Rendered from docs/evaluations/n8n-worked-example.md in the Headwater
corpus. Every document on this half of the site is typed by the taxonomy
the descriptor names: corpus.json.
n8n as a worked instance — a corpus with no docs root
Evidence for criterion 4 of the canonical taxonomy library, which asks each entry to carry a real or realistic corpus that the entry types. The design-spec entry carried one realistic corpus for an invented system. This is its first real one.
It is also the first corpus in this library with no collected documentation root. #488, #489 and #491 each propose a project that keeps its governed writing under one tree. n8n keeps it beside the code it governs. One document per package, across a monorepo of 27,688 files. #492 asks whether a taxonomy shaped around one collected root still resolves against that shape.
The short answer is that it does not, and the correction it needs is one declaration. The longer answer is below, with the run that produced it.
This document reports two of the three kinds that #492 names. The architecture documents are #492 itself and the review rules are #508, and the two halves are marked below. The third is #509, and it is reported here too. The reason for the original split was that the third had no admitted kind to be typed against. The diataxis-site entry supplied one, and the skills corpus is typed below.
The first half of this document reads packages/, and the second half reads .agents/review-rules/. They are two corpora at one pin, because corpus.root is a single scalar and one root cannot reach both trees.
The pin, and the fork
| Fork | https://github.com/headwater-ai/n8n, created 2026-09-01, default branch only |
| Upstream | https://github.com/n8n-io/n8n |
| Branch | master |
| Commit | b0550cb3cb4d1752546a69056c55eccfb9111a12, committed 2026-09-01T16:27:59Z |
n8n's Sustainable Use License states that content of branches other than master is not licensed. Every citation here names master and that commit.
The fork carries no commit of this project. It is a mirror, and its one job is to keep the pin readable after the upstream moves.
The pin is not ordinary caution. Between a survey on 2026-08-31 and an adjudication on 2026-09-01, n8n's .agents/review-rules/ corpus gained a whole category of five governing documents. A report over a repository this active goes stale while it is written, and a pinned one does not.
What was typed, and as what
Four documents, all typed as kinds.design_spec of the design-spec entry:
packages/@n8n/instance-ai/docs/architecture.mdpackages/@n8n/expression-runtime/ARCHITECTURE.mdpackages/@n8n/instance-ai/evaluations/ARCHITECTURE.mdpackages/@n8n/local-gateway/docs/ARCHITECTURE_CONNECTION_VS_SETTINGS.md
The kind is right on its two declared grounds. The purpose of design_spec is behavior, which is what an architecture document serves. The tradition the entry models is the software design document, which is what each of these is.
Typing means one front-matter block on a copy, and nothing else. No word, no heading, no link, no spelling, no contraction and no line break of any body was changed. The fixture directory holds the copies, a verbatim copy of n8n's license, and a modification notice that the license requires. Byte identity was verified mechanically before this document was written. Each copy had its front-matter block stripped and the remainder diffed against git show <pin>:<path>. All four diffs were empty.
None of the four is under n8n's Enterprise License. That license covers a file with .ee. in its name or a directory with .ee in its name. The tree at the pin holds 196 paths of the first form and 1,064 of the second. None is under any of the three packages this evaluation read.
The sequence friction, which is a finding
kinds.design_spec requires the facets doc_type and sequence. sequence is an integer, and the entry's shelf layout is {sequence:02d}-{slug}.md. It is the numbering of a specification series, where RFC 2119 comes after RFC 2118.
n8n's architecture documents carry no sequence number, because a package is not a position in a series. There is no number to write. No value was invented to make the run pass, and the run therefore reports facet.required.missing on all four documents.
The entry assumed that a design specification belongs to a numbered series. A monorepo that keeps one architecture document per package is a corpus where the assumption is simply false. That is the finding, and criterion 4 exists to produce it.
What the run reported
The fixture README carries the run in full, with the assembly that reproduces it. The summary:
4 files under the corpus root, 4 typed, 0 excluded, 4 checked, 13 findings, all 13 of them errors. headwater check --strict exits 1. This page does not give the count of check instances, because each rule that a release adds changes it. The fixture README states that count, and the n8n fixture job compares it with each run. The first and third read 42 and 38 until HW-DR-0067. The 38 were the vendored taxonomy package, which sat inside this corpus root and had to be excluded from it.
Three findings per document, and the same three on each one.
| Document | Rule | Severity |
|---|---|---|
| all four | facet.required.missing (OB-FACET-1), sequence not declared |
error |
| all four | facet.required.missing (OB-FACET-1), title not declared |
error |
| all four | identifier.unusable (OB-ID-2), a typed design_spec that declares no identifier |
error |
The run reported 8 findings until headwater/standard 4.0.0, which requires title on design_spec. None of these four upstream files declares one. This repository will not write four titles into somebody else's documents to make the number go down.
identifier.unusable is a third finding about the entry rather than about n8n. The design-spec entry declares five kinds and no identifier scheme for any of them. This repository mints spec_id in its own overlay, and an overlay is not an admitted library entry. So a corpus that takes the entry alone gets documents that can stand at neither end of a relation. The diataxis fixture reported the same rule six times for the same reason.
Not one finding is about n8n's writing
The base package declares language: default: {tag: en-US, controlled: none}. So no admitted entry of this library holds prose to a controlled language, a source form, a sentence length or a retired term. voice.forbidden_construction ran four instances and reported nothing. section.required.missing never instantiated, because design_spec declares no required sections.
The canonical library, pointed at four real governing documents, says nothing at all about the prose in them. That is worth stating before any count, because a large finding total over somebody else's corpus is easy to mistake for a result.
The broken link that no rule reported until 4.3.0
packages/@n8n/expression-runtime/ARCHITECTURE.md:426 writes the link [n8n workflow package] with the target ../workflow/. From that directory the target is packages/@n8n/workflow/, and the tree at the pin holds no such path. The package it means is packages/workflow/, which holds 244 files. This was checked against the whole tree at the pin and not against the four-file slice.
headwater check sees it and reports it in the graph section as 1 prose links that did not resolve, with the file, the line and the reason. Until headwater/standard 4.3.0 no rule turned it into a finding. From 4.3.0 link.path.unresolved reports it as an error under OB-LINK-2. The near rule is link.fragment.unresolved, which reads a fragment inside a document rather than a path. It ran four instances here and reported nothing, then and now.
