Rendered from docs/obligations/0111-the-capture-cost-surface-names-an-entry-point-and-never-the-caller.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

The capture-cost surface names an entry point and never the caller

Context

The capture-cost store took a surface term on 2026-08-14, and HW-OBL-0004 is the record that asked for it. The term separates a run of headwater new at a terminal from a run of the new tool of the MCP server. Two arms are what make Q7's comparison possible at all.

Obligation

The term records the entry point that a run arrived at. It records nothing about who drove that entry point. A person who types a tool call and an agent that emits one put the same message on one wire. The store holds one reading for each. So a difference between two rows of headwater capture is a difference between two entry points. It is never a difference between an assisted run and an unassisted one.

The corpus owes a reader of the two arms that sentence, or an instrument that separates the caller from the wire. This engine holds no such instrument. HW-OBL-0108 is the same question one level up: an agent typed the acceptance stamp of every document here, and no field records that it did.

Discharge

The store holds three readings that name an arm on the day the term arrived: one under protocol and two under terminal. One agent took all three, in one session, over one change. protocol names the run that wrote this record, over a server started with headwater mcp --write. The two terminal readings name the runs that wrote HW-OBL-0112 and HW-OBL-0113. A reader of headwater capture sees two populations, and one caller produced every reading in both.

A term that named the caller is refused for the reason the store refuses a person. A per-author number is a performance measure, and a performance measure changes the behavior that it measures. So the remedy is a measurement rather than a field. Spec 5 makes a transcript the one place where the corpus observes agent behavior rather than infers it. The run of 2026-09-11 was the first transcript that graded. A later change moved the lock under it, and since #1338 (PR #1448) its verdicts stand, graded over a moved read set. The run of 2026-09-16 is a third transcript and it observes tool calls alone, over eight sessions rather than four. It observes what two of those eight sessions wrote and committed, which is the closest this corpus has come to an agent-driven write outside a probe's own workspace, and it still records no capture-cost reading: surface is a fact headwater new writes at the moment a document is scaffolded, and a probe transcript is a fact about calls after the moment has passed. This record still waits on a measurement that no probe of this corpus takes.