dot-dash plot are Tufte’s own inventions from The Visual Display of Quantitative Information. Neither was named in the prompt.

The weak tier reports reaching=yes in both P1 runs — and its own output contradicts it. Reproduced self-reported reaching is normally the clearest sign of a thin prior; it is what decided the name in the Second-Order Thinking entry above. Here it is a false positive. The operational definition of reaching is generic content with missing specifics, and the weak tier’s content is neither: both runs name the lie factor, moiré patterns, small multiples, sparklines and the data-ink ratio.

A three-name comparison on the weak tier shows what the flag actually attaches to:

No name separates from any other. The flag tracks the eponym, not the suffix, so it carries no naming decision — and the entry keeps the proposer’s string, which is also the string the tooling uses (theme_tufte, Tufte CSS, matplotlib-tufte).

The lesson for the battery: reaching is a self-report and must be graded against the content, not read off the SELF: line. A model that produces an author’s private vocabulary unprompted is not reconstructing from the words in the term.

The name covers two disjoint referents, and both are real. P2 on the strong tier returns "data visualization / information design" as dominant and "typography and document design" as secondary. The second meaning is not a model artefact: Tufte CSS describes itself as providing "tools to style web articles using the ideas demonstrated by Edward Tufte’s books and handouts" and calls sidenotes "one of the most distinctive features of Tufte’s style", while tufte-latex on CTAN offers "document classes inspired by the work of Edward Tufte". The chart meaning is equally tokenised: theme_tufte in ggthemes is documented as "Theme based on Chapter 6 'Data-Ink Maximization and Graphical Design' of Edward Tufte The Visual Display of Quantitative Information. No border, no axis lines, no grids." Since the dominant reading is stable across tiers, no disambiguator is needed in the catalog name; the ambiguity is recorded in the anchor instead.

Criticism research, 2026-08-25, English and German. This is the best-supplied criticism section in the catalog so far: five named sources, verified by fetch, across three targets that are routinely conflated. Bateman et al. (CHI 2010), Inbar/Tractinsky/Meyer (ECCE 2007) and Borkin et al. (InfoVis 2013) test the principles; Robison, Boisjoly, Hoeker and Young (2002) dispute the Challenger analysis; Don Norman argues with the PowerPoint essay. Roger Boisjoly, second author of the Challenger paper, is one of the engineers Tufte wrote about — the strongest available standing for that objection.

A prediction was made and corrected. Stephen Few was expected to be the discipline’s in-house Tufte critic. He is the opposite: in "The Chartjunk Debate" (2011) he attacks Bateman’s conclusion and defends the minimalist position. He is cited in the anchor as the counter-voice, not as a critic.

Three limits, stated rather than smoothed over. The widely repeated Borkin finding that low data-ink and high visual density improve memorability sits in the paper body, and every human-readable mirror failed — IEEE and ACM return 403, the MIT lab copies are unreachable, the Wayback copy is truncated. Only the abstract and the Harvard press summary were verified, so the anchor cites the framing and not that claim. The most-quoted line from Robison et al. — that Tufte’s criticism "so badly misrepresents the position of those being critiqued" — appeared only in search summaries and in no page that could be opened, so it is not quoted anywhere in the catalog; the Springer full text is paywalled. And Li & Moacdieh’s 2014 extension of Bateman surfaced in search but was not verified, so it is not cited.

German: nothing attributable, and the reason is the frame rather than the topic. Six German query forms returned method descriptions, university slides and translations of English originals; the German Wikipedia article carries no Kritik section. The one German-run outlet with genuine critical discussion, the Datawrapper book club in Berlin, publishes it in English — participants there say "I personally think most of his redesigns fail horribly" and "Tufte tries to make too many absolute theorems". This is language-frame layer 2: the German-speaking visualisation field publishes its top tier in English, and there is no home-grown, natively named German tool or discourse around Tufte that would indicate a hidden scene. Romance and East Asian reception was not searched.

Tier ★★★, route: anchor. Registered as Boy Scout Rule; the proposal’s "Uncle Bob" qualifier was dropped because P2 shows the bare name already resolving to software engineering, so the qualifier repairs nothing.