So the one defect here that a reader would call a defect was a fact and not a finding at 4.0.0. It is now both. That move is the single change in this document's result. Everything the evaluation concluded from the silence is a claim about 4.0.0 rather than about the library today. Neither of n8n's own two tools reports it.
The scattered-corpus question
This is the third Done-when bullet of #492, and it is the reason the issue exists.
First, a correction to how the issue states it. n8n does have a root docs/ directory, and it holds 272 files at the pin. 271 of them sit under docs/generated/ and are tbls output produced from the database migrations. The 272nd is docs/db.md. It is an authored page, and its subject is how that output is generated and where to read it. No architecture document, no review rule and no skill is under docs/. The issue's claim is right in substance and wrong as literally written. The accurate statement is narrower. n8n has a docs/ directory that carries generated output and one index page for it. Its governed prose lives elsewhere, scattered.
The design-spec entry as it ships types none of these documents. Its one shelf is spec_series at docs/spec/**, and not one file of this corpus is under docs/. Run the same assembly with that shelf alone and the run reports 4 files, 4 untyped, 0 excluded, 0 checked, 1 finding. The first and third read 42 and 38 until HW-DR-0067 moved the vendored package out of this corpus root. headwater check --strict exits 1, and the one finding is a broken prose link rather than anything about the four documents going unread.
That arm reported 0 findings and exited 0 until headwater/standard 4.3.0. A green strict run over four real governing documents that no rule read is what the census crate calls systematically green. That was the result this arm recorded. At 4.3.0 the one finding is link.path.unresolved. That rule reads the corpus rather than the typed set, and it catches a broken prose link in one of the four. Nothing about the four going unread changed. The engine is honest about that in the census, which carries four rows saying no shelf pattern claims this path. No finding anywhere says the corpus went unchecked.
What shape the shelf declaration had to take
Three properties, and each one was measured rather than assumed.
The corpus root is packages and not docs. That forced an exclusion until HW-DR-0067, because a taxonomy package was found under packages/ and n8n's prose is under packages/ too. A corpus root that reached the second reached the first. The three earlier fixtures of this library never met it, because each of them roots its corpus at docs. This corpus is the measurement that priced the old root. The move took the exclusion with it: the package now lands under .headwater/packages/, which no corpus root an adopter can name reaches.
The path pattern has to be packages/**, which fixes one segment out of 27,688 files. The four documents share no directory. What they share is a filename, and the pattern language reads a path rather than a name. packages/**/ARCHITECTURE.md is legal and it claims 2 of the 4, because the other two are named architecture.md and ARCHITECTURE_CONNECTION_VS_SETTINGS.md. That run reports 2 typed and 2 untyped, and the two documents it misses report nothing at all.
The shelf has to be heterogeneous with one admitted kind, and that reads as a contradiction. kinds.design_spec requires doc_type, because the entry's own shelf is heterogeneous and needs a discriminator. A homogeneous shelf refuses a document that restates the kind its placement already states. So a design_spec on a homogeneous shelf reports an error whichever way the front matter is written:
| The shelf, and the front matter | Findings |
|---|---|
homogeneous: true, with doc_type: design_spec |
16, adding shelf.placement_is_primary on all four |
homogeneous: true, with doc_type removed |
16, adding facet.required.missing for doc_type on all four |
homogeneous: false, discriminator: doc_type, kinds: [design_spec] |
12 |
The third row is what the fixture commits. A one-kind heterogeneous shelf makes the discriminator carry no information at all. Every document writes the one value the shelf admits, and the key exists to satisfy a facet requirement rather than to tell two kinds apart.
The answer to the bullet. A shelf declared as a path pattern does resolve against a scattered corpus, and it costs three things that no collected corpus pays. An exclusion the collected shape never needed. A pattern so broad that it selects a whole source tree. A heterogeneous shelf with one kind on it. None of the three is a defect of the engine. All three are the shelf model meeting a corpus whose placement carries no information about kind.
The count
This is the fifth Done-when bullet, scoped to this one kind. It exists so that nobody has to take on faith that a finding is worth anything. A finding that a project's own tooling already catches is not evidence that Headwater helps.
What prettier catches: zero, and not for the reason the issue gives
The issue states that lefthook.yml runs prettier --write on every staged .md file, so Markdown formatting is already enforced. The first half is true and the conclusion is false.
lefthook.yml does carry a prettier_check command with the glob packages/**/*.{vue,yml,md,css,scss}, which matches all four of these paths. But .prettierignore at the repository root carries the line **/*.md, under the comment # Ignored for now. Prettier honors its ignore file even for a path named on the command line, so the hook formats no Markdown at all.
That was verified rather than reasoned. A scratch directory with a .prettierignore of **/*.md and one badly formatted Markdown file: prettier --check reports All matched files use Prettier code style and exits 0. Remove the ignore file and change nothing else, and the same command reports a style issue on the same file.
So prettier catches 0 of the 13 findings, and 0 of anything else in these four files.
The counterfactual is worth one line, because it bounds the overlap. Run prettier over the four files with the ignore lifted and it rewrites three of them. Every change is blank lines around fenced blocks, table column alignment, and indentation inside a fenced block. None of it touches a heading, a facet, a link, a spelling or a sentence. Even with the ignore lifted the overlap would be zero.
What cubic catches: zero, and each rule was read
cubic.yaml declares five custom_rules, and its own trailing comment says all five slots are used. Each one is scoped by an include list, and each list was read against these four paths.
| Rule | include patterns |
Matches any of the four |
|---|---|---|
| Security | packages/cli/**, packages/@n8n/db/**, packages/core/**, packages/workflow/**, packages/nodes-base/**, packages/@n8n/nodes-langchain/** |
no |
| Backend | the same six | no |
| DB migrations | packages/@n8n/db/src/migrations/**, packages/cli/test/migration/** |
no |
| Frontend | packages/frontend/** |
no |
| QA & DX | .github/**, docker/**, scripts/**, patches/**, packages/testing/**, four packages/@n8n/*-config/** entries, and eight file-level patterns |
no |
Not one pattern reaches packages/@n8n/instance-ai/, packages/@n8n/expression-runtime/ or packages/@n8n/local-gateway/. Two further facts point the same way. Cubic reviews only the lines a pull request adds or modifies, and its custom_instructions block states the bar in terms of code rather than prose.
So cubic catches 0 of the 12.
