Rendered from docs/decisions/0063-every-required-facet-of-a-generated-document-is-derived-and-the-emitter-composes-the-summary.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Every required facet of a generated document is derived, and the emitter composes the summary

Context

A generated document's kind requires what every other document of that kind requires. Two of them answered almost none of it. #780 reported the shape and understated the count. headwater explain gives the measurement at 640b51ea. docs/spec/09-open-questions.md requires seven facets, and docs/probe-results/regression-probe-transcript-for-2026-09-11.md requires five. The identity block answered three between them, so nine stayed unwritten. The three are the discriminator and the name of the register, and the name of the probe result. The sequence in the register's file name was one of the nine, and the layout is where this change now reads it.

Nothing reported those nine. The reason is a gap between two mechanisms rather than a defect in either. The census classifies a marked file as generated, so no document rule creates an instance over it. The headwater generate contract states that exemption and names generate --check as the reader that holds the file instead. That check compares committed bytes against what the emitter produces now, so it proves the emitter agrees with itself. It cannot report that the emitter produces bytes the taxonomy refuses.

The absence had one measured consequence and one visible one. relations.supersedes declares on_target: {set_state: superseded}, so an edge of it puts that state on its target. Two live registers declare such an edge into docs/spec/09-open-questions.md, and that file carried no state facet at all. Spec 13 measured it as the only one of 192 nodes with none. Both instances of the dependency rule then skipped rather than decided, so a relation stated a fact that no rule could reach. That is #227. Separately, three rows across two shelf indexes carried a label and no cue, because the documents they name declare no facet in the scent role.

Two shapes could close it, and one of them costs a ruling spec 6 already made. The block could take a member for each facet. Spec 6 refuses that shape. A taxonomy source sits outside the corpus root, where no census row covers it, no language regime binds it and no rule reads it. That is the refusal the shelf-sections emitter already made against a template. A member that took an open mapping of facets would re-admit the same prose under a different syntax.

Decision

The block stays closed at three scalars, and the engine writes every other facet the kind requires. Derivation is not declaration. It is the channel this emitter already used twice. The kind of a document on a heterogeneous shelf comes from the shelf, and the reciprocal half of every incoming edge comes from the graph.

engine/crates/generate/src/derived.rs computes five values. Each one is found through the facet role rather than through a facet name. The reason is the one the label emitter already reads a role for: title means something in one taxonomy and nothing in the next.

role	the value, and where it comes from
state	the state that an incoming edge sets through on_target.set_state. Then the terminal state of a source the emitter names, and otherwise the state whose own role is live
state_entered	the stalest date on the documents that set the state, or the newest date on the documents this projection read
freshness	the stalest date on the documents this projection read
scent	composed by the emitter, which is the one value here that nothing derives
no role	a facet that the shelf's layout names, read back out of the output path

The owner ruled on 2026-09-11 that the emitter composes the summary. An asymmetry decided it. A generated document's body is already prose that its emitter wrote and that no rule reads. A summary beside that body is no more ungoverned than the body. So a refusal of the summary alone rests on nothing that spec 6 states.

No value comes from a clock. Spec 6 rules that no generated file states when it was generated. A timestamp makes every run differ from the last over a corpus that nobody touched. A date folded from the dates the inputs already carry is not a statement about the run. So the rule stands untouched and the regeneration gate holds these bytes like every other byte of the file.

For freshness, and for a state that an edge sets, the stalest date answers rather than the freshest. A file assembled from other documents is only as fresh as the oldest thing it carries. The freshest date would state a confidence that no source supports.

Amended 2026-09-25: when no edge sets the state, the date the state was entered is the newest date on the documents this projection read. The first text took the stalest date for that case too. The owner ruled on #820 that a page is no fresher than its newest input. The page changes each time one of its inputs changes, so it entered its state with its newest input. This record is not superseded. The state, freshness, scent and no-role rows do not change. The state_entered date of a state that an edge sets is still the stalest date on the documents that set it.

Amended 2026-10-01: a probe result over a transcript at a terminal state takes the state of that transcript. This rules the judgment that the first text left open under Consequences. #1509 measured it at 4aaebb4f. All 35 results under docs/probe-results/ stood at current. 24 of the 35 graded a transcript that was not current: 23 at deprecated and 1 at draft. 5 of those 24 pages printed 6 sentences that stated a direction at the 5% level over a withdrawn run. The declaration route needs a hand-written edge on each transcript when an author withdraws it. None of the 23 edges existed, and nothing reported the gap.

The rule has three parts, in this order:

• An incoming edge that sets a state still wins, because it is a declaration about the result itself.
• Otherwise, a transcript whose state has a terminal role gives the result that state value. engine/crates/generate/src/derived.rs reads it from Composed::state, and admitted refuses it where the kind's regime does not name it.
• Otherwise, the result stands at the live value. A transcript at draft keeps the result there, because draft is the wrong word for a file that no author argues over.

The date the state was entered follows the rule for the case with no edge. It is the newest date on the documents this projection read, which is the date of the transcript. The rule reads the role of a state, and it names no state value. A shelf index and a sections file pass no source state, because one withdrawn row does not withdraw the file that lists it.

The body carries the rest. A result whose transcript is not at a live state names that state in its first paragraph, before any figure. It says that no figure is a current finding. Each comparison in it keeps the interval and states no direction at the 5% level. This covers the draft transcript too.

