Three review rounds each found a real boundary case of the same
instrument: a fixed-width lexical window over word lists, asked to
decide whether English prose attributes a hex token to a commit. It
cannot. The window is arbitrary and the vocabulary is open-ended, so
each fix exposed the next case while looking like progress.
The judgment already lived downstream. This script's own docstring says
its output is adjudication input, and an agent reads every line before
anything acts on it — but the output was worded as a verdict ("does not
resolve to a commit ... Replace with the PR number") for tokens that
were never commit citations. That wording is what made every missed
phrasing a correctness bug.
So the outcome splits by confidence rather than gating on it. Commit
context present stays a FLAG. Absent becomes a NOTE that says the script
cannot tell a session id or content hash from a commit, and notes leave
the exit code alone. Nothing is hidden and nothing false is claimed.
That demotes the cue vocabulary from a gate to a ranking heuristic: a
phrasing it misses costs one tier instead of dropping a fabricated SHA.
Only then was tightening safe, so the generic `git` cue — which
precedes every object kind equally — is gone, and the review round's
three findings became notes rather than defects. The pin boundary is
fixed too: a backticked or parenthesized owner/repo@sha is a pin.
A token noted on one line and cited on another now reports the citing
occurrence, which the earlier dedup would have swallowed.
Adjudication prose in both skills learns the two tiers. The captured
learning is rewritten around them and renamed, since it argued for
tuning the gate this replaces.
Claude-Session: https://claude.ai/code/session_01HSahUJjXzZpJt5ztL7CUyt
Trevin Chow
ccc555bb89