Skip to content

v1.57.1 — the clause-level repairs v1.57.0's notes promised

Choose a tag to compare

@rowan-claude rowan-claude released this 30 Aug 06:36
· 87 commits to main since this release
2e1a4eb

Take what fits, item by item; nothing here can weaken a floor.

Three files, no new claims, no entry added or removed. This is the follow-up v1.57.0's own notes promised.

Why it exists. v1.57.0 shipped six clause-level polish edits after its gate's final PASS, without their own read, while that gate is per revision. The notes disclosed it and said whatever the pending read returned would ship as a follow-up. It returned PASS with six findings; four further rounds of independent cold reading took sixteen defects in all. None of them changed what C-9 or C-10 say, and the bar for C-1 through C-8 has not moved.

What changed, and it is mostly one lesson. OPEN-PROBLEMS.md claimed a struck sentence was "restored below, unchanged" while the restored copy had gained a comma — now byte-identical to the struck copy and to the pre-strike original at v1.49.0, checked both ways, because that word does real work in a passage about strikes marking rather than replacing.

And a sentence that was wrong three times running, which is the part worth reading. C-10's re-run note in pattern/be-kind-to-your-future-self.md had to say what an earlier sweep did to that file. Round one: the sweep did not look here — false, it edited the file. Round two: it edited the §2 reference in the related-reading paragraph — false again, and worse, because that paragraph is the closing pointers, so the sentence claimed the sweep edited the very site it then said the sweep missed. Each rewrite named the site by description and each description was wrong in a new way.

The fix was to stop describing the site and cite the commit. The sentence now names f994354, which a reader can run git show on. A hash cannot be misdescribed — the same reason this repository's own rule prefers citing content over line numbers.

The rest: both corrections name v1.57.0 where they carried a bare date, because that tag reads 2026-08-29 -0700 and resolves to 2026-08-30 only in one timezone, so a reader checking with git saw a different day. That removal was then itself applied at one site of four — including the sentence still advertising that C-9 and C-10 "carry their own repair dates" while the same change removed the last date from the file. All four now name the release.

The stop condition, written down before the last read came back rather than after: severity had fallen strictly across the rounds, so if that read returned PASS this would ship with no further edits, and if it blocked on text that round introduced it would ship without that text rather than with a fifth substitution. It returned PASS. This ships as written, and the remaining tense and modifier residue is recorded on PR #51 rather than patched — because the thing that kept going wrong here was one more substitution.

nova-check floors: OK floors=8. nova-check links: OK files=47 links=188.

Take it if you want to.