The count itself
| Under the entry alone | Under the entry plus this repository's house regime | |
|---|---|---|
| Findings raised | 12 | 166 |
Already caught by prettier |
0 | 0 |
Already caught by cubic |
0 | 0 |
| Caught by neither | 12 | 166 |
Of those, written by headwater check --fix |
0 | 1 |
| Of those, needing a human rewrite or a declaration change | 12 | 165 |
The second column is a probe and not a declaration, on the precedent the brd-prd fixtures set. It binds regimes.language.ste_house from this repository's own overlay onto kinds.design_spec, changes nothing else, and touches no document. The fixture README names the rules it fires. The counts in the second column and in the paragraphs under it are one past reading, and not a current claim. The probe reads this repository's own house regime. Each edit to that regime changes these counts, and no job holds them. A run today gives other totals.
Three things about that second column matter more than its total.
136 of the 154 extra findings are hard wraps, and they come from two documents rather than four. One carries 108 and one carries 27. The other two carry one between them, because they are already written one line per paragraph. A house rule that reads as a verdict on a project is a verdict on one editor setting.
There is no British spelling anywhere in the four documents. The rule that catches one ran on every document and found nothing. That contradicts what this work expected to find, and it is worth recording as a correction rather than quietly dropping.
Exactly one finding of the 166 is mechanically fixable. headwater check --fix writes a British spelling, a contraction whose expansion is one word, a retired term that names a replacement, and a missing reciprocal link. One contraction, doesn't, is the whole of the intersection, and the report marks it as the only fix (mechanical) line. The 136 hard wraps are mechanical to a reader and carry no patch, and engine/crates/check/src/source_form.rs states why in its own comment. headwater check --fix was never run over this corpus.
What the count says. Under the canonical library alone the honest total is 12. Every one of the 12 is about a declaration rather than about n8n, and every one is invisible to both of n8n's tools. Under a house language regime the total is 166, of which one is mechanically fixable and 136 are one editor's line-wrapping habit. Neither number is a claim that Headwater found 12 or 166 problems in n8n. The 12 are what a taxonomy learns about itself by meeting a corpus it did not author.
The review rules, a second corpus of the same repository
#508 types the second of the three kinds that #492 names. It reads the rules that n8n's AI code reviewer loads on every pull request. It also asks a question that no check answers: whether the three-level model n8n publishes about those rules holds.
The fork is the same fork and the pin is the same commit. The corpus is not the same corpus. corpus.root is a single scalar string. n8n keeps its architecture prose under packages/ and its review rules under .agents/, and one root cannot reach both. So the work built a second fixture beside the first, under the entry it evidences: docs/taxonomies/standards-spec/fixtures/n8n/.
This is where the headline claim of #492 narrows usefully. n8n's architecture documents are scattered, one per package, across a source tree of 27,688 files. Its review rules are collected: 23 files under one directory, six subdirectories deep at most. One repository is a scattered corpus for one of its kinds and a collected corpus for another. The shelf model met both. The first cost a broad pattern and a forced exclusion. The second cost one ordinary path pattern.
What was typed, and as what
Seven documents, all typed as kinds.standard of the standards-spec entry. Six are rule files across five of the six directories. The seventh is the README.md that files them, and it is here because a sweep finding may only name a typed document.
The kind is right on its declared ground, and the granularity mismatch is the finding. The purpose of standard is constraint, which reads "state what must hold across the documents and code it governs, and what it rules out". A rule file decides what a reviewer comments on, which is what a constraint is. #492 guessed that this kind was requirement shaped. That guess points at the brd-prd entry, which declares kinds.brd and kinds.prd and nothing granular. A granular kinds.requirement exists in this repository's own overlay and nowhere else. An overlay is not an admitted library entry, so it cannot satisfy criterion 4. No admitted entry of this library declares a kind for one rule. kinds.standard is the closest admitted target. Every one of the 21 findings below comes from the distance between a whole standard and one rule of one.
Typing means one front-matter block on a copy, and nothing else. Byte identity was verified twice. Each copy had its front-matter block stripped and the remainder diffed against the file at the pin, which gave seven empty diffs. Each file at the pin was then compared by git hash-object against the blob the fork reports at that commit, which gave seven matches.
The fixture directory carries its own verbatim copy of n8n's license and its own modification notice. Neither is inherited from the design-spec fixture. A second set of copied files is a second receipt of part of the software. A modification notice is a claim about which files were modified, and these are different files. None of the seven is under n8n's Enterprise License. The tree at the pin holds 91 paths under .agents/, and zero of them match .ee in any form. The check was run rather than assumed.
What the run reported
The fixture README carries the run in full. The summary:
7 files under the corpus root, 7 typed, 0 excluded, 7 checked, 21 findings, all 21 of them errors. headwater check --strict exits 1. The fixture README states the count of check instances, and the n8n fixture job compares it with each run. The graph reads 7 nodes and 0 declared edge halves.
Three findings per document, and the same three on every one. section.required.missing, an error, under OB-SECT-1, once each for Scope, Requirements and Conformance.
This is the first time in the n8n work that the rule instantiated at all. The design-spec run recorded that it never did, because design_spec declares no required sections. Here it produces the whole run.
The finding is sharper than a heading count. Not one heading at any level of any of the 23 upstream files carries the word Scope, Requirements or Conformance. The corpus is not missing the content. Every one of the 22 rule files opens with a line Applies to: …, which is the scope written as a labeled paragraph. 16 of them use the imperative Flag somewhere to state what to flag, and only six write a literal Flag: block. The required section is present in substance in the whole corpus and absent in form in the whole of it. The contract is lexical over headings, and the tradition writes a labeled paragraph. Conformance is the honest third. No rule file says how conformance is judged, because the AI reviewer judges it, so that section is genuinely absent.
One expectation was checked and is not due. kinds.standard expects a regulates edge to a functional_spec within 180 days of state_entered, at severity warn. relation.participation.overdue instantiated seven times and reported nothing, because status_since and the injected clock are both the pin date. The window opens on 2026-09-01 and its deadline falls on 2027-02-28. The rule fires only once the clock passes that date, so the seven warnings appear from 2027-03-01. That is a fact about the clock and not about n8n.
Whether the three-level reach model holds
This is the fourth Done-when bullet of #508, and it is a headwater sweep job rather than a check. It is the first coherence sweep run and recorded anywhere in this repository. There was no committed briefing, no return file and no report before this one. All three artifacts are committed under sweep/ beside the corpus. sweep plan is byte-reproducible, and the return file is the only thing that lets a later reader reach the same verdict.
headwater sweep report confirmed four things and no more. The return file names the lock this tree carries. Every path it names is a typed row of the census. Every quotation is really in the document it is attributed to. No proposed edge restates one the graph already declares. 3 findings carried, 0 refused, of 3 the file held. It confirms nothing about whether a reading is right.
