Rendered from docs/taxonomies/decision-record/doctrine.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

The decision-record taxonomy

One decision, one document, kept forever. A team writes down the choice it made, the forces that produced it, and what the choice costs. It never edits the document afterward. When the choice changes, a new document replaces the old one and says so, and the old one stays where it is.

That is the architecture-decision-record tradition. It is the most widely practiced structured-documentation convention in software teams, and it is the second entry in this library.

This entry is also the remedy that 13 — Open obligations prescribes under Identity below the grain of a document. Two registers here hold their entries as sections of one file, so a decision is an anchor and an obligation is a bold paragraph. Neither is a document, neither reaches the identifier index, and no edge names either. The remedy is to convert each entry into a document, "which is a different tradition with a taxonomy of its own". This is that tradition, and the migration is separate work.

The item that names that remedy has no anchor, which is why the sentence above names it in prose. An entry in either register is a paragraph and never a heading, so a link into one does not resolve. That is the defect stated in the coordinate a reader meets it in.

The prior art

Nygard, 2011. Documenting Architecture Decisions is the post that started it. It supplies the three sections that everything after it copies: Context, Decision, Consequences. It also supplies the two rules that matter more than the sections. One decision per document, and a document is never edited once the decision is taken. A superseding record points back at what it replaced.

MADR. Markdown Any Decision Record generalizes the form past architecture and adds the parts that Nygard left to the author: a title that states the decision, a status, the options that were considered, and the outcome with its justification. It is the form that most tooling now writes.

adr-tools and log4brains. adr-tools is a shell scaffolder. adr new writes a numbered file from a template, and adr new -s 9 writes the successor and edits both records to name each other. log4brains does the same and publishes the log as a site. Both are evidence for one claim of this design: the succession edge is written by a scaffold, and the reciprocal half comes free when a tool writes both ends. Nobody maintains that edge by hand at any scale, and neither tool asks anybody to.

Kruchten. The relation vocabulary of spec 2 comes from Kruchten's ontology of architectural design decisions, and it reached this design from the research literature rather than from the tooling above. The two arrive at the same place from opposite directions. Obsoletes: in an RFC, Superseded by in an ADR and supersedes in Kruchten are one relation under three names. The doctrine reconciles them by adopting the published set and by claiming none of it for this entry.

What the tradition converges on

The record is immutable after acceptance. This is the property that separates the tradition from a wiki page about a decision. An accepted record states what was decided at a date, under the forces of that date. To edit it is to make every citation of it unreliable, because the reader of the citation and the writer of it then read two different texts.

Succession is the only amendment. A later record replaces an earlier one and names it. The earlier one stays, and it names its successor. That is what lets a reader who arrives at the old decision learn that it fell.

One decision per record. The temptation to violate it is constant, and the cost is exact. Two decisions in one record cannot be superseded apart. The first of them to fall takes the second with it, or it leaves the second alive under a heading that nobody reads.

The log is read as a log. A reader scans it by title and by date. That is the property that this entry meets a wall on, and the title facet is what it does about it.

The base package already carries most of this tradition

The base declares kinds.decision with purpose: rationale, voice: declarative, lifecycle: standard, an identifier scheme, and sections: {require: [Context, Decision, Consequences]}. Those are Nygard's three sections. The shelf docs/decisions/** beside it is homogeneous, which is one decision per document expressed as placement.

So this entry adds no decision kind. A decision_record kind beside the base decision would be two live paths for one tradition, and an adopter of both would meet a base kind and a base shelf that nothing fills. The design-spec entry had to do that for specification, and it recorded the cost as a finding. This entry does not repeat it, and the reason it can is a property of add rather than good luck. An add asserts that its addressed key is absent and asserts nothing about the mapping above it (spec 2). The base decision declares no facets member, so a bundle may write one.

What is left for this entry is what the base does not have: a readable name, the other register's kind, the edge that closes an obligation, and the two invariants that no check enforces.

