Skip to contents

The five-expert review synthesis adjudication (Synthesis, 2026-09-28)

Source: vignettes/articles/review-00-synthesis.qmd

The five independent reviews are in this directory, 01–05. This document is the editor-in-chief’s adoption adjudication — every review opinion is sorted into adopt (fix the documents at once) / adopt (schedule it into the roadmap) / partly adopt / do not adopt (the reason recorded). The review originals are preserved unchanged; this document is the single effective caliber; where it conflicts with a review, this document prevails.

1. The one-sentence conclusions of the five reviews

The review The conclusion
01 the computational biologist conditional trust — the three flagship claims of the science layer (MSU-Fit coverage, MSU-Corr, the multi-evidence K) would be pierced by a single page; the HRD/RNA workflows as they stand are clinically dangerous and dishonest; the M0–M8 timelines are unrealistic
02 the Rust/R expert conditional go — the contracts of the Rust runtime layer (the RNG version, concurrency/fork, panics, the FFI boundary, the distribution budget) are almost wholly missing and must be written up at M0
03 the ML/statistician major revision — the statistical claims are generally 1–2 grades stronger than the protocols support; downgrade “calibrated / distribution-free / lossless / first” to the level the protocols actually support
04 the software-development expert conditional pass — the “paper wishlist” is not a delivery plan; there is no v1 cut line, the bit-for-bit acceptance collides with the clean-room policy, as do the parallelization and the determinism
05 the Nature editor the current positioning = Genome Biology/NC, not Nature Methods — the only statistical contribution big enough for the top journal is “the calibrated statistics of exposure inference”; cut the 8 MSU-* down to 2–3; the three-act structure needs re-ordering

2. The cross-corroboration (≥2 reviewers independently raised it = must change)

  1. The CI/calibration claims lack an estimand (01#1, 03#2, 05#2/5/7)
  2. The conformal “distribution-free” is mathematically untenable under heterogeneous N (01#11d, 03#1)
  3. MSU-Corr will be read as a Cornet derivative: the gauge is unidentifiable, the baseline is missing, the ablation is missing, the m protocol is missing (01#7, 03#3, 05#2/risk 1)
  4. The determinism claims lack thread-count invariance (02#3, 04#3)
  5. “Bit-for-bit” conflates the discrete semantics with the floating-point kernels, and collides with the clean-room strategy (02#8, 04#2)
  6. The timelines are underestimated by ~3×, no minimal v1 (01’s annex note, 04#1, 05 risk 3)
  7. The handles/.state are complexity paid for a capability that does not exist (02#4, 04#11)
  8. The K selection lacks a default adjudication rule (01#9, 03#4)
  9. The “first/lossless/distribution-free” over-claims (01#11, 03 throughout)
  10. The 8 MSU-* dilute the contribution (04 implicit, 05#2 explicit)

