Rendered from docs/obligations/0001-the-promotion-fix-has-no-reading-of-the-assisted-fraction.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

The promotion fix has no reading of the assisted fraction

Context

Q4 rules that front matter is authoritative and that a relation instance is an object. It carries a claim that no run supports: the promotion fix should raise author-attributable edges without a rise in hand entry.

Obligation

The assisted fraction and the audit report are the instruments, and this corpus owes a reading of both. The assisted fraction now has a store that trends it. What stays open is a corpus that ran the verb over an edge, often enough to compare two populations.

Discharge

The first typing of this corpus supplies the other end of that scale. An agent typed 36 documents and 203 edge halves with no scaffolder and no engine, so the assisted fraction of the run is zero. What the first typing found records where the cost fell.

A scaffolder exists, and it reports the fraction of the run it just did. headwater new counts what it supplied against the four terms spec 3 names. They are the required front matter, the required sections, the identifier, and the halves of each proposed edge. This record carried a sentence that the treatment had never been applied to anything, and that sentence is false. What stays open is the claim itself, and the paragraphs below say why the verb does not settle it.

One run is not a trend, and the metric that Q4 tests is a trend. The claim is that author-attributable edges rise without a rise in hand entry. That compares two populations over time, and a verb that prints one number stores nothing.

The store exists, and it holds one reading of this corpus. .headwater/capture-cost.jsonl takes a line from every run of headwater new, and headwater capture reads it back. So the reading survives the end of a run, and this record can stop asking for a store.

The term that Q4's claim is about is the term that reads zero here. Q4 tests edges, and the fourth term of the denominator is the halves of each proposed edge. Every reading this repository holds carries edge halves 0 of 0, because no run of the verb named a relation. On 2026-09-27, a run of headwater new obligation_record that names one supersedes edge read 8 of 11, with edge halves 0 of 1. The new document opens at draft, so its far half is owed at promotion and is in neither term (HW-DR-0086). So a run at the initial state supplies no edge half, and this corpus has not exercised the term.

A taxonomy-wide reading exists and it is not this one. HW-OBL-0109 records 48 of 56 over the eight scaffoldable kinds, with no edge proposed in any of the eight. It is a reading of the denominator rather than of a trend, and it states what the denominator counts.

No reading of a committed corpus can supply the other half either. Q4 keeps created_by on the relation type, so the value states who a taxonomy expects to pay for an edge and never who wrote one. A scaffolded supersedes and a hand-typed supersedes are one string on disk. taxonomy audit counts edges by declared creator, which measures the intent of a declaration against a corpus. It cannot count what a scaffolder did. So the assisted fraction has to be taken at the moment the work happens, and every later reader of the corpus is blind to it.

The audit half of the instrument runs now, and it found that the comparison has no control arm. headwater taxonomy audit reports edge counts by declared creator, over relations rather than over edge instances, which is the grain Q4 fixes. Over this repository at --now 2026-08-14 it reports 397 halves under 9 relations: 359 under hook, 38 under scaffold, and 0 under agent. It reports that no relation of this taxonomy declares author, generator or import. Q4's claim compares an assisted creator against an unassisted one. This taxonomy declares no author relation at all, so the missing arm is a property of the schema. That is a different absence from the one the paragraphs above record.

Two capture rates are worth keeping beside it, because they are what spec 2 says the reading is for. When this record was written, traces_to stood at 80.7% of its 176 eligible documents, and governs stood below 1% of the same 176. headwater taxonomy audit prints the current capture rate of each relation. It also prints the share of the governed scope that a governs edge reaches, per pattern and in total. Both relations declare created_by: agent since HW-DR-0083, and until 2026-09-26 both declared hook, a role that HW-OBL-0105 records nothing played. No verb writes either, so every half of both is hand entry. So the spread between the two is inside one declared creator and one authoring route. The move to agent leaves the missing author arm where it was.

The one number this repository holds is from a run over a copy of its own tree. headwater new obligation_record supplied 8 of 9 units: four of five front-matter fields, three of three section headings, and the identifier. The summary was the hand entry. A run that proposed one succession edge supplied 9 of 11. The two extra units are the halves of that edge, and the near half is the target that a person named. Both readings cover one document each, and neither is a baseline for a corpus.