The kinds

Kind	Purpose	Shelf	Where it comes from
decision	rationale	docs/decisions/**	the base, with a title requirement added
obligation_record	obligation	docs/obligations/**	this entry

An obligation record is not a decision, and it needs a kind of its own. The two registers of this repository sit beside each other and look alike, so the question is worth answering rather than assuming. A decision is closed: somebody chose, and the record states what was chosen and what it forecloses. An obligation is open: the corpus owes something, and the record states what would discharge it. Their reader intents differ, and spec 2 declares purpose per kind. A reader who asks "what is missing?" and reaches a register of settled choices got the wrong answer.

The sections say the same thing in the other direction. Context, Obligation, Discharge mirrors Context, Decision, Consequences deliberately. The middle section states the thing itself, and the third states what follows. What follows from a decision is its cost, and what follows from an obligation is the instrument that would pay it.

The word obligation names two things in this model, and the kind name keeps them apart. A taxonomy declares obligations as one of the thirteen declarations: an invariant that the corpus commits to, with a severity and a disposition (spec 4). This kind is a document about work that a corpus owes. The two are unrelated, and both are called an obligation in the specification today. So the kind is obligation_record and its identifiers read OBL-, which leaves OB- to the declaration. The finding is below.

The title facet, where it went, and what declaring it did not do

docs/spec/README.md is generated, and it labeled each document with HW-SPEC-vision-and-scope where the table it replaced read "Vision and scope". #123 measured that and recorded the cause: the taxonomy declares summary, which carries the scent role, and it declared no name. A summary is a sentence, and a reader who scans a log wants a name.

A tradition whose whole point is a readable log has to answer this, so this entry declared facets.title. It is stable rather than mutable, which is the tradition's own claim: an accepted record is amended by succession, so its name changes by rename and never by edit.

The facet reached this entry's kinds and the base decision, and it reached nothing else. The gap that #123 measured is on the specification shelf, and the design-spec kinds already declare facets.require as a list. An add cannot reach into a list that exists, and a bundle holds no override. So the entry that met the defect could not fix it where it was found.

headwater/standard 4.0.0 declares the facet, and this entry declares it no longer. That is the remedy finding 1 named: the base is the one source where vocabulary that several traditions need composes. The facet-role registry of spec 2 already carried name as its fifth member, so no meta-schema change was needed and none was made. This entry now requires the facet on decision and on obligation_record, and the design-spec kinds, evaluation and probe require it too (#427).

Declaring the facet was never the whole of what #123 measured, and this is the part the entry got wrong. The defect #123 named is on a generated shelf index, and the emitter that writes one read the identifier and never the facet in the name role. So docs/decisions/README.md listed 47 records as HW-DR-nnnn for as long as this entry declared title, and every one of those records declared one. The declaration was necessary and it was not sufficient, and nothing here noticed, because this doctrine states what a taxonomy declares and no reader of it opens the emitter. #427 found it by predicting that the shelf index would move on the day the specification shelf gained titles, and then reading a file that had not moved. headwater_generate::label is now one function that both emitters call.

The waiting_on facet, and why it is a state rather than a property

A record on this shelf states what the corpus owes and what would discharge it. It never states what it is waiting on right now, and that is the one question a reader of the whole shelf asks. HW-DR-0030 settles the shape and this paragraph states why the declaration is here.

It is not a value of status. status answers whether a reader may rely on the document. A record that waits on a ruling is fully current, because it states what holds now. The blocker belongs to the subject of the record and not to its authority, and one facet cannot hold two orthogonal things.

It is a present-tense state and not a property fixed at filing. Over the 129 records of the repository that wrote this entry, 38 of the 39 open records that wait on a ruling are followed by a build. One is followed by nothing and none by a measurement. So a property that had to name the act that discharges a record would be wrong about 38 of them. The value names the act that comes next, and the person who writes the ruling down is the person who moves it.

The set holds four values and no residual. A residual value means none of the others, so it changes on a record that nobody edited on the day a new class arrives. Spec 13 refuses an editorial facet over this shelf on exactly that ground, and the refusal holds. Each of these four names an act positively. Two of them hold populations that a two-valued set has no cell for: 22 records wait on an outside corpus and 8 wait on a run of an instrument that already ships.