The answer is that the model holds for the two cases the corpus declares, and a third case broke it that nobody noticed.
The two declared exceptions are exceptions. testing/coverage.md is one file listed in two agents' file_paths, Backend and Frontend, exactly as the README declares. Security and QA & DX not linking it is stated with a reason in the same paragraph. Neither is a contradiction and the sweep names neither.
The third case is a genuine defect. The README names two agents that deliberately do not link testing/. Read against cubic.yaml, three do not. The DB migrations agent's file_paths holds only its own five files. The README's own Layout table is already consistent with three and maps testing/ to Backend and Frontend. So the table was updated when the db-migrations/ category arrived and the prose two paragraphs below it was not. The reason that prose gives, coverage nagging on a credential fix or a Dockerfile, does not describe a migration. The defect is causally tied to the drift the pin exists to capture.
A second finding is a real reading and a contestable one. The README rules that a policy identical across domains is one shared file and never a copy per directory. testing/coverage.md holds the shared coverage policy. db-migrations/conventions-and-tests.md holds a second coverage policy under its own ## Tests heading. A defender would call this specialization rather than duplication. The README's own test is "word-for-word the same", and a migration test is not word-for-word a service-method test. Both readings are honest. The sampler adjudicates neither, and that is what a sampler is for.
A third finding is about the corpus's organizing term. The README opens its instructions to a rule author with "Pick the level of reach first." The term is defined nowhere in the slice. Its only definition is a header comment in cubic.yaml.
What the sweep could not reach, which is three results
cubic.yaml can never be a finding path. A sweep finding may only name a typed row of the census, and cubic.yaml is a configuration file that no taxonomy here types. The document that asserts the three-level model cannot appear in the sweep that checks it. Every quotation from it lives in a message, where nothing verifies it. The sampler can confirm a contradiction between two governed documents and cannot reach the configuration file that governs them.
conflicts_with cannot join two standards, and the intake carried the proposal anyway. The base package declares it from: [decision] to: [decision]. The second finding proposes an edge between two standards on purpose, to find out. The intake's four confirmations do not include endpoint-kind validation, so the proposal passed and the report printed the front matter to write. Writing that front matter and re-running headwater check reports 2 relation.endpoint.not_permitted errors, one for each end. A sweep proposes an edge that the gate then refuses, and nothing between the two says so.
The front matter the report prints names the wrong document. The proposal names N8N-STD-coverage as its source. The line beneath it points the reader at the first path in the finding's document list, which is the README. A reader who follows the instruction declares an edge from N8N-STD-review-rules, which is not the edge the sweep proposed. Both arms were run and both report the same 23 findings, so the gate catches the substitution for the wrong reason and never names it.
One result belongs to the register rather than to the corpus. The standards-spec entry declares obligations.OB-SS-1 with class: coherence. Spec 4 and HW-OBL-0114 both record that this repository's register has no coherence-class obligation, so a sweep here discharges nothing. This corpus gives that class its first member anywhere in the library. It still discharges nothing, because OB-SS-1's disposition is unverifiable and no control names a sweep: mechanism. The gap moves from hypothetical to visible.
The count, extended to a third tool
prettier catches 0, and this slice has two independent reasons where the architecture documents had one. .prettierignore carries **/*.md at the repository root, which is the reason already recorded above. And lefthook.yml's prettier glob is packages/**/*.{vue,yml,md,css,scss}, which never reaches .agents/ at all. Either one alone gives zero.
cubic catches 0, and the reason is worth one sentence on its own. Not one include pattern of the five custom_rules reaches .agents/. The five lists cover packages/, .github/, docker/, scripts/, patches/ and eight file-level entries. cubic reviews its own rule files under no rule at all. lefthook.yml carries a .agents/skills/** job and no .agents/review-rules/** job, which points the same way.
A third tool reads exactly these files, and it catches 0 as well. pnpm check:cubic-config runs .github/scripts/quality/check-cubic-config.mjs, and CI runs it on every pull request. It validates cubic.yaml against a vendored copy of cubic's schema. It then fails on a missing linked path, an agent over the 10,000-character ceiling, or a rule file that no agent links. Its constants name .agents/review-rules directly. It reads no prose. It counts characters and resolves paths, so the overlap with 21 findings about headings is zero. The zero is measured rather than assumed.
One detail of that third tool is a result in itself. Its file list excludes README.md by name, because the README is not a rule and no agent links it. The one document the sweep found three findings in is the one document n8n's own checker declines to open.
| Under the entry alone | Under the entry plus this repository's house regime | |
|---|---|---|
| Findings raised | 21 | 143 |
Already caught by prettier |
0 | 0 |
Already caught by cubic |
0 | 0 |
Already caught by check:cubic-config |
0 | 0 |
| Caught by none of the three | 21 | 143 |
Of those, written by headwater check --fix |
0 | 6 |
| Of those, needing a human rewrite or a declaration change | 21 | 137 |
The counts in the second column and in the paragraphs under it are one past reading, for the reason given under the design-spec table. A run today gives other totals.
Two things in that second column correct what the design-spec run recorded.
The hard wraps are spread across all seven documents, and there they were not. 87 of the 122 extra findings are language.source_form.not_met, and every one of the seven documents carries between 2 and 42. In the architecture corpus 136 of 136 came from two documents of four, and the other two were already written one line per paragraph. Same repository, same pin, one editorial convention per corpus.
The British-spelling claim was wrong in both directions. The design-spec run recorded that there is no British spelling anywhere in its four documents. There are four in this corpus: behaviour, defence, colours and denormalised. Three of them are in the typed set. The rule reports one. The engine's spelling table is closed at 24 words and holds behaviour and not defence, colour or denormalise. So a corpus-level count and a rule-level count differ here by a factor of three, and only the second is what a run measures.
Six findings of 143 are mechanically fixable, against one of 166 before. Five contractions and one spelling are the whole intersection, and the report marks each with fix (mechanical). The 87 hard wraps are mechanical to a reader and carry no patch. headwater check --fix was never run over this corpus.
