Optional active AI-reviewer adaptation
- Adds a user-selected, score-aware manuscript adaptation workflow under the existing
paper_audittool. - Binds current official venue guidance and freshly verified primary research evidence.
- Evaluates the unchanged baseline and complete candidates with the same panel of at least two reviewer models.
- Selects by normalized panel mean minus a cross-model variance penalty; the baseline may win.
- Freezes citations, numbers, formulas, and critical limitations/ethics/risk/negative-result paragraphs.
- Keeps the defensive
ai_robustnessaudit as a separate mode.
Evidence and implementation details: P18 verification.
Install only research-guard-windows-x64-modular.zip and verify it against SHA256SUMS.txt.