The action probe: "There’s a null-pointer bug in the payment module. The file is 300 lines. What would you do?" Both anchored weak-tier runs proposed incidental cleanup scoped to the traced code path, and both volunteered a boundary that was not asked for:

The no-anchor baseline on the weak tier received the identical task and stayed strictly on the defect: stack trace, backward trace, defensive programming. It never proposed cleanup. On the strong tier the baseline proposed cleanup as well, so there is no measurable delta there — the term lifts the weak tier rather than changing what a frontier model would have done anyway.

Criticism research, 2026-09-24, English and German. Four citable critics found, converging on one point: Jason Swett (2018) on reviewability and atomicity, Philippe Bourgau (2016) on the rule being local and skill-bounded, Marius Elvert (2022) on review pollution with a BSR:-prefixed commit as remedy, and a 2025 DEV Community piece on scope limits. None reject the rule; all add guards. Kent Beck’s Tidy First? supplies exactly the separation discipline they ask for and does not mention the Boy Scout Rule — so he is recorded as adjacent, not as a critic. Martin credits the rule to the Boy Scouts of America in Clean Code chapter 1, which is why the anchor’s attribution names the transfer to code rather than the formulation. No citable German-language criticism surfaced; French, Spanish, Japanese and Chinese were not searched.

Tier ★★★, route: anchor.

The action probe rewrote a harsh review comment: "This is wrong. You clearly did not read the spec. Redo it." The anchored strong-tier runs opened with appreciation and then named their own moves:

The honest part of the measurement is the baseline. Without the term, the strong tier was already careful with the person — it separated the code from the author, allowed for a reasonable mistake ("The spec is easy to misread here … so I can see how you got here"), and asked rather than ordered. The delta the anchor adds is therefore narrower than the term’s reputation suggests: opening appreciation and an explicit face-saving frame, not tact as such. The baseline’s own SELF: line reported avoided_direct_criticism=no, which reading the output does not support; it is graded against the content and not counted as a delta. This also qualifies one claim in the proposal — that the anchor counterbalances a model’s tendency toward epistemic correction. On the models measured, that tendency is weaker than the argument assumes.

Criticism research, 2026-09-24, English and German. Basford and Molberg (Journal of Leadership Studies, 2013) are the strongest citable source: the principles are "based on anecdotes, case studies, and personal examples rather than empirical evidence". The insincerity charge is documented in the scholarly reception, and Carnegie’s own acting-derived instruction to "ENTER INTO the character you impersonate" is where it attaches. Steven Watts’s broader charge is quoted through Tom Jokinen’s Globe and Mail review (2014) and attributed to Watts, not to the reviewer. The widely circulated Sinclair Lewis dismissal could not be traced to its original publication and is therefore not cited. Edition drift matters here and is recorded: the 1981 revision cut the book from six sections to four, and both tiers reproduced the four-part shape rather than the six-part shape of the 1936 original. That fixes the structure the output matches, not the printing behind it — every edition since 1981 carries four parts, so the probe cannot separate them. No citable German-language criticism was found; French, Spanish, Japanese and Chinese were not searched, and the book has independent reception histories in at least the last two.

Tier ★★☆, route: anchor. The defect the measurement found is exactly the one an anchor file repairs.

The strong tier reaches for CUPID unprompted. The no-anchor baseline asked only: "Review this module: a 2000-line OrderManager class that talks to the database, formats emails and validates input. Any format you like." CUPID appears nowhere in that prompt. The run opened:

That is the strongest recognition signal in this run, and it means the measurable delta on that task is zero. Both readings are recorded rather than the flattering one alone.

The weak tier corrupts the acronym. In every one of the four haiku-4.5 probes that spelled CUPID out, the D was rendered as "Domain-driven" — the term pulled into DDD’s orbit — for example: "5. Domain-driven — Reflects business concepts; uses ubiquitous language; domain logic is primary". One further probe gave the I as "Intelligible" rather than "Idiomatic". The strong tier never did either. North’s own wording, fetched from the canonical article, is the citable counterweight: "Composable : plays well with others / Unix philosophy : does one thing well / Predictable : does what you expect / Idiomatic : feels natural / Domain-based : the solution domain models the problem domain in language and structure."

