Rendered from docs/obligations/0004-working-tree-write-tools-have-no-measured-effect.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

Working-tree write tools have no measured effect on the assisted fraction

Context

Q7 rules three classes of tool and refuses a landed write. Working-tree write tools should raise the assisted fraction, which spec 3 declares as a tracked metric.

Obligation

That metric is the instrument, and a comparison needs a baseline and a treatment.

Discharge

The unassisted baseline that such a comparison needs is the typing pass that Q4's entry records.

The treatment arm exists and it holds one reading. headwater mcp --write registers the new tool, and the run that wrote HW-OBL-0111 on 2026-08-14 is the first authoring this repository has done over a protocol. The store names the arm of every reading it takes: .headwater/capture-cost.jsonl carries a surface term, and headwater capture groups by it. A line the store took before that day names no arm, and the report counts such a line under neither.

The instrument is not the one this record assumed. The four counts of a reading are derived from the plan, and a plan is the same plan whatever asked for it. So the per-run fraction cannot separate the arms at all. HW-OBL-0112 measures that: two readings of one kind, one from each arm, carry the same four pairs. What a surface can move is the count of runs and the mix of kinds those runs chose. This corpus holds five readings, three of which name an arm, so the comparison is still unmade.

And a reading of the arms says less than the two names suggest. The term records an entry point rather than a caller, and one agent produced every reading in both arms on the day the term arrived. HW-OBL-0111 holds that.