A record that fits none of the four is a defective record. Its Discharge section does not say what happens next. Two of the 129 met that test on the first reading, and the change that declared the facet wrote the next act into each of them. That is the set with no residual working as designed rather than a missing fifth value.

On a record at status: discharged the value reads as what discharged it. A record leaves the queue by status and never by this facet. That a facet applies to one value of another facet has nowhere to be said, which is HW-OBL-0123, and this is a third instance of that record rather than a reason to invent a value.

The declaration is here because the engine put it here. An overlay that reaches add_to.kinds.obligation_record.facets.require is refused: does not commute with add.kinds.obligation_record in docs/taxonomies/decision-record/bundle.yml. The same refusal meets facets.optional. That is HW-OBL-0031 on a real change, and it means an adopter's overlay cannot extend this kind's facet contract either. So the four values answer to any corpus that keeps obligation records, and not to one repository.

A required facet with a closed value set and no role makes the scaffolder stop. headwater new obligation_record refuses to write a record without a value and names the four. Every record filed after this declaration states what it waits on, or it is not written.

The lifecycle ladder maps, and two of its rungs do not

The tradition's ladder is the reason this entry was chosen second. Proposed, then accepted, then superseded or deprecated, is what an ADR log is for. The base declares five states in vocabularies.lifecycle_state, and the four that regimes.lifecycle.standard names are the same graph. The fifth is discharged, which the base names in regimes.lifecycle.obligation and which kinds.decision therefore cannot reach.

The tradition	The base state	What it means here
Proposed	draft	Written, complete, and awaiting a ruling
Accepted	current	The team decided, and the record is now immutable
Superseded	superseded	A later record replaced it, and names it
Deprecated	deprecated	The decision no longer holds, and nothing replaced it

This entry adds no lifecycle state and no lifecycle regime, for the reason the design-spec entry found first. The states are a list under vocabularies.lifecycle_state, and an add introduces a key rather than an element. Two rungs pay for that.

It binds a second regime that the base declares, on one kind. kinds.obligation_record names regimes.lifecycle.obligation, which is standard and one edge into discharged. A record here is a debt the corpus wrote down, and it has an ending that a decision does not: the corpus paid it, and the record is kept as the receipt. To bind a declared regime is not to add one, so the add-only rule is untouched.

The names are the tradition's and the corpus cannot write them. A team that keeps ADRs writes proposed and accepted. Under this entry it writes draft and current. The core is satisfied either way, because the core is semantic and the roles survive the renaming (spec 2). What is lost is recognition, which is criterion 1 of the admission criteria: somebody who works in the tradition is meant to recognize it.

rejected collapses into deprecated, and the two are different facts. A rejected record was proposed and refused, so nothing ever relied on it. A deprecated record was accepted and later abandoned, so things did rely on it and may still. A reader who meets deprecated cannot tell "we considered this and said no" from "this used to hold". Both are terminal-retained, which is why the collapse is silent rather than an error.

The fix is one element in one list, and no bundle may write it. The finding is the second data point for an item that 13 — Open obligations already carries, so it sharpens that item rather than opening a new one.

Relations, and the one this entry adds

Succession and adjudication belong to the base, and this entry claims neither. Q18 ruled that an adjudication is a decision carrying overrides against the document whose effect it displaces, and spec 2 carries the whole Kruchten set with families assigned. The tradition that invented succession does not get to redeclare it.

One correction to the reading that this issue was filed under. It records that overrides needs no bundle, because "one line of overlay enables it" through $package.optional.overrides. That line does not resolve today. The base package manifest declares no optional block, nothing on disk holds the eight unenabled relations, and the resolver refuses a $package reference and names the open question as the reason. Enabling overrides therefore means writing the declaration out in full, in an adopter overlay. That is a fact about the package library rather than about this entry, and 13 — Open obligations already carries the shape of optional as an open item.

A decision reaches its evidence through the base traces_to. The relation is in the evidence family and runs from a governed document to a governed document, so a decision that names the evaluation which settled it needs no declaration here. An obligation record reaches the decision that produced it the same way, which is the structure that spec 13 already has in prose: each item sits under the decision that left it behind.