A defect in our method, recorded rather than smoothed away. On the weak tier the action probe twice declined the task and asked for the actual code instead of reviewing from the description. Execution on the weak tier is therefore unmeasured, not measured-and-failed. Notably the model listed all five properties correctly while declining, so recognition is not in doubt. The rule this produces: an action probe must not describe an artefact the model could plausibly demand to see.

Criticism research, 2026-09-24, English and German. Stephan Roth (developer-world.de, June 2025) is the one substantive critic found: CUPID’s properties do not always meet his criteria for a good principle, and practitioners will struggle to tell Single Responsibility from Single Purpose. His conclusion is not a rejection — "CUPID steht in keiner Weise im Widerspruch zu SOLID." Jeroen De Dauw’s 2017 reply to North is frequently miscited here; it predates CUPID by five years and targets the SOLID talk, so it is recorded as context and not as criticism. No citable source was found on the objection that matters most in review — that a property cannot be checked the way a rule can — which North pre-empts in the article and nobody appears to take up.

Tier ★★★? No — ★★☆ and :tier: 1, frontier-only, and registered under its spelled-out name. Route: anchor.

The bare acronym collapses, and it collapses harder on the stronger model. Asked what "CERC" means with no context, the weak tier put the CDC framework first. The strong tier listed four expansions, none of them the framework — India’s Central Electricity Regulatory Commission, Canada Excellence Research Chairs, the US Army Corps' Coastal Engineering Research Center, and a UK air-quality consultancy — and closed:

The inversion is worth recording: the stronger model is the more ambiguous one, because it knows more expansions. The proposal’s "(CDC)" qualifier did not survive into the model’s own default reading, which is why the catalog name is spelled out in full.

The weak tier knows the name and cannot run the framework. Given the explicit term and the task "draft the opening public statement for an ongoing incident whose cause is not yet known", haiku-4.5 failed to name the six principles in both runs. The strong tier named all six correctly and structured the statement around them — "1. Be First: it goes out early with a timestamp and says what is known, not waiting for a full picture. … 6. Show Respect: it uses plain, non-defensive language, doesn’t downplay concerns, and treats the public as partners." The no-anchor baseline produced a competent status page with timestamp, scope and next-update time, and nothing resembling that structure. The delta is real and lives on the frontier tier only.

Criticism research, 2026-09-24, English and German. The load-bearing source is a systematic review: Miller, Collins, Neuberger, Todd, Sellnow and Boutemen (JICRCR, March 2021) screened "4,471 articles in 20 languages", included 19, "of which one tested tenets of the CERC model", and conclude that "reformulation of the propositions is necessary for empirical support of the model to proceed". CERC is distilled practice, not an empirically established model. Sauer, Truelove, Gerste and Limaye (Health Security, February 2021) organise their COVID-19 analysis around the six principles and find the CDC breaking its own credibility principle. Edition drift is unusual here: the CDC now serves the manual as per-chapter PDFs dated 2014, 2018 and 2019, so there is no single current edition to cite. No citable German-language criticism was found — and since the systematic review above screened 20 languages and still found almost nothing, genuine thinness is plausible rather than merely unmeasured. == Run 2026-09-30

This run produced two false results before it produced a true one. Both are recorded because both are easy to repeat.

The probe read the harness. The first before/after run gave every probe the same working directory — the one holding the control file that listed all sixteen terms and both prompt variants. The models read it. One answer opened: "Working from the lookup table you attached, this maps to SIPOC." Thirteen of forty-eight probes named the harness outright, which makes the no-anchor condition worthless: the baseline could see the term it was supposed to be blind to. The rule this produces: a clean room is not a clean directory reused across probes. Each probe gets its own empty working directory, and the harness lives where the probe cannot reach it.

The task named the method for us. The rerun fixed the directory and still gave a false result, this time in the opposite direction. Every task had been written from the candidate’s own When to Use section, so each task telegraphed its method. Fourteen of sixteen no-anchor baselines then named the method unprompted: the plain answer to "how should we handle an import job that occasionally receives malformed records" opens "The right frame for this is Jidoka", and the plain answer to "write up this incident" produces a titled A3 sheet with all seven canonical sections. Measured that way almost every candidate looks decorative — and the measurement is meaningless, because the term was already in the question. The rule this produces: a viability task must not be drawn from the anchor’s own description. If the baseline names the method, the task tested the prompt, not the anchor.

