Nothing to upgrade for. references/, scripts/, SKILL.md and AGENTS.md
are byte-identical to
v1.4.2 — verified
by diffing the exported tree. This tag exists so a validation run is citable; if
you are on 1.4.2 you already have this content.
What was recorded
- Full suite: 25/25 against the installed 1.4.2. The first run since
evaluation/stopped being installed in 1.4.2, which confirmed the point of
that change: no agent cited the suite, and the one that had previously found
and quoted its own expected answer reached the same refusal from the references
alone. The 1.4.2updatedAfterfix surfaced in two answers in the words the
fix added. - Two run-hygiene findings, neither affecting shipped content:
- The repo's git history still serves claims that were later retracted. One
agent rebuilt part of an answer from a superseded commit message and
presented a mechanism the current docs explicitly contradict. Its conclusion
was right, but this cannot be fixed by excluding a path — history is
immutable and agents read it. - An eval agent wrote a 316-line file into
scripts/, which is part of
the exported tree; an unnoticed commit would have shipped unreviewed code.
Untracked and removed.
- The repo's git history still serves claims that were later retracted. One
The method note in evaluation/scenarios.md now says to install via
git archive rather than cp -r, to check git status after a run, and to
distrust any answer whose sources include commit messages.
Full detail in CHANGELOG.md.