One relation is new, because one event has no form. An obligation is discharged when an instrument runs and reports. That is not a citation and it is not a derivation. It is the event that closes the entry, and this corpus has seventeen unmeasured claims waiting for it.

relations.discharges:
  family: evidence
  from: [evaluation]
  to:   [obligation_record]
  inverse: discharged_by
  reciprocal: required
  created_by: author
  evidence_at: from

The evaluation is the evidence, so the relation declares evidence_at: from. The edge points from the evaluation that substantiates to the obligation that it substantiates. The evidence family cannot say which end that is. Without the member, the rule that reads an evidenced claim against the warrant of its evidence read this relation the other way (HW-DR-0103).

The creator is the author, for the reason the design-spec entry gives for cites_evidence. A person types the front-matter line, and no verb, hook or agent proposes it. The owner ruled author on 2026-09-26, and HW-DR-0083 records the ruling. No verb, hook or agent writes a discharges edge, so the value hook would name an actor that nothing plays (HW-OBL-0105).

This entry declares no participation expectation, and the absence is the finding. An expectation states that a document of a kind, in a state, should acquire a named relation inside a window. The two failures this tradition actually has are neither. An accepted record that somebody edited has all its edges. A proposal that nobody ruled on is a document that never left a state, and no edge is missing from it. Spec 2 requires a window on every expectation, and an honest window does not exist for the other candidate either: an obligation that waits on a first adopter waits for as long as it takes, and this corpus says so about nine of its items.

The two invariants, declared as obligations with no control

Spec 4 makes an obligation data, and it gives every obligation exactly one disposition. An obligation that no control discharges states a gap with an owner, or states that no mechanism can exist. This entry declares three, and not one of them has a check.

OB-DR-1, immutability after acceptance. The tradition's central rule, and a check class that nothing else in this library needs. It reads a content change against a lifecycle state, so it needs the prior version of the document. Spec 12 supplies that input as needs_prior, and warrant.promoted is the first rule to declare one. So this is a gap with a named target and a worked precedent rather than a wish.

OB-DR-2, a proposal reaches a ruling. The reported failure of ADR practice is a log that fills with proposals. The rule is over state dwell time, which taxonomy audit already measures and no check reads.

OB-DR-3, one decision per record. This one is unverifiable rather than a gap, and the reason is the boundary that spec 1 draws. To count the decisions in a document is to reason about what its prose asserts. The homogeneous shelf and the section contract each carry part of the rule, and a second decision written under the same Decision heading is invisible to both.

Declaring the three costs nothing and buys two things. The register states what this taxonomy holds itself to, which is the whole reason obligations are data. And a later engine that gains the first two rules finds the obligations already written, with the argument beside each.

The core, declared

Criterion 5 of the admission criteria asks each entry to state its relationship to the invariant core. This is a full entry, and it satisfies the core through the base.

• state, state_entered, freshness and scent are the base facets status, status_since, last_verified and summary. Both kinds inherit all four from governed_document. title carries no engine role, and it claims none.
• rationale is served by the base decision, which this entry reuses rather than replaces.
• behavior is served by the base specification. This entry declares no behavior-serving kind, and it does not need to, because it removes none.
• Succession stays lifecycle-sensitive and untouched.

obligation is a fourth purpose, and the evidence-and-obligation entry declares it since HW-DR-0044. This entry uses it and declares requires: [evidence-and-obligation], which is the next section.

What this entry deliberately does not declare

A projection. The log wants an index by date and by state, and projections is a list in the base. The same rule that blocks a lifecycle state blocks this. It also blocks the artifact that made this issue urgent: the redirect map of #68 is a projection, and it has to come from an adopter overlay rather than from this entry.

A second identifier scheme for decisions. The base declares decision_id as {namespace}-DR-{seq:04d} and binds it to the kind. A tradition that renumbers is a tradition that breaks citations.