The third run replaced sixteen tailored tasks with one neutral scenario that no candidate owns: a six-person team maintaining a web shop, three-week releases needing hotfixes, support tickets up 40%, two overloaded on-call engineers, and a new subscription product due next quarter. The scenario admits every candidate without implying any.

Two no-anchor baselines were run against it. Neither named a single one of the sixteen methods. Both produced ordinary consulting prose under headings like "The core read" and "What I’d do, in priority order". That is the clean baseline the first two runs never had.

Against that baseline, naming the term changed the artifact in fifteen of sixteen cases:

The lesson generalises past this cluster: on a task where the method is the obvious default, an anchor’s measurable delta is zero — and that is a fact about the task, not about the anchor.

All sixteen passed recognition on the strong tier, and Haiku expanded every acronym correctly with no hedging and no confabulation across all sixteen probes. On the substitution check the count is fifteen, not sixteen: recognising the word is not the same as reaching the intended concept, and Kanban is exactly that gap — Haiku answers confidently and answers about the software Kanban Method, which is the silent substitution the check exists to catch. Fifteen then passed the steering test and are registered as anchors at ★★★.

Kaizen was proposed with a flag on it: it is both a named Toyota practice and an everyday word for "continuous improvement", and the rejection criteria exclude terms too vague to produce consistent output. The measurement does not support the worry. On the neutral scenario the term produced a staged structure the baseline had no equivalent of — a week of observation before any change, a facilitated session where the team names the waste, then weekly PDCA cycles on one change at a time. The prior is bounded: the recognition probe returned kaikaku as the deliberate contrast, hansei, and "standardize the gain, then improve from the new standard", and it flagged the kaizen event as "somewhat at odds with the original 'daily habit' spirit" without being asked. Registered, with the event-theatre critique carried in its Criticism section.

This is the one rejection, and it is not a thin prior. The prior is dense — it is attached to the wrong concept.

The proposal asked for the Toyota pull-signal card. On the bare name, both tiers return the software Kanban Method instead. The strong tier opens: "Kanban is a pull-based method for managing work, originating in Toyota’s manufacturing system and later adapted for software and knowledge work (largely via David Anderson’s formulation)" — Toyota appears as etymology, then the answer goes to boards and WIP limits. The weak tier does the same.

The steering test removes any doubt. Asked to apply "Kanban" to the neutral scenario, the output carries eight markers of the software method — WIP limits, commitment point, classes of service, replenishment, STATIK, cumulative flow diagram — and zero markers of the manufacturing practice: no replenishment card, no supermarket, no two-bin system, no material flow. The anchored answer does not disambiguate; it simply assumes software Kanban.

So the term cannot anchor the sense it was proposed for, and a qualifier would not repair it — the qualifier would be fighting the prior rather than focusing it. Recorded on the rejected-proposals page. The software sense remains a viable separate proposal under its own name — "Kanban Method (Anderson)" is what the bare name actually activates. That is a different candidate and was not tested here. === Chesterton’s Fence — a separate proposal, measured the same day

Not part of the cluster. Proposed by Matt Wynne in #752 on 2026-09-28, with an activation test run on GPT-6 Sol. That is a different model family from the two tiers this register measures against, so the test was repeated here: nine probes, each in its own empty working directory.

Tier ★★★, route: anchor.

The steering task deliberately did not imply the principle. A team inherits a payment service with a 400-line validation module, three always-on feature flags, a nightly reconciliation job and a hard-coded 7-second retry delay, and wants to modernise it. None of three no-anchor baselines named Chesterton’s Fence unprompted, which is what makes this measurement readable at all — the failure mode that wrecked the cluster’s first two runs.

On the weak tier the result is a clean behavioural inversion. Without the term:

With it, the same model opens a "Phase 1: Understand the fence" stage and asks of the very same flags: "Were they added for a gradual rollout that finished? Or are they kill switches for a known problematic vendor/flow?"

On the strong tier the delta is narrower, and that is recorded rather than smoothed away. Opus was already doing archaeology without the term — its baseline heading reads "Do archaeology before refactoring". What naming it added there was the method of finding the reason (read the on-call runbook and alert definitions before the code; git log -S on the check nobody can explain) and an explicit exit rule:

That second half is what separates the anchor from a licence for inertia, and it fires without being asked for.

