Standalone study repository, migrated from the model-selection article series.
- Latest detailed article with corrected fusion results — preserved before the shorter blog replacement
- Original detailed article
- Executable companion
- Protocol status and index, including the frozen historical protocol, the protocol with amendment A1 and the freeze receipt
- Review fixes and known limitations
- Source provenance
- Manuscript (PDF) and LaTeX source
The original measurements (runs of 3–5 October 2026) and code are imported. The study followed a protocol frozen on 2026-10-04 before any test-fold score existed; that protocol is archived as historical evidence under protocol/historical/. It is not an approval under this repository's harness.
A cached-score correction on 6 October 2026 recomputed fusion separately within each candidate pool and recalibrated its thresholds on validation. The selected mass weight (0.55), official-deduplicated fusion Recall@5 (64.8%), all single-method results and all paired contrasts are unchanged. Expanded-pool fusion Recall@5 changes from 43.9% to 42.7%. At the validation-fitted false-nomination threshold, fusion covers 17.4% of target-present queries, with 88.4% Recall@5 among nominated queries and 10.4% false nominations on target-removed queries.
The corrected evidence index binds numerical tables, comparisons and receipts to their checksums. The paper uses these corrected outputs and unchanged historical supplementary evidence. Original archives and articles remain intact. See correction protocol and review fixes.
Independent training reproduction remains pending. This correction reuses saved component scores, nonfusion ranks and candidate pools; it does not retrain models, regenerate pools or establish original data authenticity. The 28 cached input files (401.7 MB) are locally available and hash-identified, but are not distributed in Git. Their manifest source URLs document provenance; they are not direct downloads of those files.
| Target | What it does | What it does not do |
|---|---|---|
make verify |
Inexpensive checks: SHA-256 of imported files and archive members against recorded digests, extraction of evidence from the archive, and companion regression tests on small synthetic inputs | Does not download data, train, score or recompute any benchmark number. Passing it is not scientific verification. |
make paper-imported |
Validates historical and corrected saved evidence, formats tables into paper/generated/ and compiles paper/main.tex (system Python, local TeX Live, rsvg-convert) |
Runs no experiment, data download, model inference or environment creation. A successful build is not reproduction. |
make corrected-analysis |
Recalculates fusion and analysis from the hash-verified cached inputs | No retraining or new pools. |
make reproduce-corrected |
Repeats the cached-score correction from inputs supplied through MSMS_CACHED_INPUT_ROOT, compares deterministic outputs and rebuilds the paper |
Requires the original cached files and local TeX tools; does not reproduce training. |
make smoke, make reproduce |
Intentionally disabled. They stop with an error explaining that no approved, implemented reproduction plan exists. | They do not run the historical workflow. |
Validation status is recorded in docs/review-fixes.md. The historical benchmark workflow is documented command by command in companion/REPRODUCE.md. It needs external data and hours of training, and it is not wired to any make target.
The repository will hold the detailed methods and paper; the blog will provide a shorter accessible explanation. Original third-party licences and notices remain applicable; no blanket relicensing is applied.