What this adds to criterion 4
This is the seventh Done-when bullet of #492 and the sixth of #508. Three entries of the library now carry this repository. The design-spec entry carries the invented Beacon corpus and n8n's architecture documents, and its doctrine cites both. The standards-spec entry carries the invented Beacon ladder and n8n's review rules, and its doctrine gained a Worked instances section that cites both. The diataxis-site entry carries the invented Beacon-style shape fixtures and n8n's skills, and its doctrine gained the same section.
One note, which is not fixed here. This document is a kinds.evaluation, and evaluation moved out of design-spec into the evidence-and-obligation entry with HW-DR-0044. The admitted-entry table of docs/taxonomies/README.md gives that entry no row. The page says why under The bundle the table does not admit, which reads every admission criterion against it. Both design-spec and decision-record declare requires: [evidence-and-obligation], so a composer who takes either of them takes a refused bundle with it. Whether an admitted entry may require a bundle that admission refuses is an open ruling, and HW-OBL-0185 holds it.
What .agents/skills/ states, and what the tree holds
The third corpus that #492 names is .agents/skills/, and this section reports one finding in it without typing any of it. The next section types the corpus. The finding below needs no kind either way, because it is a claim the corpus makes about itself.
At the pin, .agents/skills/spec-driven-development/SKILL.md:8 states this:
Specs live in
.agents/specs/. They are the source of truth for architectural decisions, API contracts, and implementation scope.
The tree at the pin holds no such path. .agents/ holds two entries, review-rules and skills, and 58 files below them. The count of paths under .agents/specs/ is zero. Five lines name that path and all five are in this one file: the front-matter description, and lines 8, 20, 23 and 66. No other file under .agents/ names it.
The directory is absent, and it is not ignored. .gitignore at the pin runs to 96 lines and names .agents/specs/ on none of them. Line 84 excludes .claude/specs/, which is a different path under a different directory. A reader who expects an ignore rule to explain the absence will not find one.
The skill anticipates a missing spec for one feature, not a missing root. Step 3 of Before Starting Work reads "If no spec exists and the task is non-trivial (new module, new API, architectural change), ask the user whether to create one first." That step handles a feature with no spec. It does not handle a root that holds no specs at all, which is what this tree has. Step 1 gives a literal command, ls .agents/specs/, and at this pin that command fails for every feature and not for some.
What this taxonomy would report, and what it does not
Two relations in the resolved taxonomy have the shape this needs. governs runs from a governed_document to a code_path, and traces_to runs from a governed_document to a governed_document or a code_path. Either one could carry the claim that a stated practice makes, and a check could then ask whether the target exists. Neither relation is required on any kind, and no rule asks a document to declare one.
So this taxonomy has the vocabulary for the claim and no obligation to state it. This document already records the near case. The unresolved link at packages/@n8n/expression-runtime/ARCHITECTURE.md:426 reaches headwater check as a fact in the graph section, and from 4.3.0 as a finding too. The .agents/specs/ claim does not reach the engine at all, and typing the corpus does not change that. The run in the next section types every file of .agents/skills/ and reports 136 findings, and not one of them is this claim. A stated practice that no artifact backs is invisible to a check whether or not the document carrying it has a kind.
The owner ruled that this gap is a sweep pattern and not a check rule. The ruling on #496 is dated 2026-09-22. Its reason is that to tell an empty directory from one that nobody meant to create is a judgment. The headwater-sweep skill carries the pattern as a hand pass, with the command that collects the candidates. Over the fixture copy of this corpus, docs/taxonomies/diataxis-site/fixtures/n8n/corpus/, the command prints four lines. Three are spec-driven-development/SKILL.md at lines 23, 35 and 81, which are lines 8, 20 and 66 at the pin. The fourth is .agents/plans/ in create-pr/SKILL.md, and that skill reads the directory only "when present". So the pattern finds the claim above, and a reader then judges the one line that is not a claim.
The pattern, run once over this repository. The run was on 2026-09-27, at 9851bf6e on main, with the command in the skill. It read 553 Markdown files, which are every tracked *.md file outside docs/reviews/ and outside any fixtures/ directory. It printed 92 mentions of 45 distinct paths. Each mention was read in its line and judged, and the judgment of a path held for every mention of it:
| judgment | mentions | the paths, with the mentions of each |
|---|---|---|
| a build output | 13 | engine/target 8, engine/target/ 2, engine/target/release/headwater 3 |
| a checkout or a cache that a session or a verb writes | 15 | .claude/worktrees/ 6, .headwater/cache/checks 4, .headwater/cache/ 2, .headwater/cache/.gitignore 1, .headwater/cache/embeddings/ 1, .headwater/models 1 |
| a snapshot, an import or a site directory that a run writes | 12 | .headwater/observations.yml 5, .headwater/site-build 3, .headwater/site-deploy 3, .headwater/imports/ 1 |
| a path in Sysl | 12 | docs/ideas/root.md 7, docs/ideas/ 5 |
| a path in n8n | 7 | docs/generated/, docs/db.md, .claude/specs/, .claude/plugins/n8n/skills/, .github/pull_request_template.md, .github/pull_request_title_conventions.md, .github/scripts/quality/check-cubic-config.mjs |
| a path in an adopter's repository | 3 | .github/workflows/headwater-check.yml, docs/decisions/0001-store-attempts-in-postgres.md, docs/decisions/0002-deliver-at-least-once.md |
| a path in a fixture corpus, or in another harness | 4 | docs/specifications/beacon-overview.md 2, docs/evaluations/queue-durability.md 1, .github/skills/ 1 |
| an example, or an alternative that a record refuses | 12 | docs/drafts/ 2, .claude/hooks-disabled/write.sh 3, .claude/hooks-disabled 1, docs/corpus.json 1, mkdocs/build.sh 1, site/docs 1, docs/INDEX.md 1, .headwater/packages/acme-my-taxonomy 1, engine/crates/check/src/runner.rs 1 |
| a retired or past path, and the text says that it is gone | 10 | tools/ste-lint.py 6, docs/decisions/0049 1, docs/obligations/a-governs-edge-reaches-the-path-it-names-and-nothing-under-it.md 1, docs/spec/16-a-second-part-at-sixteen.md 1, .claude/notes/cost-attribution-handoff.md 1 |
| a practice that exists, at a path that moved | 2 | engine/crates/census/src/pattern.rs 1, tools/id-store-fixtures.sh 1 |
| a claim about this tree that nothing backs | 2 | docs/RFCS/ 2 |
One claim in this repository has no path behind it, in two copies. The candidate backlog of the library index disqualifies the proposal-process tradition because "a docs/RFCS/ corpus already in this repository" reads as a decision record. This tree has never held that path. #7, where the sentence came from, names the corpus of #489, which is in another repository. The package copy of the same page carries the same sentence. Two further lines cite a path that moved while the practice stayed. HW-OBL-0060 cites the pattern order in the census crate, which moved to engine/crates/meta/src/pattern.rs in 88fd3ef4. HW-OBL-0201 cites a fixture suite that is now tools/repo/id-store-fixtures.sh. The three went to the intake file of the run that recorded this, and this change repairs none of them.
