Research report 08: the fact-check audit report of the existing research documents
Source:vignettes/articles/research-08-fact-check-audit.qmd
ℹ️ Document nature: this piece is a research snapshot (archival snapshot; the verification date is in the text). Live decisions are governed by architecture.qmd §0 (D1–D16) and review-00-synthesis.qmd; the method names in the text (e.g. MSU-Extract, MSU-Corr) are the wording of the time, the current naming is in ARCHITECTURE §5; the benchmark metrics are governed by capability-matrix.qmd and ARCHITECTURE §8.
Audit date: 2026-09-28. Every load-bearing claim of docs/ was checked against Crossref, PubMed E-utilities, GitHub raw/API (the current HEAD), crates.io, CRAN, PMC, Europe PMC and the local R/rustc. The auditor may not modify the files — this report is the basis of correction (the corresponding fixes were executed by the mainline).
§1 The discrepancy adjudications (the canonical citation strings, quotable directly)
| # | The item | The adjudication |
|---|---|---|
| 1 | The Medo DOI |
10.1038/s41467-024-53711-6 is correct; -53574-6 is a Crossref 404, non-existent. Nat Commun 15:9467 (2024); PMID 39487150, PMC11530434. The citation string of the original 02 report was corrected. |
| 2 | The Degasperi article number |
eabl9283 is correct; abl9111 does not exist. Science 376(6591):eabl9283, 2022-04-22; PMID 35949260. |
| 3 | Jiang/Wu/Rozen | The canonical citation = Brief Bioinform 26(1):bbaf042, 2025 (epub 2024-11-22); doi:10.1093/bib/bbaf042; PMID 39910776, PMC11798676. PMID 39487150 belongs to Medo. The recommendation is to uniformly label it 2025 (epub 2024). |
| 4 | MuSiCal | Verified: Nat Genet 56:541–552 (2024), doi:10.1038/s41588-024-01659-0 (PMC10937379). |
| 5 | PCAWG 2020 | Alexandrov, Nature 578(7793):94–101, doi:10.1038/s41586-020-1943-3 (PMID 32025018). Note 10.1038/s41586-020-2623-x is a 404 — wherever it appears it must be replaced. Li Y, Nature 578:112–121, doi:10.1038/s41586-019-1913-9 verified (PMID 32025012). |
| 6 | Everall | Officially published: Nat Genet 58(3):570–581, 2026 Mar (epub 2026-02-13), doi:10.1038/s41588-025-02474-x, PMID 41688639. The abstract confirms 10,983 patients, 16 cancer types, 134 signatures, 26 new COSMIC entries, the SV reference set. |
| 7 | Steele | Verified: Nature 606(7916):984–991 (2022), doi:10.1038/s41586-022-04738-6; 9,873 TCGA, 21 CN signatures. |
§2 The per-claim verification table (the summary)
| The claim | The verdict | The remarks/corrections |
|---|---|---|
SPMG: scripts/MutationMatrixGenerator.py ~4,057 lines; the indel ~:1169–2100; BSD-2 |
verified | exactly 4,057 lines; catalogue_generator_INDEL_single :1169, forward_homology :1486–1518, dinuc_tsb_ref :484; BSD-2-Clause (the Bergstrom/Alexandrov Lab); pushed 2026-09-22 |
MuSiCal: nnls_likelihood_bidirectional, LTH 0.001, SIGS_ASSOCIATED without SBS7, no MWPruning/MDL |
verified |
nnls_sparse.py:201 (thresh_backward=0.001, per_trial=True); the relaxed one at :434 (:410/:461 are the LOO blocks, the line numbers drift); utils.py:72 the three groups, no SBS7; MWPruning/MDL 0 hits |
| mSigAct: the χ²₁ LRT; nbinom.size 8/50/100; GPL-3 | verified (branch-conditioned) | the default branch v3.0.3-branch; the old master behaves differently (the multinom default, nbinom.size=5) — replicate against the default branch |
| Sonata: exists, MIT, the model form of cornet.py | verified |
compute_exposures = exp(α + β + sig_emb @ sample_emb.T).T matches exactly |
