What moved between versions of the coder instrument, and why.
What it holds
v31 to v32 (2026-09-09): emit.thread refuses a read
out_of_scope disposition of a thread with ten or more
in-window items unless excluded_quote is what
cite returned; C5 counts in-window items from
coder_corpus.jsonl. v30 to v31 (2026-09-08):
close_run.py --preflight, C11 (a coded thread
keeps a live code), C12 (branch-B notes repeat the gate), and a guard on
branch-B raw reads before step 4. Earlier history: v29's case lead and
two branches, the forest, the facts files, the audit, and the runs that
tested them.
Who writes it, and when
The instrument's author, at each version change. Dates are given per section.
Why it exists
The instrument README says: "CHANGELOG-v32.md says what
moved and when". The v32 section gives the reason: an audit "found a
systematic C5 failure in completed v30 rooms". The rationale beside the
instrument, rationale/V32-C5-PREVENTION-RATIONALE.md, gives
the same reason and adds that the method, evidence standard and outputs
did not change.
Where
CHANGELOG-v32.md; V32-C5-PREVENTION-RATIONALE.md.
Checked against: CHANGELOG-v32.md, README.md (instrument), V32-C5-PREVENTION-RATIONALE.md