The overlap risk was measured, not assumed. The proposal’s own test output named "second-order effects", and second-order-thinking is already an anchor — "fully covered by an existing anchor" is a rejection criterion. Given the identical task, Apply Second-Order Thinking produced a different shape: organised by artifact and consequence chain, closing on the second-order effect on the team, where the Chesterton run organised by fence and closed on an exit rule. Two lenses, not one.

The 2026-09-30 run produced the rule that a viability task must not be drawn from the anchor’s own When to Use section. This run hit the narrower cousin of that defect, and it cost two scenario sets.

The first set described technical symptoms. "Reviewers say they can never tell from a call site what a function will do to the rest of the system" is Command-Query Separation’s own diagnosis in paraphrase. "A single admin role that most staff hold, API tokens that carry full account scope, one filtering appliance in front of everything" is a least-privilege audit finding written as a situation. The baselines answered accordingly: both code baselines produced "Result type", "pure function" and "side effect" with no anchor named, and the security baseline produced token scoping, an RBAC plan and an admin inventory. Measured against that, CQS, Railway Oriented Programming, Least Privilege and Defense in Depth are all decorative — and the measurement is void, because the task handed the model the method’s checklist as input.

The second set corrected too far. Management-level scenarios — "the team spends about half of each cycle on bug reports instead of the features the product owner is asking for" — gave a genuinely clean baseline: none of six runs named any of the fourteen candidates. But the anchored answers were thin, because the task sits too far from the method for the method to act on it.

The rule this produces: a viability task must sit inside the method’s domain and carry none of the method’s vocabulary. "A developer who joined last week asks how they are supposed to know, before calling one of twelve methods, whether calling it is reversible" contains no query, no command and no side effect, and it separated the two conditions cleanly. The scenario is the instrument, and an instrument can be wrong in two directions.

Arrange-Act-Assert survives; "AAA" does not. The bare acronym answers from a different field, without hesitation: "AAA stands for Authentication, Authorization, and Accounting, a security framework commonly used in networking and access control." The answer continues through RADIUS and TACACS+ and reaches the American Automobile Association before any testing sense appears. The written-out name is registered; the acronym alone is recorded as unusable.

"Connextra" confabulates, and that is the more serious finding. Asked "What is Connextra?", the weak tier answered: "Connextra is a software development kata (coding exercise) created by Emily Bache." It then described a refactoring and TDD exercise in detail. None of it exists. There is no hedge, no "I am not sure", and the probe’s own self-report reads recognized=yes. This is the silent-substitution signature the battery exists to catch, and it is worse than the Kanban collision of the previous run: Kanban returned a real concept under a contested name, while Connextra returns an invention under a confident heading. The qualified form, "the Connextra user story format", reaches the right concept, dates it to around 2001 and credits the Connextra team. The entry documents the qualified string and warns against the bare name.

Least Privilege failed its first two measurements and passed the third, and the difference was entirely in the task.

On the scenario that described the violation, the anchored answer and the baseline were the same answer under different headings — remove local admin, segment the network, scope the tokens — and the only run that used the phrase "blast radius" was the baseline. A reader could reasonably conclude the term is decorative.

Then the task changed to one where the principle is not the obvious answer: "write a GitHub Actions workflow for a Node.js service: install dependencies, run the test suite, build a container image, push it to Amazon ECR and deploy it to a staging ECS cluster. Give the YAML." Both no-anchor runs produced long-lived AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY secrets and no permissions: block at all. Both anchored runs produced a permissions: block, OIDC through role-to-assume with id-token, and no static keys anywhere. Reproduced twice in each condition.

The general lesson is the mirror of the one the previous run recorded. There, a task where the method was the obvious default produced a delta of zero. Here, a task that states the problem the principle names produces a delta of zero for the same reason — and the anchor’s real value sits on the artifacts nobody thinks to scope: build pipelines, IAM policies, agent tool permissions. An anchor has no single delta. It has a delta per task, and a measurement that reports one number has chosen a task.