A language regime, or a voice regime. The base declares declarative and binds it to decision, which is right for this tradition: a record states what was decided, and it does not narrate the argument as it unfolds. An adopter that wants a controlled profile binds one in its own overlay.

A template placeholder syntax. The templates here write "{{today}}" in quotes. Every template of the first library entry writes it unquoted, and 13 — Open obligations records that a YAML loader reads the unquoted form as a flow mapping and refuses the whole block. The quoted form parses. It also makes the substituted value a quoted scalar, which Q2 never typed. This entry takes the reading that parses and records the assumption, and it does not edit the other entry's templates, because a finding about somebody else's package is a separate change.

The worked instance corpora, and what criterion 4 now reads

The fixture corpus is invented, and the corpus beside it is not. fixtures/corpus/ holds a miniature ADR log for a system called Beacon, and its own page opens "Everything here is invented, including the names." That corpus exercises the declarations of this entry. It cannot test the tradition, because one set of people wrote the rules and the prose.

Criterion 4 asks for a corpus this project did not author, and there is one now. fixtures/sources/inspect-evals/ holds the 10 files of adr/ from UKGovernmentBEIS/inspect_evals at 360484a06383f9260279938262d78ed646ddbca1, under the MIT license and unedited. tools/repo/decision-record-fixtures.sh assembles a corpus root from them at each run, adds front matter, and changes nothing else. CI runs it on every pull request. The fixtures page records the source, the revision, the paths and the run. The evaluation reports what the run found.

The first result belongs here, because it is a claim this doctrine makes. The base decision kind requires Context, Decision and Consequences. All 10 of the 10 real records carry all three, and no body was edited. Every one of them also carries Status and Date, and 7 carry Considered Options. The diataxis-site entry had to refuse all 4 of its 4 external documents on the same rule. The difference is that Nygard's three sections are a convention this tradition carries, and Diátaxis constrains a reader's purpose rather than a writer's headings.

The second result is the ladder above, measured where this doctrine predicted it. 4 of the 10 records write a status that the five states of vocabularies.lifecycle_state have no cell for. Three read Accepted — implementation deferred and one reads Accepted — Stage 0→1 active, Stage 2+ deferred. The front matter writes current, and the clause after the dash is gone.

What that clause names is this entry's own waiting_on facet, at the value build. The guidance on that value reads "the answer is known and somebody has to write the code or the prose", which is what those three records say in prose. A real ADR log invented, in a status line, the distinction this entry added as a facet. The facet is required on kinds.obligation_record and declared on no other kind, so a decision record of this entry cannot state it. This entry takes no declaration change on that reading. To widen kinds.decision.facets moves bundle.yml, which moves the lock and the package, and that argument belongs to a change of its own. 13 — Open obligations carries the finding, and HW-OBL-0123 is the adjacent record: a status that means "accepted, and not yet built" is one value of status that carries a second dimension.

The entry's second kind now has an external corpus too, and it satisfies no section it names. fixtures/sources/sysl/root.md holds Sysl's docs/ideas/root.md at 43c498362ce2951e7f7442430f52f780e8bb0b56, under the Apache-2.0 license and unedited. #830 named it for one reason: Sysl's docs/ideas/ carries no proposal convention at all, unlike an ADR log, so it is the corpus that tests whether obligation_record's own section contract holds against prose nobody wrote with this taxonomy in mind. It does not. The document is one bullet list under a single ### Future enhancement ideas heading, and it carries none of the three sections the kind requires. tools/repo/decision-record-fixtures.sh reports section.required.missing three times over the one document, once per absent heading, and the runner does not repair it: rewriting the heading to read Context, Obligation or Discharge would make the corpus say what this taxonomy wants to hear rather than what a team with no such convention actually wrote. The fixtures page records the run in full, including the constants the runner writes where the source supplies no value: status: draft, status_since and last_verified at the pin date, and waiting_on: build.

Findings

These go to 13 — Open obligations, on the route that the library index fixes. Nothing here reopens a closed decision. Three of the eight are new items in 13, three sharpen an item that is already there, one belongs in the glossary, and the last is a confirmation rather than a finding.

