The closure

Across this chapter the unit of work has been a rung: a thin slice of value, specified only enough, built and proven before the next begins. The word that says it is proven is done, and on this project done has a single, mechanical meaning. Each rung carries a closure — a finite, named set of acceptance checks, fixed up front from the spec (A1.04.2). The rung is done when the set is closed: every check has executed and passed, and nothing is outstanding.

The contrast is with “looks done.” “Looks done” is a statement about appearance, made by a person glancing at the work. “Closure complete” is a statement about a set you can enumerate and re-run on demand. The first is a feeling; the second is a fact a machine can confirm. Once the closure is the definition, “is it done?” stops being a debate and becomes a count.

The tempting error is to read a high pass-count as “mostly done.” It is not. A closure is closed or it is open; there is no “eighty percent closed.” Four checks of five green is not done — and the one outstanding check is, more often than chance, the sad path everyone hoped to skip: the duplicate, the empty input, the second click. The count that matters is not how many are green; it is whether any is outstanding.

Move the slider to set how many of the five checks are green, and mark whether the sad-path check (C3) is among them. Only 5/5 reads done. Every count below it — however high — reads not done, and a 4/5 that is missing C3 is the common false “done” that ships the duplicate to a real learner.

Take the canonical rung “enrol in a course.” Its closure is the five acceptance checks the figures use: browse lists courses, enrol records an enrolment, enrol twice creates no duplicate, opening a lesson returns its content, and progress advances and persists. The code can compile, and C1, C2, C4 and C5 can be green, and the rung is still not done while C3 is outstanding — because “enrol twice creates no duplicate” is exactly the invariant a learner hits in the wild.

The Author (the Claude agent) builds the enrol path to the spec; the Operator (the human) accepts the rung against the executed checks, not against a glance at the screen. The closure is closed only when C3 is proven too — at which point the rung is correct, because that is what correct was defined to mean.

# The closure for the rung "enrol in a course" — five acceptance checks.
#   C1 browse       → the catalogue lists courses
#   C2 enrol        → an enrolment is recorded
#   C3 enrol twice  → no duplicate            (the idempotent sad path)
#   C4 open lesson  → its content is returned
#   C5 progress     → it advances and persists

# 4/5 green with C3 outstanding is NOT done — the set is not closed.
done? = Enum.all?(closure, & &1.passed?)   # => false while C3 has not executed green
# done only when every member has executed and passed.

Done is the state of a closed set: a finite, named closure of acceptance checks, fixed from the spec, with every member executed and passing and nothing outstanding. “Looks done” is a feeling about appearance; “closure complete” is a count you can re-run. A majority is not a closure — one outstanding check, usually the sad path, leaves the rung not done. The next dive sharpens what it means for a check to be a real member of the set: it must be proven, not merely asserted.

Primary sources for “done” as a set of passing checks, and where it connects in the course.