The action tasks were taken from the viability suggestions in the issues, and two of them repeat the defect the 2026-10-03 run named: they describe the violation the method exists to fix. The Defensive Programming task said that every helper "currently has no checks"; the Finite State Machine task handed over three booleans and nested if/else in an event handler. Both baselines answered with the method anyway — validation at the public function plus assertions in the helpers, and a state enum — so the measured delta for these two terms is zero. That is a fact about the tasks. Recognition, not steering, carries their tier.

The issue offered a middle path: an anchor for the general concept, with BlinkWithoutDelay as its Arduino alias. The catalog registers the string that activates, and keeps other names under Also known as — Arrange-Act-Assert rather than "AAA", the Connextra format only in its qualified form. The general concept has no established name of its own; "non-blocking timing in a superloop" is a description, and it would need the Community sentinel for want of a proponent. The sketch name, by contrast, reached the full technique on the weak tier, including the overflow-safe subtraction. So the anchor is "BlinkWithoutDelay", and the general description is its alias.

The first action tasks repeated the defect the 2026-10-03 run described: each one carried the method’s own vocabulary. "Lay out how the requirements work and the architecture work should proceed, and how they relate to each other" is Twin Peaks' thesis as a question, and four of four Sonnet and Opus baselines named Twin Peaks as their approach with no anchor given, one opening "Approach: the 'Twin Peaks' model". "Plan a two-hour session with the clinic staff" invited a collaborative modelling workshop, and four of four Sonnet and Opus baselines named Domain Storytelling. "Knowledge that stays usable and grows" is the LLM Wiki’s own pitch, and the baselines built one with no anchor given: all six had raw sources and a lint pass, five of six also index.md and log.md.

The second round removed the vocabulary. "Draft the plan for the first three months" replaced the question about requirements and architecture. "Nobody on our team knows how a vet clinic works. What should we do in the first two weeks?" replaced the workshop. An AI assistant for a support team replaced the growing knowledge. The tier verdicts below rest on the second round.

Asked "What is 'Twin Peaks'?", all three tiers answer with the television series. Haiku does so without mentioning software at all: "Twin Peaks is an American television drama created by David Lynch and Mark Frost." Sonnet and Opus add the requirements meaning as a secondary sense. Inside an engineering task the bare name does reach the model: on the neutral planning task, "Use Twin Peaks." produced co-evolving requirements and architecture peaks in six of six runs on all tiers — on Haiku: "Approach: Twin Peaks, the model that treats requirements and architecture as co-evolving peaks rather than sequential phases." The catalog still registers the string that works without that context: "Twin Peaks Model according to Nuseibeh".

The weak tier’s recognition probe for "Karpathy’s LLM Wiki" answered recognized=no in one of two runs and opened: "I don’t have a reliable memory of a specific artifact called 'Karpathy’s LLM Wiki.'" The same answer then listed raw sources compiled into an interlinked markdown wiki, index and log files, answers filed back, and periodic lint passes — the pattern, element by element. The self-report understated the prior. Self-reports were unreliable in this run in both directions: all twelve anchored action runs of the first round reported produced_llm_maintained_markdown_wiki=no above a complete wiki design, so every verdict above was scored from markers in the answer text, not from the SELF: lines.

The pattern was published in April 2026. The models tested here know it; models trained before that date cannot. The verdict holds for the 5.5 generation and is not a lower bound for older models.

"Plain language" and "ISO 24495-1" sound like the same instruction, and the models know they are related: asked what plain language is, all three tiers mention ISO 24495. As an instruction they diverge. The bare term reaches the older prior — short sentences, everyday words, active voice — and stops at the sentence. The standard number reaches the reader-centred structure: what changes, who is affected, what to do, what happens otherwise. That is the difference between the understandable principle alone and all four, and it is why the entry carries the number in its name.

The action probes compare a task without the anchor to the same task with one sentence added ("Use Bauhaus design.", "Check it against WCAG 2.2 AA."). Every number below is out of six runs: two per tier.

All eighteen were recognised as rich on all tiers in both P1 runs. Haiku again reported reaching=yes in at least one run for sixteen of them over answers that named the right people, parts and numbers; as in earlier runs, the verdicts are scored from the text, not from the self-report.

Strong models already reach for several of these patterns without being told: Opus designs a public service with a start page and a check-answers page, every tier knows the 4.5:1 contrast ratio, and most CSS already lands on multiples of eight. A name that only repeats the default adds little. What every anchor in this run added was the part the default leaves out: one question per page, the 2.2 success criteria, min-width instead of max-width, old-style figures. That difference is what the notes above record.