1. Composition between two library entries had no add-only form. This is the finding that a second entry exists to produce, and the entry met it three times in one bundle. HW-DR-0095 (Q67) later ruled the form. An entry may write into the keys of an entry that it names in requires, with add_to or add, and the dependency applies first. That discharged HW-OBL-0040. The three cases below are the record that the ruling answered, and each one now has that form at the cost of a dependency. case group 8 of the diataxis fixtures runs it.

• A shared purpose. obligation_record serves the obligation purpose, and the design-spec entry declared it. Criterion 6 refuses a second declaration at that address, and the resolver refuses it too, because both operations write the same leaves. So the remedies are a rename, which puts two names on one reader intent, or a dependency. This entry declared requires: [design-spec] for one declaration, and it inherited six kinds and three shelves of a tradition that an ADR-keeping team need not use. HW-DR-0044 closed the half of this that a split could close: purposes.obligation moved to evidence-and-obligation, alongside the one other address this entry ever reached into design-spec for, and this entry now declares requires: [evidence-and-obligation] in place of the line above. What it inherits now is one kind and two purposes, not six kinds and three shelves.
• An endpoint list of concrete kinds. The design-spec cites_evidence runs from: [design_spec, decision_register, obligation_register]. obligation_record is not among them, and no add reaches into a list that exists, so a kind of this entry cannot cite evidence at all. An abstract kind at the source end would have kept the endpoint open, and spec 2 says that abstract kinds are what make a bundle possible. HW-DR-0044 did not close this one: cites_evidence stayed in design-spec, because two of its three from kinds stayed there too, and obligation_record still is not among them.
• A facet requirement list. The same rule stopped title from reaching the design-spec kinds, which is where #123 measured its absence. HW-DR-0044 did not close this one. #427 closed it, by the third of the three resolutions the paragraph below names: the facet moved to the base in headwater/standard 4.0.0, and each entry now requires it on its own kinds. The requirement list stayed unreachable and the address stopped being one an entry owns.

The three share a root. Vocabulary that more than one tradition needs cannot live in an entry, because criterion 6 makes the first entry to claim an address its owner. The base is the only place where shared vocabulary composes. Either a purpose and a name facet belong there, or the resolver admits two identical declarations at one address and the confluence proof takes an exception. The split records one address that moved to a third place instead of either resolution, which is available whenever the address is a purpose or a kind and not a list.

2. The two checks of this tradition are about content and state over time, and no declaration expresses either. Immutability after acceptance reads a content change against a lifecycle state. A proposal that never reaches a ruling reads dwell time in a state. Every declaration in the language that carries a rule about time is a participation expectation, and an expectation reports one thing: a missing edge inside a window. Both failures here have every edge they need. The obligations OB-DR-1 and OB-DR-2 state the invariants with gap dispositions, and each names an engine input that already exists, which is what makes them gaps rather than requests.

3. A rejected decision and a deprecated one are one state. This sharpens the item that 13 already carries about list extension in an add-only overlay. The design-spec entry hit the limit twice and could name no case where the collapse loses a fact. This one can. rejected and deprecated are both terminal-retained, so the collapse is silent, and a reader cannot tell a refused proposal from an abandoned decision. The tradition also cannot write its own two rung names, which costs criterion 1 the recognition it asks for.

4. A generated index labels a document with its identifier, and the fix did not reach the shelf that showed the defect. This sharpened the item that #123 opened. facets.title answered it for this entry's kinds and for the base decision, and it could not answer it for the specification shelf, for the reason finding 1 states. #427 closed it in two parts, and only the first was the part this finding predicted. The facet moved to headwater/standard 4.0.0, and every kind whose documents a generated list renders now requires it. Then the shelf index still read HW-DR-0001, because the emitter read the identifier and never the name role, so the declaration this entry made had never reached the file that showed the defect. The remedy this finding proposed, a name role in the closed facet-role registry, was already in the registry and already unread by one of the two emitters that needed it.

The lesson this finding is kept for. A taxonomy declaration is not a rendered page, and this entry's doctrine could not tell the difference. Every sentence here was true and the reader still saw an identifier. A finding about what a reader sees is closed by opening the file the reader opens.