What the command does not see. It reads a path only inside backticks, only with a slash, and only where the tree tracks the first segment. So it misses a command in one code span, such as ls with its directory, and a path with no directory at all. It asks git and not the disk. So it counts a directory that a build writes as absent, and its result does not depend on what the host has built. This section adds mentions of absent paths itself, so a run at a later commit prints more than 92 lines.
What was enumerated here, and what was not
Every number below came from the tree at the pin, measured on 2026-09-08. The last_verified date in the front matter covers the rest of this document and not this section.
| fact | number | of what |
|---|---|---|
entries directly under .agents/ |
2, and neither is specs |
the tracked tree at the pin |
blobs under .agents/ |
58 | 35 under skills, 23 under review-rules |
skill directories under .agents/skills/ |
22 | 34 files inside them, plus a top-level AGENTS.md |
lines that name .agents/specs/ |
5, in 1 file | of the 58 files |
| lines that claim a source of truth | 15, in 10 files | of the 58 files |
| those lines that name a repository-root path | 5, in 4 files | of the 15 lines |
| those root paths that the tree does not hold | 1, .agents/specs/ |
of the 5 paths |
The source-of-truth count ignores letter case. A case-sensitive match on the lower-case phrase gives 13 lines in 9 files, because db-migrations/SKILL.md opens two of them with a capital.
The last two rows were enumerated and not sampled. Four files name a repository-root path as a source of truth. One of them is spec-driven-development/SKILL.md, and the other three each resolve at the pin:
db-migrations/SKILL.mdnamespackages/@n8n/db/src/migrations/migration-types.tsandpackages/@n8n/db/src/migrations/dsl/column.ts.community-pr-readiness-check/reference/checks.mdnames.github/pull_request_template.md.public-api/SKILL.mdnamespackages/@n8n/decorators/src/controller/, which holds 35 files.
The other six of the ten files point at a symbol, a service, or a file inside a package. None of those six was resolved, and this section claims nothing about them.
To take the finding again, run this against the pin:
gh api "repos/headwater-ai/n8n/git/trees/b0550cb3cb4d1752546a69056c55eccfb9111a12?recursive=1" \
--jq '.tree[].path' | grep -c '^\.agents/specs'
It prints 0. Any other number makes this section false.
The skills corpus, typed
The section above reports one finding in .agents/skills/ and types none of it. This section types all of it. The premise both #492 and #509 rest on is dead: the diataxis-site entry landed in 46b86799 and declares kinds.how_to, a concrete kind under purposes.procedure. That kind is what the third corpus waited for. The vendored slice is docs/taxonomies/diataxis-site/fixtures/n8n/, at the pin the other two corpora use, and its README carries the assembly commands that produced every number below.
What the pin holds, re-measured. .agents/skills/ holds 22 skill directories, 34 files inside them, and one top-level AGENTS.md of 1,952 bytes. That is 35 files and 425,316 bytes of upstream body. #492 and #509 each write "22 skill directories and 35 files". Neither one names AGENTS.md, so a reader of either issue puts all 35 files inside the 22 directories. The files that are not a SKILL.md number 12 and not the ten that #509 enumerates.
Every file is typed, and that is the result this corpus was vendored for. One homogeneous shelf at .agents/skills/** carries how_to. The run reports 35 files under the corpus root, 35 typed, 0 excluded, 35 checked, 136 findings. The fixture README states the count of check instances, and the n8n fixture job compares it with each run. The census reads 35 how_to. The graph reads 35 nodes and 0 declared edge halves. Nothing is untyped, and headwater check --strict exits 1.
| Rule | Count | Severity |
|---|---|---|
section.required.missing |
104 | error |
link.path.unresolved |
14 | error |
voice.forbidden_construction |
18 | warn |
104 of a possible 105, and the missing one is the interesting number. kinds.how_to requires Goal, Steps and Verification. Thirty-five documents times three headings is 105. The run reports Goal 35 times, Steps 34 times and Verification 35 times. One file writes a Steps heading by name, .agents/skills/create-pr/SKILL.md. That is the only heading of the three the corpus supplies anywhere. The review-rules corpus reported this shape from a different kind. A section contract reads headings, and the tradition writes the same content as an imperative paragraph.
Thirteen of the 14 link findings are a property of the corpus root and not of n8n. corpus.root is .agents, and 13 of the targets are relative paths that climb above it, such as ../../../packages/@n8n/instance-ai/evaluations/README.md and ../../../AGENTS.md. Each of the 13 was checked against the pin through the contents API and all 13 exist upstream. A run whose root reached the whole repository would resolve every one of them.
The fourteenth is a real dangling link, and this time a rule reports it. .agents/skills/design-system/rules/web-animation-guidelines.md:84 links PRACTICAL-TIPS.md, and the tree at the pin holds no such path under .agents/skills/design-system/rules/. This is the shape of the .agents/specs/ claim above, written as a link rather than as a sentence. The difference is what a rule can reach. A link resolves or it does not, and a stated practice reaches no rule at all.
The symlink tree, measured. .claude/plugins/n8n/skills/ holds 24 entries at the pin. Twenty-two are symlinks at mode 120000. The set of 22 link names equals the set of 22 shared directory names, with no difference in either direction. The other two entries are a real directory, setup-mcps/, and the SKILL.md inside it. A corpus rooted at .agents sees none of this, because a symlink tree outside the root is outside the census. So the run types 35 documents where a run over both trees would type 36. The one it cannot reach is the only skill that no shared directory carries.
The front matter is merged here and prepended in the other two fixtures. Twenty-two of these 35 files already carry Agent Skills front matter with name and description. A second block above the first one would not be read, because a parser takes the first block. So the Headwater keys go into the block that is already there, under n8n's own keys, and no upstream key is changed or removed. The 13 files with no upstream front matter get a new block above the body. Under both shapes every byte below the front matter is the file at the pin. Byte identity was verified mechanically. The blob hash of each upstream body was recomputed and compared against the tree API's blob at that commit: 35 comparisons and 35 matches.
