Rendered from docs/decisions/0028-q28-whether-an-evaluation-is-governed-prose.md in the Headwater
corpus. Every document on this half of the site is typed by the taxonomy
the descriptor names: corpus.json.
Q28 — Whether an evaluation is governed prose
Context
Q27 bound the house language regime to decision and left one arm open. evaluation was the arm. An evaluation binds the narrative voice regime, and every bound kind binds declarative. Q27 adopted that line and said what was wrong with it. It is a line about voice rather than about lexical rules, and nothing states that a narrative document is free of a word limit.
docs/evaluations/ holds 15 records and 68,266 words of prose. No lexical rule read one word of it. #247 measured what a binding would report and counted 432 findings. A count is not a characterization, which is the bar Q27 set for itself, and this record meets it for the shelf Q27 deferred.
Decision
An evaluation is governed prose. .headwater/overlay.yml binds kinds.evaluation.language to ste_house and the lock records it. Four reasons follow, and the first three are measurements.
First, the 432 are one rule at one severity, and almost all of them are real. Every finding is language.controlled.not_met and every one is advisory. 422 name a sentence past the 25-word limit and 10 name a semicolon in running prose. The four error classes are a contraction, a British spelling, a hard-wrapped block, and a retired term that names a replacement. This shelf produces none of the four, so headwater check --strict exits 0 with the binding in place.
The findings fall on running prose rather than on a table. 408 fall on a paragraph, 21 on a list item and 3 on a table cell. None falls on a heading. The sentences are long rather than runaway. The median is 30 words and the longest is 52. 236 of the 422 sit within five words of the limit.
Second, the quotation reading accounts for 15 findings in 432. HW-OBL-0086 records that an inline quotation reaches every lexical rule as the author's prose. An evaluation quotes an external standard more than a decision record does, so the reading had to be priced here rather than inherited. Every inline quotation of five characters or more was replaced by one word, which is 159 quotations across the 15 files. The count moved from 432 to 417. Q27 measured the same experiment on the decision shelf at 64 to 63. The reading is real on both shelves and it is small on both.
Third, the reading that explains the rate is not authorship, and this shelf is what shows it. Q27 found 60 of its 64 findings on the five most recent records and read that as an era of authorship. This shelf refutes that reading directly. Every one of the 15 evaluations was drafted by an agent, and 14 of the 15 carry accepted_by: j.baxter, so a human read and accepted them. They still carry 432 findings.
The measurement that separates every group in this corpus is whether a lexical rule was reading the prose when it was written. Each row counts prose words as whitespace-separated tokens outside front matter, outside a fenced code block and not indented four spaces. No verb reports the figure, so the method is stated here and the table is a reading of the tree this record lands on.
| group | documents | prose words | findings | per 1000 words | was a rule reading it |
|---|---|---|---|---|---|
docs/spec/ |
18 | 119,133 | 9 | 0.1 | yes, throughout |
docs/obligations/ |
127 | 34,511 | 0 | 0.0 | yes, carried through the conversion |
docs/decisions/ 0001–0021 |
21 | 16,161 | 4 | 0.2 | yes, as prose of spec 9 |
docs/decisions/ 0027 |
1 | 1,462 | 0 | 0.0 | yes, written under Q27 |
docs/decisions/ 0022–0026 |
5 | 5,872 | 60 | 10.2 | no |
docs/evaluations/ |
15 | 68,266 | 432 | 6.3 | no |
This record is a seventh row and it reports zero, which is the same reading as the fourth. A denominator for it is omitted, because a word count of this document inside this document moves whenever the sentence stating it moves.
The four clean groups differ by a factor of 40 from the two dirty ones. Every document in every row was drafted by the same model family. Acceptance does not separate them either. Records 0001 to 0021 are accepted and clean, and 14 of the 15 evaluations are accepted and are not. What separates them is the history of the file. Records 0001 to 0021 were carried out of docs/spec/09-open-questions.md, which the editorial pass of 2026-08-09 rewrote under the regime. The obligation records carried their prose off a bound register, and the overlay moved the binding with it. Record 0027 was written after Q27 bound the kind. The two dirty groups were written where no rule was looking.