5. obligation names two things. Spec 4 declares obligations as a taxonomy block: invariants that a corpus commits to. Spec 13 holds obligations as owed work. The two are unrelated and the specification calls both an obligation. This entry writes obligation_record and OBL- to keep them apart in one corpus, and the glossary is where the general fix belongs.

6. A declared invalid_when reaches no check, so two live contradictory decisions pass in silence. The base declares conflicts_with with invalid_when: {both: {status: current}}. Spec 2 lists that state first among the four checks that the decision-relation vocabulary brings, and HW-EVAL-adjacent-work calls it a deterministic, blocking-eligible check that the design held all along. The engine reads the declaration in one place, where it makes the status facet count as read for the relevance canon, and no rule reports the state. The fixtures plant two current decisions that contradict each other, and the run is clean. This is a check that two specifications promise and no code performs, which is a different class of gap from a form that nothing states.

7. This entry adds no override and no remove, and it is admissible. That is not a finding, and it is recorded here because criterion 3 asks the question. Every operation in bundle.yml is an add at a key that the base and the design-spec entry both leave absent. headwater taxonomy resolve and headwater taxonomy validate confirm it over the merged result, with this bundle selected beside the first.

8. The entry's own two kinds gave two different answers to a question 13 already holds open. The decision kind's real corpus, inspect_evals, satisfies Context/Decision/Consequences with no edit to any body — the first result recorded above. The obligation_record kind's real corpus, Sysl's docs/ideas/root.md, satisfies none of Context/Obligation/Discharge: tools/repo/decision-record-fixtures.sh reports section.required.missing three times over the document's one heading, ### Future enhancement ideas, which is none of the three required names. The source was chosen for exactly this: #830 named Sysl's docs/ideas/ as the corpus "with no proposal convention at all", so the failure is the honest measurement of that absence rather than a defect in the assembly, and the runner does not repair it. 13 — Open obligations already holds the open question this sharpens: whether a kind of a library entry may decline to state a section contract. diataxis-site answered it once, from outside this bundle. This entry now answers it twice, from inside one bundle, at its two different kinds: a tradition that names its own headings satisfies a contract built from them, and a tradition that names none does not, whichever entry's kind asks.

What a migration inherits

This entry is the taxonomy, and the conversion of this repository is separate work. Both have run now, in two passes, and this section records what the entry predicted against what the passes measured.

#124 converted the decisions. Twenty-one records on docs/decisions/**, each of the base decision kind, with HW-DR-NNNN and the three Nygard sections. #126 converted the obligations. 101 records on docs/obligations/**, each of this entry's obligation_record kind, with HW-OBL-NNNN and the three sections that this entry requires. The estimate above said sixty for the largest list and the list holds sixty-four, so the count was low by four and the total was low by one.

Ready, and it was. The kinds, the shelves, the identifier schemes, and the sections. Neither pass changed a declaration of this entry. The second pass added one line to the adopter overlay, and it added no member and no kind here.

The two register documents became indexes, and one of the two is still authored. 9 — The decision register and 13 are indexes over the entries now, and a human writes each one. A projection would write either, and this entry may not declare one. What holds them back is content that no facet carries, which 13 states for both.

The citations held, and the rule that was to report a break does not read them. link.fragment.unresolved selects a link whose destination begins with #, so it never reads file.md#anchor. Both passes kept every cited heading, so nothing fell, and a script rather than a rule is what proved it. The second pass resolved 1,649 cross-file fragments by hand and found none broken.

09-open-questions.md is generated. #127 supplied the shelf_sections kind and #131 made a generated file a node, so the tombstone is a projection over the decisions shelf.

What a migration inherits that this section did not predict. The dates. An entry of a register carries no date, and the register's own history carries one for each entry. The second pass derived status_since per record from the earliest commit whose version of the register held that entry's opening words, and it set no date in bulk. The first pass had no such source and wrote one date across twenty-one documents. So the recoverable fact is a property of the artifact that held the entry, and not of the volume of the migration.