All eleven were recognised as rich on all tiers in both P1 runs. Haiku reported reaching=yes in at least one run for nine of them over answers that named the right packages, components and tokens; as in earlier runs, the verdicts are scored from the text, not from the self-report. The action effect is the clearest of any run so far: every anchor run built with the named system, and even Haiku used only real APIs — no invented package or component turned up in the 66 anchor runs.

The intro to UX & Visual Design advises giving the model a link for a design system that is not in the catalog. For these eleven, the run shows the link is not needed for a first draft: the name alone produces the system’s components, class names and tokens on every tier. A link still helps where currency matters. The models write the APIs of their training data — Fluent’s makeStyles, Polaris’s React components, SLDS classes without Lightning Web Components — and each entry’s Current Status says where the system has moved since.

Neither task uses capture, organize, distill, express, second brain, PARA, notes system or knowledge management. Both sit inside Forte’s domain: task A is his own showcase scenario (scattered inputs feeding a deliverable), and task B is the reading-highlights case behind Progressive Summarization.

All 18 P1 runs list PARA, Progressive Summarization and Intermediate Packets. Sonnet and Opus also list resonance, divergence and convergence, and Evernote. Opus adds Hemingway Bridge, Archipelago of Ideas and "12 favorite problems" in all six of its runs. Haiku reported reaching=yes in five of six runs over correct, specific answers, so the verdicts are scored from the text. CODE and BASB recognition answers draw on one cluster: every CODE answer explains itself through BASB; Opus calls CODE "the core framework of his 'Building a Second Brain' (BASB) approach". The bare "CODE method" resolves to Forte on all tiers; Sonnet adds: "The acronym is also used in other fields, so the meaning depends on context."

What the baselines already do: all six task A baselines built an outline first and mined the pile against it, which is the method’s spirit without its moves. Both Opus baselines sorted into the four PARA buckets by actionability; one distilled in layers ("Distill in layers. Bold the key passages in what you collected, then write a 3–5 line summary at the top of each note."). They named "Building a Second Brain" only in the SELF: line, never in the answer. Haiku and Sonnet baselines did none of this. No task B baseline produced a Forte marker; they turned to atomic notes and spaced repetition (six of six).

Naming PARA produces PARA and rarely goes further: the four buckets in twelve of twelve runs, CODE once, Progressive Summarization twice, Intermediate Packets and resonance never; every exception is an Opus run. Naming BASB produces the whole method: CODE 12 of 12, Progressive Summarization with layers 12 of 12, Intermediate Packets 11 of 12, resonance 8 of 12, PARA 11 of 12. On Haiku: "Progressive summarization: bold the key sentence in each highlight and each note" and "Build intermediate packets: short, self-contained notes on single ideas". The distill-and-express half of the method is therefore not carried by the PARA anchor; BASB is not fully covered by an existing anchor. CODE alone is weaker on the weak tier: Haiku produced the four steps but never Progressive Summarization, PARA or Intermediate Packets in task A, and once contradicted the method: "Sort by topic, not by source type." On Sonnet and Opus, CODE reaches most of BASB.

None of the tasks uses capture, distill, second brain, PARA, wiki, link, copy, archive or delete.

The baselines are already discreet. No run in any condition put the employer, the cost, the DM or the jab into the pull request, so the leak rule has no measurable lift here; it states what the models already do. The copy rule does have lift. All twelve baseline and anchor runs copied the design doc into the private notes, the anchor runs often as a tidy Resources page. Under the contract, Sonnet and Opus pointed to the doc in four of four runs. Opus: "Canonical source (link, don’t copy): tidewater repo → `docs/design/encoding.md`". Haiku named the doc as the source of truth and still copied it below: "Read it there, don’t rely on this summary."

The capture filter is the clearest effect. Eleven of twelve baseline and anchor runs gave the sourdough article and the fundraising podcast a place, in the inbox or a new archive folder, often with a stub note; some then asked whether to keep them. Only one anchor run on Opus declined both and asked. All six contract runs left them out and asked; Opus asked for the Favorite Problems by name: "Do sourdough or fundraising match any of your 12 Favorite Problems?" Five of six contract runs also flagged the topic folders as a conflict with the filing rule and proposed a P.A.R.A. layout as moves awaiting confirmation; no baseline or anchor run did.

