1.9.3rc11: upkeep keeps its quotes

2026-10-20

This is the historical rc11 release account; the latest preview is listed on Releases. rc9 shipped two REM bugs, and rc11 fixed both.

Daily upkeep could rewrite a page and lose what it had. In one replay, a page with twelve quotes came back with none. Upkeep now checks a rewrite against the page it replaces, the way enrich and muse already did.

The pronoun repair guessed. It turned a sentence about the subject's manager into one about the owner's. It now only rewrites he, his and she, and never the form the page uses for its own subject.

How each one was found is in The repair that guessed. The deepen picker changes are in An empty page is not a reason.

python -m pip install --upgrade 'connectonion==1.9.3rc11'

27 October: keeping the words twice

Preserving evidence can make a page worse in a different way. A saved nightly candidate and the current page differed by two extra lines, consistent with a review that recorded two restorations: an almost repeated History event and an old lead below Uncertainties: Unknown. That comparison does not establish the exact promotion postimage or rule out a later edit. A second repeated obligation was already in the candidate, so the restoration code could not explain it all.

An invented example isolated one deterministic cause. Nora's earlier quote was “I will reply after Friday”; the new quote kept those words and put a full stop inside the marks. The guard compared whole quoted strings, called the old quote lost, and put the old sentence back. Asking the writer to distill again would leave that guard unchanged. Removing the guard would again lose evidence.

The candidate repair instead recognizes literal words inside one longer quote with the same immediately bound originals. It does not erase spaces or join separate quotations; able to pay is not retained inside unable to pay. Renumbered citations can name the same original, but an ID elsewhere in Sources cannot lend its identity to a different quote. Truly missing introductory prose returns to the introduction, retaining the new lead too.

The tests separate preservation from truth. A matching span does not prove that its new surrounding claim is accurate, and it cannot fix a duplicate the writer already supplied. The original review gate remains. This is an unreleased follow-up to #2687, not a repair of the owner's saved notebook or proof of current model usefulness. Native page and source-entailment acceptance remain separate.

29 October: publication checkpoint

The bounded repair above ships in 1.9.3rc27. Version PR #2700 passed its exact-head checks; the immutable tag then passed the normal release workflow. The public wheel and source archive match the workflow artifacts byte for byte. The frozen offline suite passed 17,655 tests with 336 skipped. These are package and synthetic regression results. No real notebook was rewritten for this trial; real-model usefulness, source truth and native desktop/phone acceptance remain open.