Two Models That Agree Beat the Best of Them Alone: Parameter-Free Connected-Component Consensus That Beats the Baseline under the Official BraTS-2023 Metrics

Version 11 (2026-10-05) — link and provenance correction of the v10 record; no result changes. The archive is byte-for-byte the published v10 (record 22904810, zip md5 b555d816e488fbeecbc705fcc5b084d0) except for the eleven files listed under 'What changed in v11', and that claim is machine-checked: check_delta_v11.py compares the staged bundle to the v10 inventory file by file and fails on any difference that is not declared. It carries the result-first title 'Two Models That Agree Beat the Best of Them Alone: Parameter-Free Connected-Component Consensus That Beats the Baseline under the Official BraTS-2023 Metrics', both authors (Guillaume Cassez first, Stanislas Larnier second), full English/French manuscript parity (30 sections, 9 figures, identical table counts), and the multi-seed robustness tables of §4.6 reported from the complete, fragment-cleaned evaluation run (n = 720 valid pairs; data/multiseed/stats.json, md5 fe5d513e563a0bc946db983c93e8013f — the numbers published since v7), cross-checked cell by cell by scripts/check_5_6_provenance.py. Three corrections were measured on 2026-10-05. (1) Repository links: the two Hugging Face model cards cited by both manuscripts now resolve to the guillaume-cassez namespace; the namespace they used to live under has been withdrawn and its models migrated with their content verified file by file. Twelve dead references in the v10 archive are corrected (four per manuscript, two README badges per language), and a fail-closed lock in the release tooling blocks the zip if any file of the bundle still points at the withdrawn namespace; the withdrawn identifier itself is spelled out nowhere in this archive, so the criterion stays binary and re-checkable by anyone with a plain text search. The generator that produces the provenance note is itself shipped, so that the corrected figure can be audited from the archive. (2) Dataset provenance, one number: analysis/DATASET_EXCLUSIONS.md stated that the patient-level exclusion caught '0 cases of the unified set'; the true value is 50, and the CSV shipped in the same archive already listed those 50 cases (entree_exacte = non). The 0 was not a measurement but a silent fallback of the generator when the dataset volume was not mounted; the generator now computes the figure twice and refuses to write on divergence or on an impossible scan. The 1251 → 1196 study set, the 55 dropped cases and every published number are unchanged. (3) Build recipe: header.tex loads booktabs defensively, because a build whose AST no longer contains a table otherwise dies on \toprule with 'Undefined control sequence'. The archive bundles the FR + EN manuscripts (markdown and PDF, 25 and 23 pages), the 9 figures, the analysis and data artifacts, the demonstration patients C1-C6 re-selected on 2026-08-31, and build.sh, a self-contained pandoc + XeLaTeX recipe that recompiles both PDFs from the shipped sources.