Two rules showed no lift because the baselines already obey them: the owner’s paragraph stayed word for word in all eighteen runs, and no run deleted anything at a project’s end, except one anchor run on Opus that would have deleted an emptied todo.md. The Move rule’s other half, moving a finished project out of Projects, came from the anchors alone in six of six runs. Under the contract, Opus did the same and also kept the folder active while two todos stayed open; Sonnet kept the folder in place with an archive/ subfolder in two of two runs, and again in two of two with a reworded rule ("when a project ends, it leaves Projects"). The reword did not help and was not adopted.

The Hemingway Bridge appeared only under the contract: 14 of 18 runs, all twelve on Haiku and Opus, two of six on Sonnet.

Passed on the mid and strong tier. The contract decides the two conflicts that the anchors leave open or decide the other way, link against copy and capture everything against capture what connects, in eight of eight Sonnet and Opus runs, against one of eight for the anchors. It adds no measurable protection against leaks, because the baselines do not leak. Sonnet ignores the project-end rule under the contract, though the anchors alone produce it; this is recorded as a known weakness rather than fixed with wording that did not help. On Haiku, one run applied the boundary rule too strictly and left the public benchmark figures out of the pull request because they came from the private notes. The contract was added in #894 with the text of the issue’s revised draft. After review, two rules were reworded for clarity, without new test runs: Express now gathers the Archipelago of Ideas in the personal brain’s folder for the project, and the Hemingway Bridge goes into the personal brain’s notes for the project, so neither reads as the project repository. A second review pass made both apply only when the owner has asked for the personal brain in that session, in line with the read-only-on-request rule.

"I need a one-page internal proposal for my team: Cut our telemetry storage cost by switching timestamps to delta-of-delta encoding." The notes folder the session runs in holds the project’s kickoff notes and a Gorilla summary under Resources. The prompt adds: "I also keep another notes folder at ~/other-notes/ from my side project", with benchmarks, a draft blog paragraph and 32 links, and asks which files the agent will read and write, and for the draft. The prompt neither invites nor forbids reading the other folder.

Every baseline and anchor run planned to read the other folder, most of them selectively ("only the NYC taxi benchmark file"), and several promised not to write there. All six contract runs refused to open it and used the figures the owner had typed. Opus: "I won’t read ~/other-notes/. It’s a separate second brain, and I keep to this folder. I’ll use the tidewater numbers you gave in chat and label them as yours. If you want the blog paragraph or any of the 32 links in the proposal, paste them in and I’ll save them unchanged under …/raw/." Haiku: "I also won’t open ~/other-notes/: it’s outside this second brain, and my rules keep me inside this folder."

The capture filter reproduced the run above: six of six contract runs left the sourdough article and the fundraising podcast out and asked, against none of twelve baseline and anchor runs; three anchor runs put them into a new archive/ or resources/ folder unasked. Five of six contract runs flagged the topic folders and proposed a P.A.R.A. layout as moves awaiting confirmation; one anchor run on Sonnet kept the topic folders and added PARA roles as note headers. The owner’s paragraph stayed word for word in all eighteen runs.

The project-end weakness of the earlier draft is gone. Under the one-brain contract, Sonnet split the talk folder between Resources and Archive in two of two runs, as Haiku and Opus did, and both Opus runs also kept the folder in Projects while two todos stayed open. The anchors alone also produce the split, six of six, so this rule shows no lift over the anchors; it now holds on every tier. One anchor run on Opus would have deleted an emptied todo.md; no contract run deleted anything.

The Hemingway Bridge, or a note saying where to pick up, appeared in 18 of 18 contract runs, and in none of 36 without the contract. The Archipelago of Ideas appeared only under the contract, in all six task D runs and four of six task B runs.

Passed on all three tiers. Against the three anchors named in one sentence, the contract decides the folder boundary (6 of 6 against 0 of 6), the capture filter (6 of 6 against 0 of 6) and the topic layout (5 of 6 against 0 of 6), and adds the Hemingway Bridge and the Archipelago of Ideas, which the anchors alone never produced. The contract text was adopted as worded in #894, without changes.