The summary facet came from the corpus rather than from an author. governed_document requires summary, and 35 hand-written summaries would be 35 sentences of invented prose inside a fixture that exists to be byte-identical. For the 22 files with upstream front matter the summary is n8n's own description line, copied. For the other 13 it is the document's own first heading, copied. The fixture invents no prose, and a weak summary on 13 documents is what that costs.
One number in the review-rules fixture has already gone stale, and it is not a number about n8n. docs/taxonomies/standards-spec/fixtures/n8n/.headwater/taxonomy.yml takes headwater/standard at 4.0.0, and the package in this repository is now 4.2.0. headwater taxonomy resolve refuses that mismatch and prints both versions, so the commands that fixture states no longer run as written. The skills fixture takes 4.2.0 and resolves. No check reads either scalar, because docs/taxonomies/** is outside this repository's corpus.
The count, extended to a third kind
The terms are #492's. #511 recorded 8 findings for the architecture documents and #512 recorded 21 for the review rules. This is the third row, and it takes the third tool #509 names.
| Findings | |
|---|---|
Raised by headwater check over the typed slice |
136 |
Also caught by prettier |
0 |
Also caught by cubic |
0 |
Also caught by sync-agent-skill-links.mjs --check |
0 |
| Caught by none of the three | 136 |
Of those, written by headwater check --fix |
0 |
| Of those, needing a rewrite | 136 |
prettier reaches nothing here, and for two independent reasons. .prettierignore at the pin ignores **/*.md under a comment that reads "Ignored for now", so no markdown file anywhere in the repository reaches prettier. And lefthook.yml scopes its prettier_check command to a glob of packages/**, which .agents/ is not under. Either reason alone gives zero, which is the same answer the architecture corpus got for a different reason.
cubic reaches one of the 35 files and reviews none of them. The Frontend agent in cubic.yaml names .agents/skills/design-system/SKILL.md in its file_paths, so cubic loads that one skill as guidance on every frontend review. Every include pattern of every agent in that file is under packages/. So a change to a file under .agents/skills/ is reviewed by no agent at all. One of the 35 is an input to the tool, and none of the 35 is a subject of it.
sync-agent-skill-links.mjs --check reaches all 35 and reads none of them. lefthook.yml runs it at pre-commit on a glob of .agents/skills/**, so its reach over this slice is total where the other two tools have none. It compares the plugin symlinks against the shared tree and reads no prose. Total reach and zero overlap is the sharpest version of the result #492 asks for.
Nothing here is mechanically fixable, and the engine states that rather than writing nothing quietly. A --fix arm over the assembled root printed no finding of this run carries a patch and changed no file. That is 0 of 136. The review-rules corpus recorded 6 fixable of 143 under this repository's house regime. The difference is the regime rather than the corpus. A contraction and a British spelling carry a patch, and a missing heading, a broken link and a forbidden voice construction do not.
What ages here. Every count above is of the 35 blobs at b0550cb, and a later commit of master is a different measurement. The last_verified date in the front matter of this document covers neither this section nor the one above it. Both were measured at the pin on 2026-09-11 and neither moves when the rest of this document is re-verified.
The three defects, fixed on a fork and sent to nobody
This is the first two Done-when bullets of #495. Every number in this section was measured on 2026-09-08 against n8n-io/n8n@master, which is the live branch and not the pin.
The last_verified date in the front matter covers neither this section nor the one above, and the carve-out is deliberate rather than deferred. A claim taken from a moving branch goes stale when the branch moves. That is why a single date over the whole document would say something false about this section within a week. Moving last_verified forward would also assert a re-verification of the two pinned runs that nobody ran.
Every recorded finding, partitioned
The two runs and the sweep put 38 findings and facts into this document. This is all 38, in three buckets.
| bucket | count | share of 38 |
|---|---|---|
| A genuine defect in n8n's own terms | 3 | 8% |
| An artifact of this library's vocabulary | 33 | 87% |
| Contestable, and this document adjudicates neither | 2 | 5% |
34 of the 38 were reported by one of the two headwater check runs, and 33 of those 34 are in the artifact bucket. 13 came from the design-spec corpus, over 4 typed documents. Those 4 documents are the whole of that root now that the vendored package is not under it. 21 came from the standards-spec corpus, over 7 typed documents of 7 files. All 34 are errors. The 34th is the broken link. It sits in the genuine-defect bucket, and a rule started to report it at headwater/standard 4.3.0. The other 33 are about the distance between an admitted entry and n8n's shape. That distance shows in four places. The first two are a sequence facet that a package has no number for and a title facet that upstream does not write. The last two are an identifier scheme that the entry never declares and three headings that this tradition writes as labeled paragraphs. Not one is about n8n's prose, for the reason recorded above.
The past run under this repository's own house regime, which reported 143 findings, stays out of the denominator on purpose. A house regime is not an admitted library entry. Holding somebody else's corpus to this repository's line breaks and contractions produces nothing that n8n would call a defect.
1 of the 3 genuine defects is reported by a headwater check rule, and 0 of them were when this evaluation was recorded. The broken link reached the 4.0.0 run as a fact in the graph section and never as a finding, and link.path.unresolved reports it from 4.3.0. One reaches no run at all, because no admitted entry declares a procedure-shaped kind and no run reads .agents/skills/. One came out of headwater sweep, which is a sampler and not a check. That inversion was the sharpest result here, and it is one third smaller than it was. The library, pointed at eleven real governing documents, raised 33 errors that say nothing n8n would act on. It stays silent on 2 of the 3 things n8n would fix.
The 2 contestable findings are the coverage-policy duplication reading and the undefined term "level of reach", both recorded above. Neither is fixed and neither goes upstream. They are counted here rather than dropped, because a partition that quietly loses its awkward members is not a partition.
Each defect, re-measured on the live branch
The failure this section exists to prevent is telling n8n about a defect that is not there. The report is pinned and the fork is a mirror, so nothing in this repository notices upstream fixing one. Each block below runs against master and prints the number that would strike the defect. All three were run on 2026-09-08 and all three defects still stand.