So a binding does not clean a shelf and it was never going to. It stops the next document arriving dirty, and that is the whole of what this line buys.
Fourth, an evaluation is the evidence of governed prose, and the taxonomy declares that rather than this record observing it. The cites_evidence relation runs from: [design_spec, decision_register, obligation_register] and to: [evaluation]. So evaluation is the only kind that can stand at the evidence end of a citation from prose the regime already binds. review_record, review_prompt, probe and probe_result cannot be an endpoint of it at all. The corpus agrees with the declaration. 14 of the 15 evaluations carry cited_by and 12 are cited by path from docs/spec/. No review record and no probe carries one.
Q27 excused docs/reviews/ as text this repository recorded rather than wrote, and excused a probe as the instrument of a measurement. Neither excuse reaches a document that the bound prose cites as its own evidence, and no other unbound kind is reachable by that relation.
Consequences
What a writer meets. An evaluation now answers to the six rules that read prose. A hard wrap or a British spelling in one stops a commit. headwater check --fix writes the spelling and leaves the wrap, because the source-form rule carries no patch. The word limit and the semicolon stay advisory, and the ste-editor skill is the route to prose that clears them.
The 432 stand, and nothing works them down. This record promises no cleanup. The corpus has never worked an advisory count down and the count has never fallen on its own. What the binding changes is the 16th evaluation, and record 0027 is the measured case of that effect on the other shelf.
A reader meets all 523 or none, and that is a ruling rather than a gap. Q12 refused a flag that scopes which findings count, on two grounds that survive this change. A flag that decides which findings count makes two runs over one tree disagree, and it turns an unchecked document into an unreported one. HW-OBL-0080 records that no binary ever carried one. headwater check --change carries a manifest and it reports the whole corpus, which was measured on this branch at 523 findings with one document named. So an advisory count here is read in full, and the number a reader acts on is the error count.
Why no adoption pair holds any of them, and the reason is measured rather than argued. The adoption block of the lock holds a (document, rule) pair under a task with an owner and an expiry. A pending finding does not fail a strict run. The grain of the pair is the whole rule. language.controlled.not_met carries the word limit and the semicolon at warning, and the contraction and the British spelling at error. A pair taken out for the word-limit backlog therefore covers the errors too.
That was measured on a scratch copy of this tree. A British spelling was appended to docs/evaluations/language-choice.md. With the binding and no pair, the strict run exits 1. With one pair on that document and the same rule, the strict run exits 0 and the error is reported as migration-pending under the task. So the debt mechanism would buy quiet by returning the one thing the binding is for. Q27 declined a pair because a pair would suppress nothing. The sharper reason is that a pair would suppress something.
What the binding refuses that nothing refused before. The same British spelling on the unbound tree moved the corpus from 91 findings to 91, and the strict run exited 0. On the bound tree it moved 523 to 524, and the strict run exited 1. The instrument is the injection rather than the count.
Where evaluation sits, and which arms are now empty. Every kind this regime binds also binds the declarative voice regime, except this one. evaluation is the first narrative kind under a controlled profile, which is the line Q27 adopted and this record retires. Six kinds stay unbound. docs/reviews/ exempts review_prompt and review_record by design. The words of a probe are the instrument of a measurement, which covers probe and probe_result. probe_transcript and specification hold no document here.
What reopens this. Two conditions, and each one is an event.
The first is an evaluation that has to quote at length. 15 findings of the 432 come from an inline quotation, and a record that quotes a standard for a paragraph would move that share. A block quotation reaches no lexical rule, and a directive naming reason=false_positive states the exception where a reader meets it. An evaluation that can use neither is the case that reopens this.
The second is the next evaluation. It lands under this binding. If it still carries findings at the rate of the 15 below it, the third reason above is wrong and the binding only reports. Record 0027 is the same test on the decision shelf and it came back at zero.
What this record does not settle. Whether an agent may write the acceptance stamp of a document it drafted is HW-OBL-0108. Whether warrant: proposed is a value of that facet at all is HW-OBL-0125. The grain of an adoption pair is a defect of the debt mechanism rather than of this ruling. It is filed rather than repaired here.