This amendment opens again on one condition. 13 documents outside docs/probe-results/ name a result over a transcript that is not current, by its path or by its HW-RESULT- identifier. They are this record, 2 specification parts, 7 obligation records, 1 evaluation and 2 transcripts under docs/probe-runs/. The condition is that the terminal state raises findings on them that a prose edit cannot clear. The fallback is then a recorded decision that every result stays current, and the first paragraph of its body says why. Over 4aaebb4f the change raised no lifecycle finding. headwater check reported none before the change and none after it.

A facet outside all of that is refused rather than defaulted. The refusal in identity::front_matter now reads the whole required set against everything written into the file. What it guards is a taxonomy that requires a facet in no role this engine reads and in no shelf layout. No author can add such a facet, because a generated document's only writer is this engine. No check reads one either, because of the exemption above. An invented value is worse than a refusal, because somebody will cite it.

A section that the kind requires and no emitter writes is refused too. This record ruled on facets, and #409 asked the same question of sections.require. A generated body is composed from the corpus rather than from a contract. So a heading that answers one is an accident, and a rename undoes it. The refusal reads the composed body rather than the contract alone, so a body that does answer the contract is written. The repair is the kind rather than the declaration. A kind whose population a projection writes states the section contract that the projection keeps, or states none.

Consequences

#227 closes. docs/spec/09-open-questions.md states superseded, derived from the two incoming edges that put it there. engine/crates/check/fixtures/corpus.checks loses the line that recorded both dependency instances as skipped for want of a state there. Neither instance reports a finding. The edge is the cause of the state rather than a dependency on it, which is the reading that rule already took.

Three rows gained the cue they had none of. They are the row for the open-questions register on docs/spec/README.md, and both rows of docs/probe-results/README.md. That closes a remainder HW-OBL-0044 recorded when it discharged, which was one row of sixteen with a label and no summary. That record already stands at discharged, and this change does not edit it.

The exemption does not end, and that is deliberate. No check reads a generated document, so these nine facets are written and unchecked. They are written for three reasons. A reader meets them, and an index and an export carry them. The third is that a relation which put a state on its target had nowhere to put one.

A projection now reads another projection's front matter, so the verb reaches a fixed point in two passes. The shelf index carries the summary of every document on its shelf, and one of those documents is generated. The first pass writes the summary and the second pass carries it into the index. This is not new with this change, because the name that #627 admitted had the same property. A committed corpus holds the fixed point, and generate --check reproduces a fixed point in one pass. That is why the gate never saw it. Since #842, one headwater generate does both passes. It reads the tree again after each write and stops at the first pass that writes nothing. A run that still writes on its fourth pass exits non-zero and says that it did not settle.

One judgment could go the other way. Each probe result is written as current over a transcript that is not. One transcript stands at draft and the other at deprecated, so the gap is two states wide rather than one. The result is an accurate statement of what the corpus derives now, and its own body says the run was refused. The alternative couples a projection's state to the state of its inputs. draft is also the wrong word for a file that no author argues over. The taxonomy already holds the mechanism for the other answer: a relation from the transcript that declares on_target.set_state is honored by this change. So the general question is a declaration rather than a behavior of the engine, and this record does not rule on it. The amendment of 2026-10-01 under Decision rules it for the probe result.

One adjacent gap stays open and this change does not reach it. corpus.checks still records one skip over the same document. warrant.evidence.unsupported cannot say whether an evidenced claim rests on it, because the document declares no warrant. Spec 3 rules that the engine derives regenerated from the marker and never from a declaration. So the warrant of a generated document is known, and the rule reads the declaration instead of the derivation.

Ten of the eighteen kinds in this repository's lock declare sections.require, and neither generated kind is one of them. So the section refusal reports nothing here, and engine/crates/generate/fixtures/unheld.taxonomy.yml is the whole of the evidence for it. That is the standing that unrolled.taxonomy.yml already has. every_identity_this_repository_declares_supplies_what_this_refusal_reads is the guard, and it prints its numerator and its denominator. A kind that gains a section contract and a generated member fails it.

One clause of the same question is #945. Whether a derived state is one that the kind's own lifecycle regime admits was unmeasured, and this change did not measure it. The state that the amendment of 2026-10-01 reads from a source goes through the same admitted refusal. probe_result binds the standard regime, which names deprecated.

Two documents of the specification series now state sequence 9. They are docs/spec/09-decisions.md and the generated docs/spec/09-open-questions.md. Both file names carry that number, so the derived value is true of the path it was read from. The tombstone carried no sequence at all before this change, so the pair is newly visible rather than newly wrong. No rule reads a shelf layout after birth, which HW-OBL-0106 already records, so nothing reports the collision.

A derivation never folds the file it is writing, and two filters hold that. A generated file that declares an identity is a node of the census. So a projection whose output sits on the shelf it reads finds its own last version among the documents it folds. A fold over that set is a function of its own previous answer, whether the fold takes the minimum or the maximum. A committed date that no source supports would then stay the answer on every later run. The regeneration gate would hold it, because the emitter agrees with itself. That is the same failure the census exemption assumes a second reader for. Neither declaration in this repository reaches the shape, so a fixture taxonomy builds it and one test watches the date fail to move. This was found by a review of the pull request rather than by the build.

Two crate dependencies moved. headwater-generate now names headwater-resolve and headwater-scaffold as dependencies rather than as dev-dependencies. The first carries the one hardened YAML scalar escaper in this workspace, which had four escape routes found by running the binary against it. The second carries the tokenizer that writes a file name from a shelf layout. Reading a path back for the facet it names is the same grammar. A second copy of either would let this engine disagree with itself about bytes it wrote.