| The COSMIC v3.6 mirror: the file-name patterns, RNA-SBS GRCh37 only, SV GRCh38 only, the column counts | verified + corrected | ID and CN are also GRCh37 only; the mouse/rat have SBS+DBS only. The column counts: SBS 102 columns (Type+101), DBS 23, ID 26, CN 27, RNA-SBS 6, SV 13 |
| The CN48/SV32 labels | verified | all the 48/32 labels match exactly |
| SUITOR: PLoS CB 18(4):e1009309 2022; the entry-wise CV; ECM≡KL-MU; 300 initialisations | verified | the original-paper original-sentence quotations confirm it |
sigminer: CRAN “MIT + file LICENSE”; extdata v3.1; py_path hard-coded; the Broad BSD-3 header |
verified |
R/sig_extract.R:38; 34 extdata files (no SIGNAL-named file) |
| rust-NMF: GPL≥2; 76×/~215–280× | verified | the LICENSE file is authoritative; the README’s 16-thread table agrees with our numbers |
| the extendr/savvy/rextendr versions | corrected | extendr-api latest = 0.9.0 (2026-04-16 unyanked); 0.8.2 is the last of the 0.8 line. savvy 0.11.0 ✓; rextendr 0.5.0 ✓ |
| The CRAN-policy quotations | verified + a wording caution | the original sentences confirmed; there is no literal “networking forbidden at install time” clause — the actual content is “external resources minimized/https/graceful degradation/precompiled downloads as a last resort, needing CRAN’s consent”. The conclusion (vendor everything) is unchanged |
| the twobit crate 0.2.2, MIT, jbethune | verified | updated 2026-03-27 |
| The HRDetect coefficients | verified | the means/sds/weights/intercept −3.364 all exact |
| A spot check of 5 PMIDs | verified 5/5 | 35949260/31278357/42351273/41688639/35668106 all correct |
| The local environment | verified | R 4.5.2; rustc 1.97.1; rextendr 0.5.0; sigminer 2.3.1; MutationalPatterns not installed |
| (the extra check) the SPA default cosmic_version=3.6, the forced background signatures, the exclusion dictionary :376–410, the probabilities :1694 | verified | decomposition.py:257/:272/:376–410 (14 subgroups incl. Treatment_signatures, which the earlier summaries missed) |
| (the extra check) the Cornet preprint | verified | Europe PMC PPR1324264; 2026-09-21; doi:10.64898/2026.09.14.751548 |
§3 The other findings
- The 02 report’s Medo DOI error (the only hard flaw) — corrected.
- The 00/04 extendr “latest version” was stale (0.9.0) — corrected.
- The per-build reference-library availability was over-generalized (the 05 §1, ARCHITECTURE §4.2 and ROADMAP M4 “five classes × 7 builds” does not hold) — the 05 was corrected and the bundling plan adjusted in the architecture v2 (the missing builds are self-derived or converted across builds by labels).
- The CRAN-network wording softened — corrected.
- The Jiang year unified to 2025 (epub 2024) — the 00 was corrected.
- The mSigAct branch footnote — the 04 was corrected.
- Details: the 02’s exclusion lists missed the
Treatment_signaturessubgroup; the 00 mentioned sigminer’s extdata having SIGNAL, unconfirmed (the 34 files have no SIGNAL-named one). - Not rechecked (the low-risk annotations): the Wu 2022 DOI form, the BayesPowerNMF/BASCULE DOIs, the remaining ~60 PMIDs of the 03 report (the spot check 5/5 passed), the DBS78 group-order details, SPMG’s “~13,700 LOC” package total.
§4 The overall conclusion
The archive is exceptionally accurate at the source-engineering layer (every sampled engineering claim, incl. the file line numbers, defaults, coefficients and column counts, matched the current upstream bit-for-bit). The required corrections before use as an implementation reference total 4 groups (§3.1–3.4), all executed. The rest can be trusted as-is.