Rendered from docs/obligations/0002-declarative-voice-is-called-detectable-at-useful-precision.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Declarative voice is called detectable at useful precision, and the sample is unreachable

Context

Spec 8 calls declarative voice "mechanically detectable at useful precision". The measurement that closed Q5 covered three ASD-STE100 structural rules, used as proxies, and none of the three survives. The passive, progressive and auxiliary detectors lived in tools/ste-lint.py and retired with it. Q5 measured their false-positive rates at 8%, 83% and 56%, and no rule blocked on any of the three.

So the measurement that closed Q5 does not run here, and the instrument that produced it is recoverable rather than lost. tools/ste-lint.py stands intact at a39e3fb^, which is commit 9f0d1d8. The evaluation records the method beside it. The seed is 20260811, and the samples are 60 of 449 passive findings, 25 of 54 auxiliary, and all 12 progressive. So a reader can repeat that measurement.

A rebuild of the four detectors adds nothing to this claim. The evaluation states that it did not measure future_intent, change_narration or phased_rollout, which are the categories that the declarative regime forbids. A detector nobody acts on is a warning count, and its reading is already recorded.

The edge to that evaluation is traces_to and not discharges, and the difference matters. The decision-record entry declares discharges for the event that closes an entry, which is an instrument that ran and paid the debt. This record is open, and the evaluation says itself that it did not measure the three categories above. So discharges would state a closure that no run reached. What the taxonomy has no relation for is the honest link, which is an open record and the evidence it rests on. cites_evidence says that and runs from the three register kinds rather than from a record.

Obligation

The instrument is a run of the three categories over this corpus, with an adjudicated sample of at least 50 findings each.

Discharge

voice.forbidden_construction implements future_intent, change_narration and phased_rollout, which are the categories that the declarative regime forbids. So the implementation exists, and what this record holds open is the sample.

The reading of 2026-09-11, on commit 58f46a8d. change_narration reports 36 findings, and the other two categories report none. So no category reaches the bar of 50. Edition two of the change_narration set reported 60 findings on the day it landed. This record then carried that figure through two merged rewrites that removed most of them. Write a count here as a dated reading that names the commit. A bare count goes stale on the next rewrite, and nothing reports it.

All 36 findings were adjudicated one at a time, under the two labels spec 4 declares. 13 are genuine. 5 stand in a section where the narration is the content a reader wants, such as the ## Context of a decision record. 18 are false for good, and each one is a way that English states something other than a change. The five modes are a definition, a conditional inside a hypothetical, an inline quotation, another system's behavior, and a heading. The loosest pattern of the set carries 24 of the 36 findings, and 9 of those 24 are genuine. That pattern stays, and the argument is the absolute count rather than the rate. The rate moves in both directions. Over genuine findings alone it falls from 13 of 36 to 4 of 12. Over the findings that are not permanently false it rises from 18 of 36 to 9 of 12. The yield does not move, because 9 genuine findings go and no numerator returns them. engine/crates/check/src/voice.rs holds the ruling, one sentence for each of the five modes, and the trend that would retire the pattern. None of the 18 carries a headwater allow directive, because a directive requires an until and none of these sentences becomes wrong.

The zero from the other two categories is saturation rather than a clean corpus. On that same commit, of the patterns that future_intent declares, 0 of 14 occur in the 291 documents whose kind binds the declarative regime. phased_rollout stands at 0 of 13 over the 293 documents that bind the declarative or the prospective regime. Both denominators come from the kind-to-regime binding of the lock, applied to the 320 documents of the graph. The corpus holds one occurrence of will become and two of in the first release, and all three sit under docs/reviews/ or docs/evaluations/, which bind narrative and forbid nothing. HW-OBL-0168 states why a zero from a pattern set carries two readings rather than one.

So this claim is short for all three categories rather than for two, and the reading above is the measurement that says so. A wider corpus is one way to reach the sample, and this record waits on one. A pattern set curated against the prose that the two silent categories are written in is the second way. It asks nothing of an adopter. Neither route is taken here.

The reading of 2026-09-27, on commit 0ee88a2d. The reading of 2026-09-11 counted the body alone. Edition four of the rule also reads the summary facet, so this reading counts the body and the summary together. The rule reports 23 findings over 376 rule instances. change_narration reports 22, future_intent reports 1, and phased_rollout reports none. So no category reaches the bar of 50, and change_narration is further from it than on 2026-09-11.

All 23 findings were adjudicated one at a time. 7 are genuine, and each one narrates a change to this repository. 4 stand where the narration is the content a reader wants: 3 in a ## Context section and 1 in a dated measurement. 12 are false, and no rewrite repairs them. 11 of the 12 are change_narration, in the five modes above. The twelfth is the one future_intent finding, and it is a sixth mode: a general prediction that states a property of a thing. It is in docs/taxonomies/diataxis/doctrine.md. So the first finding that future_intent has reported on this corpus is not evidence that the set detects the fault.