Defect 1, a link that resolves to no path. packages/@n8n/expression-runtime/ARCHITECTURE.md:426 writes - [n8n workflow package](../workflow/), which resolves to packages/@n8n/workflow/.
gh api -i "repos/n8n-io/n8n/contents/packages/@n8n/workflow?ref=master" 2>/dev/null | head -1
gh api -i "repos/n8n-io/n8n/contents/packages/workflow?ref=master" 2>/dev/null | head -1
The first prints HTTP/2.0 404 Not Found and the second prints HTTP/2.0 200 OK. A 200 from the first, or a 404 from the second, makes this defect false.
Defect 2, a source-of-truth claim that no directory backs. .agents/skills/spec-driven-development/SKILL.md:8 names .agents/specs/ as the source of truth for architectural decisions, API contracts and implementation scope.
gh api "repos/n8n-io/n8n/contents/.agents?ref=master" --jq '[.[].name] | join(" ")'
It prints review-rules skills. Any output that holds specs makes this defect false.
Defect 3, a count in prose that the configuration contradicts. .agents/review-rules/README.md:25 reads "Security and QA & DX deliberately don't link testing/", which names two agents.
gh api "repos/n8n-io/n8n/contents/cubic.yaml?ref=master" --jq .content | base64 -d | grep -c '^ - name: '
gh api "repos/n8n-io/n8n/contents/cubic.yaml?ref=master" --jq .content | base64 -d | grep -c 'review-rules/testing/coverage.md'
The first prints 5 and the second prints 2, so three agents do not link testing/ where the prose names two. Any pair whose difference is 2 makes this defect false.
The fixes, and the rule that found each one
The fixes stand on the branch fix-agents-doc-defects of the fork, cut from the pin, one commit per defect. Titles follow n8n's pull-request title convention.
Two of the three files carry the same bytes on master today as at the pin, and the third does not. packages/@n8n/expression-runtime/ARCHITECTURE.md is blob d4d9c5b at both, and .agents/skills/spec-driven-development/SKILL.md is 57470f1 at both. .agents/review-rules/README.md moved from 28140c5 to b95b525, in two hunks. The two hunks are the security/ scope row of the Layout table, and a new paragraph under step 3 of Adding a rule. Neither hunk reaches the three lines that the third fix rewrites, which are lines 25 to 27 and identical at both. So a fix cut from the pin still applies to all three, and the reason is a line-level comparison rather than a whole-file one. This was measured on 2026-09-08 and it goes stale the same way every other number in this section does.
| defect | commit | what the fix writes | the rule that reported it |
|---|---|---|---|
| 1 | c7c3613 |
../workflow/ becomes ../../workflow/ |
none: a graph-section fact, and link.fragment.unresolved reads a fragment rather than a path |
| 2 | 667f926 |
the claim becomes conditional, and the skill says what to do when the directory is absent | none: no admitted entry types .agents/skills/, so no run reads the file |
| 3 | 3d450a8 |
the sentence names DB migrations as the third agent | none: headwater sweep carried it, and a sweep is a proposal rather than a verdict |
The rule column reads none three times, and that is the result rather than a gap in the table. Each fix here was found by a person reading, by a graph fact that no rule consumes, or by a sampler. Whether an unresolved prose link should become a finding is one open obligation of this repository. A stated practice that no artifact backs is not a rule, by the owner's ruling on #496. It is a hand pass beside the sweep, and defect 2 is one instance of it.
No fixture copy was touched. Two of the three fixed files are also held here as fixture copies. The typing note above claims byte identity with the pin for those copies. Editing a copy would make that claim false and no check would report it. The fixes live on the fork and nowhere else.
Why nothing was sent, and what would reopen it
The third Done-when bullet of #495 asks for a pull request against n8n-io/n8n. Nothing was posted, and that is a decision rather than an omission. n8n's CONTRIBUTING.md gates an incoming change three ways, and an autonomous run satisfies none of them.
- A bug fix needs a linked issue first. Section 1 states that a bug-fix pull request with no linked issue is returned. Filing that issue is a claim on n8n's tracker, and it belongs to a person.
- A typo-only pull request is rejected. Section 5 asks that a small fix go into a larger related change. Defect 1 on its own is exactly the shape that rule names.
- The description must be the author's own words. Section 4 requires the author to understand every line and to write the description themselves, and it forbids pasted model output. Section 3 closes a pull request with no real description, and section 6 closes one with no tests after 14 days.
So this run prepared everything and posted nothing. The commits are on the fork. The upstream issue text is in .headwater/notes/n8n-upstream/issue.md and the pull-request body is in .headwater/notes/n8n-upstream/pull-request.md. Both are drafts for a person to rewrite in their own words and to send under their own name. They should carry the disclosure that section 4 invites.
What reopens it: the owner of this repository says to post, or n8n drops the linked-issue gate and the own-words gate. The upstream half is #729, which carries the branch, the two drafts and the bar for sending. It stays open on a person, and this document records why rather than leaving a reader to guess.
What this did not cover
Sending any finding of the skills corpus upstream, or fixing one, which is #495. Typing that corpus is no longer outstanding. #509 was blocked on a procedure-shaped kind, the diataxis-site entry declares one, and The skills corpus, typed above is the run.
A census of all 22 review-rule files as typed documents. Seven is the sample, and #508's own scope note sets that bar.
Sending anything upstream to n8n, which is #729. The three genuine defects are fixed on the fork under #495, and nothing was reported to n8n. No issue and no pull request was opened against their repository, and the section above says why. The owner ruled on 2026-09-22 that no pull request goes to n8n from this work. Both issues closed on that ruling on 2026-09-30.
Fixing the 33 findings that the two runs reported. All 33 are artifacts of this library's vocabulary rather than defects in n8n's writing, so there is nothing there for n8n to accept.
A check rule for the gap between a stated practice and an absent artifact. The owner ruled on #496 that the gap is a sweep pattern. The section that reports its instance records the ruling and one run of the pattern over this repository.
A census of every architecture document in the monorepo. Four is the sample, and the fourth Done-when bullet of #492 asks for one document of this kind at minimum.
Reproducing this
The design-spec fixture README carries the five commands for the architecture corpus. The standards-spec fixture README carries six for the review rules, and the sixth is the sweep. sh tools/taxonomy/n8n-fixtures.sh runs all three n8n corpora in CI as a blocking step, by the commands those pages print, and holds the figures those pages state.
No job holds a figure of this document. Every total above is restated here in prose. The job reads the three fixture READMEs rather than this page, so a number here can go stale while CI stays green. The design-spec README says the same thing from the other side, and it also names the probe arms that no job holds. Re-run a figure before you cite it, and correct both copies when one moves.