3. The adoption adjudications (fix the documents at once)

# The adjudication The adopted content The landing site
A1 adopt The statistical contract: the three-layer estimand (presence=the LRT χ²₁; magnitude=the support-conditional absolute exposure; zeroing=a decision rule with separately measured composite coverage); the calibration under both the multinomial and NB generative models; the coverage as a curve of N×Shannon; the Fisher intervals interior-only; “boundary parameters get tests, not intervals” ARCH §7
A2 adopt The conformal down-scoping: Mondrian by TMB bin + a studentized score; the claim rewritten as the per-stratum empirical coverage; the quantile tables shipped with the package ARCH §5/§7
A3 adopt The method-naming contraction: keep only the MSU-Fit / MSU-LowCount / MSU-Infer named methods (the headline cluster); the extraction pipeline / the correlation engine / the damage tempo / the tensor / the catalog IR all demote to framework components (losing the MSU- prefix) ARCH §5
A4 adopt The correlation-engine downgrade: pre-registered falsiable claims (simultaneously beating MvNMF+correlated refit and the original Cornet over r∈{0.8,0.9,0.95}); the gauge-fixed reporting (the embedding vectors are forbidden to interpret); m<K via held-out; the volume regularization and the embedding separated by ablation; the five modalities reduced to the SBS+DBS demonstration ARCH §5/ROADMAP v1.x (the deferred items)
A5 adopt The stateless FFI: delete the handle registry and the .state property; the S7 embedded object becomes a snapshot summary (the dimensions + the channel-table hash + the refdb version); keep of the caugi pattern only the read-only aesthetics ARCH §2/§3
A6 adopt The self-developed RNG: the PCG64 + SplitMix64 stream layout (written into the paper’s Methods), independent of rand’s RNG surface (which carries no across-version guarantee); the across-version golden tests ARCH §2
A7 adopt The concurrency contract: parallelism only across independent units; a fixed reduction order within a unit; the thread-invariance identical() tests; the per-call ThreadPool (abandoning the global pool); the msuiter.threads option (2 during checks); Makevars -j2; R_CheckUserInterrupt at the main-thread chunk boundaries + the workers’ AtomicBool ARCH §2/§9
A8 adopt The two-layer acceptance terminology: the semantic bit-for-bit (the channel arithmetic / the protocol steps / the tie rules / the default parameters — the integer and label goldens, a hard acceptance) vs the numerical protocol equivalence (the floating-point kernels — fixed seed + tolerance); the R NMF brunet cross-check = a one-off fixed-platform experiment, not into the CI gate ARCH §7/ROADMAP
A9 adopt The benchmarking upgrade: add purity/caller/class-specific count intervals (SV 10–10³, CN 10–10⁴) / the hypermutant arm / the joint adversarial cell / the off-catalog signature cell; the three-layer calibration-selection-evaluation scenario isolation (different generators); the primary metric = the Hungarian one-to-one (the greedy + the threshold sweep as an appendix, splitting penalised); the protocol frozen and shipped with the package ARCH §8
A10 adopt The K adjudication rule: the hierarchical decision (SUITOR argmin → the stability veto → the Wilcoxon downgraded to diagnostics); the split/merge semantics of the Hungarian unmatched; the meta-rule evaluated over the full grid ARCH §7
A11 adopt The catalog-completeness contract: the MNV 2–5/>5, complex/double-indel routing table + the skip ledger; the DBS reconnection algorithm for split VCFs; the MNV golden fixtures; the explicit routing of complex events ARCH §7/ROADMAP M1s/M1c
A12 adopt The workflow-scope safety: the HRDetect coefficient mode WGS-only + the allele-specific-CN-caller whitelist + WES marked experimental; the RNA workflow’s minimal caller requirements + the pseudogene panel + the matched-DNA-validation estimand + the research-use annotation ARCH §7/ROADMAP M6s
A13 adopt The refdb governance: the precedence (by default the bundled one prevails, the provenance pins the version); the schema_version forward rejection; updates pinned to a commit SHA + a burned-in manifest + fail-closed; the update tests fixture-ized; v1 bundles only the human SBS/DBS/ID; a missing build = a build_independent: true declaration (a relabel, not derived) + an honest notice at query time ARCH §6
A14 adopt The API freezing: method = ms_nmf(...) as a first-class S7 object + the string sugar; the custom generics uniformly ms_-prefixed (the hardhat naming stays a documentation concept only); the new §3.5 Error & lifecycle (the rlang class/i-j-c/bare stop forbidden/the lifecycle policy/the ffi-surface.md freezing procedure) ARCH §3
A15 adopt The CI completion: the matrix adds R 4.3 (Linux at least) + devel (allowed to fail); the valgrind/ASAN scheduled job; the CI benchmarking gate downgraded to a sanity bound (<5s), the speedup-ratio acceptance moved into the pinned-hardware bench; dependabot + cargo audit; codecov (R≥85%, Rust reports without a gate); the fixture refresh script + the PR-template check item; the rhub/–as-cran pre-submission ARCH §9
A16 adopt The dual-track roadmap: v0.1 (~6 weeks: SBS96/192/384/1536+DBS78, KL-NMF+NNLS, a single method, CI, r-universe) → v0.2 (~12 weeks: +ID83, the K selection, the three fitting methods, bootstrap CIs, refdb v1, 6 viz figures) → v1.0 (~5–6 months: the MSU-Fit calibration, the HRD report, pkgdown); ID89/Infer/Tempo/Tensor/Lda/Supervised/CN-SV fitting/RNA/therapy deferred to v1.x; M8 split into M8a/M8b ROADMAP v3
A17 adopt The governance: the registry gains the engine_version/contract_version/certified tri-states; the non-certified engines get a benchmarking warning + a trust-boundary statement; the monthly reverse-dependency matrix; CONTRIBUTING/CoC/templates ARCH §3.4
A18 adopt The honest hypermutator naming: ms_stratify_hypermutants() (classify+exclude+forced refit, not “normalization”); common/rare promoted into the atlas workflow; the Cornet 10–40× recalibration demoted to a correlation-engine-internal hyperparameter (sweep-verified + de-rescaled + forbidden on the NMF path) ARCH §5/§7
A19 adopt The paper repositioning: the headline = “Calibrated inference of mutational signature exposures” (Framing A) + the ecosystem support (B); the two-week pilot coverage experiment (2 tools × a TMB grid) front-loaded before any narrative writing as a kill-criterion gate; the case studies upgraded by three paths (treatment records/RNA-only/low-count panels) needing cohort collaborators; all seven kill criteria accepted ARCH §10/ROADMAP the G gates
A20 adopt The rest: β=1 as the default is the multinomial MLE (delete the “three-noise” marketing or provide the protocol); the zero-pseudo-count policy; the SUITOR fold/imputation semantics replicated bit-for-bit (the semantic layer); hurdle/structural zeros for the DM; the twobit over Cursor<Mmap> ADR; the column-major+NA hard FFI contract; panic=Result with the boundary converting to error; the workspace dependency-direction assertions + forbid(unsafe_code); the MSU↔︎function mapping table + ms_sitrep(); the docs-sync CI; the ADRs split into their own directory everywhere

