NEWS

---

inferguidR 0.0.0.9000

• Standardize the named v5 package logo across README, documentation pages and available application headers, with matching light and dark artwork.

Development version; numerical and release qualification remain incomplete.

• Numbers typed into the calculator and the workbench, and numbers read back from canonical JSON, are converted with correct rounding on every platform. R's own conversion missed the nearest double for many inputs with 16 or more significant digits or a large exponent where long double has no extra precision (macOS on Apple silicon), which changed calculator results and made retained report hashes fail there.
• The unconstrained theoretical Youden optimum reports the analytic stationary point instead of a numerical estimate next to it.
• Exact unconditional pooled, continuity-corrected pooled and Mantel-Haenszel calculations now preserve rational equality classes at discrete critical boundaries. Some criterion results can change where the previous calculation split mathematically tied tables through floating-point rounding. Exact compromise remains guarded pending an attained-size solver contract.
• advisor_power_spec() creates explicit same-test planning specifications for mapped Boschloo recommendations. Variant, direction, probabilities, allocation policy and numerical controls are retained. Rules version 2026-10-26.2 requires explicit replay of older saved sessions; existing recommendation documents remain readable.
• Classifier probability diagnostics now expose calibration and net-benefit panels, explicit decision probabilities and partial-AUC settings. Changed or invalid evaluation settings make old results ineligible for report export.
• The planner retains method-specific input drafts when dynamic controls are rebuilt, including invalid and explicitly empty entries with their input locale. Methods without effect helpers no longer raise an indexing error during a method transition. Responsive plot labels and legends remain visible at narrow browser widths.
• create_report_snapshot(include_probability_summaries = TRUE) can retain explicitly selected aggregate probability diagnostics in report schema 0.4. Default exports continue to omit these summaries and raw scored rows. Boschloo power reports now retain both conditional count lattices, including their finite support, rejection membership and probability masses. Import and export use the frozen grids without rerunning the region solver.
• Report exports stage and validate a complete sibling file before replacing the destination with a checked rename. Failed rendering or commit leaves an existing report intact; this follows filesystem rename semantics rather than promising crash durability or isolation from concurrent writers.
• Canonical JSON string handling avoids repeated scans while preserving the serialized bytes and hashes. Report normalization explicitly preserves NULL values across supported R versions.
• Public numerical campaigns, method/mode/option inventories, installed release preflights and source/resource audits provide reproducible scoped evidence. A bilingual statistical glossary explains interpretation and uncertainty.
• Platform workflows use a verified shared source archive, fresh candidate libraries and a separate minimum-R dependency profile. Passing these checks does not establish complete external-reference parity or release approval.