Each finding carries one label. genuine is a finding to repair. accepted_deviation and false_positive are the two labels that spec 4 declares, and this reading spends no directive on either.

document and line	pattern	label	reason
docs/decisions/0008-probe-cost-and-cadence.md:26	there was no	accepted_deviation	history in the ## Context section
docs/decisions/0012-migration-path-for-an-existing-corpus.md:42	no longer	false_positive	the present behavior of another tool
docs/decisions/0017-governed-access-and-the-solution-layer.md:94	did not exist	genuine	tells the change that a ruling made
docs/decisions/0022-q22-the-integrity-posture-of-a-published-package.md:26	did not exist	accepted_deviation	history in the ## Context section
docs/decisions/0024-q24-readability-and-what-a-sweep-can-be-asked-about.md:54	no longer	false_positive	a conditional inside a hypothetical
docs/decisions/0039-q39-how-a-figure-reaches-a-hand-built-page.md:54	no longer	false_positive	the condition that makes a script fail
docs/decisions/0051-q51-what-licenses-a-term-into-the-public-glossary-and-what-licenses-a-sixteenth.md:48	did not exist	accepted_deviation	a dated measurement in the ## Consequences section
docs/decisions/0054-the-upper-bound-of-a-reconcile-first-allocator-is-the-corpus-and-a-claim-store.md:49	no longer	false_positive	a definition
docs/interfaces/headwater-probe.md:33	no longer	false_positive	a definition
docs/obligations/0129-spec-12-calls-two-phase-a-outcomes-structural-findings-and-the-engine-emits-none.md:29	no longer	false_positive	a definition, in the ## Context section
docs/obligations/0130-the-publication-date-of-this-repository-is-unset-and-the-owner-alone-sets-it.md:28	there was no	accepted_deviation	history in the ## Context section
docs/spec/07-distribution-and-federation.md:159	no longer	false_positive	a conditional inside a hypothetical
docs/spec/glossary.md:592	no longer	false_positive	the definition of a retired term
docs/spec/glossary.md:811	used to	false_positive	a heading
docs/taxonomies/README.md:225	used to	genuine	tells what a fixture README said before
docs/taxonomies/README.md:233	has since	genuine	tells the history of a milestone
docs/taxonomies/decision-record/doctrine.md:69	no longer	genuine	tells a move of a facet to the base
docs/taxonomies/decision-record/doctrine.md:100	no longer	false_positive	a table row that defines a state
docs/taxonomies/design-spec/doctrine.md:137	has since	genuine	tells a change to spec 2
docs/taxonomies/diataxis/doctrine.md:46	will eventually	false_positive	a general prediction that states a property
docs/taxonomies/diataxis/doctrine.md:137	has since	false_positive	a conditional inside a hypothetical
docs/taxonomies/diataxis/doctrine.md:141	has since	genuine	tells that a demonstration ran
docs/taxonomies/evidence-and-obligation/doctrine.md:28	used to	genuine	tells what an older kind carried

The rate of each pattern is its genuine findings over its findings. no longer is 1 of 10. has since is 3 of 4. used to is 2 of 3. did not exist is 1 of 3. there was no is 0 of 2, and both are accepted_deviation. will eventually is 0 of 1. Two labels are close calls. The line of 0051 could be genuine, because it tells a change to the site. The line of the design-spec doctrine could be accepted_deviation, because its section records what a draft assumed. Neither call moves a ruling below.

The loosest pattern now yields 1 genuine finding of 10. no longer carries 10 of the 22 change_narration findings. Since 2026-09-11 its genuine count fell from 9 to 1, and its false count fell too. The quotation exclusion of #783 removed one false finding, and rewrites made for other reasons changed the words of others. So the false count is not a floor that no rewrite reaches. The ruling in engine/crates/check/src/voice.rs retires the pattern when a reading finds the genuine count at zero and the false count unmoved. This reading sets the baseline for that ruling: 1 genuine and 9 false. A later reading compares its false count with 9, not with the figures of 2026-09-11.

The two quiet categories are still at saturation. Of the patterns that future_intent declares, 2 of 14 occur in the 378 documents whose kind binds the declarative regime. Both occurrences of will be are inside a quotation, which the rule does not read. The other pattern is the false finding above. Of the patterns that phased_rollout declares, 0 of 13 occur in the 382 documents that bind the declarative or the prospective regime. change_narration has 5 of its 23 patterns in the same 382. Both denominators come from the kind-to-regime binding of the lock, applied to the 421 documents of the graph. 39 of those documents bind narrative.

No voice regime can name the sections that it reads, and this record prices that ruling. A decision record keeps history in its ## Context section, and the rule reads that section as it reads every other section. 4 of the 23 findings stand in a ## Context section. 3 of them are narration that a reader wants there, and 1 is a false definition. A section scope clears those 4 findings. It does not clear the fourth legitimate narration, which stands in a ## Consequences section. So a schema change buys 4 findings, and it is declined. The convention for an author is a block directive with reason=accepted_deviation, an until and a note. The convention costs 4 directives today, and this corpus spends none. A directive requires an until, and a sentence about history does not become wrong on a date.

This record still waits on an adopter. The second route is a curated pattern set, and build would say that its answer is known. No candidate pattern for future_intent or phased_rollout has been measured against this corpus. So that route waits on a measurement that nobody has designed, and the wider corpus is still the route with a known shape.