4. Partly adopt / do not adopt

The review opinion The adjudication The reason
04#1 suggests deferring the SUITOR CV to v0.2 optional partly adopt the MSU-Fit headline leans on the calibration narrative but the K-selection rules lean on the CV — v0.1 has no CV, v0.2 adds it as a switch (agreeing with 04); but the K-adjudication-rule documentation is written in full once at v0.2
05 “name only MSU-Fit/LowCount/Infer” adopt but keep the component names the components still need engineering names (the extraction pipeline etc.), they just no longer appear as “methods” in the paper abstract
01#3 lifts the MNV fix into M1 adopt the MNV routing is a semantic-correctness issue (without the fix the DBS/SBS catalogs are systematically distorted), not a deferrable feature; it lands at M1s (see A11)
05 “cut MSU-Tensor out of the abstract” adopt + keep the exploration slot the TensorSignatures biological motivation is real; it demotes to exploratory (held-out likelihood for the K + the Kruskal checks), scheduled for v1.x
04#9 the ADRs into their own directory adopt (lightweight) establish vignettes/articles/, migrate D1–D16 in first (16 entries already existed when the adjudication landed, incl. D15/D16); the research documents gain the archival statement
02#11 the self-developed ~300-line twobit reader partly adopt the M0 spike tries both roads (the crate over Cursor<Mmap> vs self-developed), deciding the ADR by the byte-swap/big-endian validation results; the default inclination is the crate

5. The organizing action items of the editorial suggestions (needing the author’s own decisions/driving, not solvable by documents)

  1. The two-week pilot coverage experiment (the G0 gate): the nominal vs actual coverage of 2 competing tools × a TMB grid — verifying the ground of the “the competitors’ CIs are miscalibrated” narrative, before all paper writing.
  2. The statistics co-author (the conformal/CoDA theory parts) + the clinical/cohort collaborator (one of the treatment records / RNA-only / FFPE panel) — locked before M3, otherwise the case studies are destined to be decorative (the editorial kill criterion #4).
  3. The co-maintainer: found before M1s (the bus factor + CRAN + the reviewing triple need; i.e. the G1 gate).
  4. Review the competitive landscape quarterly (the publication dynamics of Cornet/MuSiCal/SPA), triggering the ROADMAP’s kill-criteria reassessment.

6. The documents in effect

  • ARCHITECTURE.md v2.2 (the implementation of this document’s adjudications)
  • ROADMAP.md v3 (the dual track + the G gates)
  • The review originals 01–05 are kept as the evidence and the paper’s rebuttal material