Old words are not an unchanged claim

Repair for #2639 in 1.9.3rc17. This journal records its development; availability requires the matching public PyPI package and GitHub prerelease.

REM protects accepted content while new evidence is added: a small source packet cannot re-prove the whole relationship on every turn. But the source audit used an unsafe shortcut. It removed every space and citation, then checked whether a refuted sentence appeared somewhere in the previous Markdown. That could turn a FAIL into PASS after a condition disappeared, a citation changed originals or a historical statement moved into the current state. Even now here and nowhere looked identical. A heading or code example could supply the matching characters.

Word boundaries alone would leave the conditions and evidence problems. Another prompt would leave the code that overrides the model's verdict untouched. Whole page equality would prevent small updates. The selected boundary is complete unchanged prose context: preserve word separation, title and section placement, and each numbered citation's original identity and relative location. Every candidate occurrence must be covered. Removing or reordering an old neighbouring paragraph or bullet invalidates the section's exemption; qualifications may live outside the same sentence. A whole old block can remain before a new sentence when the earlier context and evidence still match.

Independent AI source review caught two holes in the first implementation. We kept source order but discarded the recorded citation positions, allowing a citation to move to another sentence. We also excluded the main title, allowing a pronoun assertion to inherit protection after the page subject changed. Two review-level tests reproduced these before the correction. They join the initial nine failing controls and the later adjacent-qualification cases.

Protection is not verification. An unchanged finding retained under the existing policy now leaves an insufficient report and a visible deferred count in the promotion record. It does not become a factual pass. A further review found that an inconsistent model response could label itself PASS while listing a deferred finding. Three promotion regressions reproduced this and active findings paired with PASS or INSUFFICIENT. Report normalization now keeps active findings failed and deferred ones insufficient regardless of that headline. Active new failures still reach the ordinary repair and refusal path. Missing or ambiguous source bindings cannot establish continuity. A conservative unmatched rewrite receives ordinary review; there is no fuzzy claim-equivalence model in this change.

These tests exercise report classification and synthetic notebook promotion, not a real model's ability to discover every source error. Syntactic continuity also cannot prove that added surrounding prose has the same meaning or that an old accepted statement is correct. No real notebook was rewritten, no new model or provider run performed, and current native product acceptance remains open.

23 October: published preview

1.9.3rc17 is now a public prerelease with a matching PyPI package. The release workflow passed and verified the public artifacts against its build. This makes the bounded manual trial installable; it does not establish real model usefulness, repair saved pages or complete native product acceptance.