# Independent VAST-TOOLS referee verdicts Pinned commit: `5598e1f1b0067611dc9e5abb285b4eacaf543ee3` (VERSION `2.5.1`). Checked 2026-11-06. | ID | Verdict | My repro path | Effect on users | | --- | --- | --- | --- | | VT-001 | **CONFIRMED** | [repro/VT-001/](repro/VT-001/) | A positive count accompanying rounded cRPKM `0.00` becomes zero in all three count exports; independently supplied count 5 became 0. | | VT-002 | **CONFIRMED** | [repro/VT-002/](repro/VT-002/) | Renaming the same two samples changes merged MIC PSI/counts from missing to numeric: `NA` versus PSI 75 and counts 4/12. | | VT-003 | **CONFIRMED** | [repro/VT-003/](repro/VT-003/) | Column order overrides requested biological pairing, suppressing an eligible event and scoring an event with no eligible named pair. | | VT-004 | **CONFIRMED** | [repro/VT-004/](repro/VT-004/) | A one-event processing chunk aborts `diff`; a final singleton leaves an incomplete result table. | | VT-005 | **CONFIRMED-MINOR** | [repro/VT-005/](repro/VT-005/) | Group identifier `0` causes an existing sample to be reported absent; renaming that group avoids the fatal error. | | VT-006 | **CONFIRMED-MINOR** | [repro/VT-006/](repro/VT-006/) | Zero retained events cause a fatal summary division by zero, although the correct header-only result table has already been written. | Ranked by effect on real users: 1. **VT-003 — incorrect biological pairing.** Silent inclusion/exclusion of differential-splicing results under a documented analysis mode; no unusual labels or data values needed. 2. **VT-001 — raw-count loss.** Silent alteration of exported observed counts. Restricted to positive counts with rounded-zero expression; downstream count analyses may use these exports. The default `compare_expr` expression threshold normally excludes such low-expression rows. 3. **VT-004 — incomplete differential-splicing output.** A normal run can halt after many successfully written events. With default chunks of 5000, source inspection gives the condition `event_count % 5000 == 1`; default singleton and a three-event/two-line-chunk run reproduce it. 4. **VT-002 — order-dependent microexon missingness.** Incorrect numerical estimates can enter subsequent PSI/quality calculations. Mixed missing/numeric contributions require different mappability; the demonstrated strand-protocol mixture is supported but explicitly discouraged. Frequency in users' datasets is unknown. 5. **VT-006 — empty-filter workflow failure.** Small subsets or strict filters can legitimately retain nothing. The fatal exit can interrupt automation, but it does not corrupt the already-written event table. VT-005 is a lower-priority label-specific failure. ## Reproduction and independent answers Only the hunter's `BUGS.md` was read. No hunter repro scripts, output artifacts, or patches were opened, executed, imported, or copied. Fixtures and calculations were independently authored from upstream source/formats. Each bug has its own fresh remote clone at the pin; all six tracked source trees remain unchanged. [run_referee.py](repro/run_referee.py) creates the minimal cases; [reachability.py](repro/reachability.py) exercises public commands and produces MIC inputs with the actual upstream analyzer. Every command saves its arguments, working directory, exit status, stdout, and stderr beside its input. Numerical answers are in each bug's `independent-answer.json`. **VT-001.** `MakeTableRPKMs.pl:79–87,107–114` initializes counts to zero and only copies them for nonzero expression. In my input, LOW has cRPKM `0.00`, count 5; HIGH has `10.00`, count 200, and ZERO has `0.00`, count 0. Unpatched outputs are 0/200/0 instead of the correct 5/200/0. All three exports fail; the documented public `combine --exprONLY --C --TPM` route also exits 0 with the wrong counts. The correct raw count is the observed integer, independent of expression rounding. Independently, `5 * 10^9 / (100000 * 20000000) = 0.0025`, which rounds to `0.00`; upstream `expr_RPKM.pl:143` uses this two-decimal format. Count 200 and genuine zero remain correct. Reachability: [README expression/count instructions](https://github.com/vastgroup/vast-tools/blob/5598e1f1b0067611dc9e5abb285b4eacaf543ee3/README.md#L543-L549), `RunDBS_2.pl:210–214,481`. This is not an intentional expression filter: the input explicitly contains a raw count and no count-filtering option was requested. **VT-002.** `MergeOutputs.pl:451–452` adds numeric values to an accumulator even when it already contains `NA`. My all-missing plus 4-exclusion/12-inclusion example yields numeric PSI `100*12/(4+12)=75` only when the missing file sorts first. Reversing sample names yields all `NA`. The numeric-only control correctly gives 8/24 and PSI 75. The producer case independently constructs mappability files and runs unmodified `Analyze_MIC.pl`: strand-unspecific zero effective lengths emit all missing values, while strand-specific effective lengths 35 with 12 inclusion/4 exclusion reads emit the numeric row, both at read length 50. Merging these actual producer outputs reproduces the name-order dependence. A Perl numeric-coercion warning is emitted, but merge exits 0. Unknown contributions cannot establish the complete pooled counts; if the intended policy were instead to discard unmappable contributions, it must still be invariant to names. The existing last-NA and MULTI behavior favors preserving missingness. Reachability: [README strand-specific merging](https://github.com/vastgroup/vast-tools/blob/5598e1f1b0067611dc9e5abb285b4eacaf543ee3/README.md#L369-L372) permits the mixed protocol but warns against it; `Analyze_MIC.pl:123` emits missing counts and `Add_to_MIC.pl:49–96` carries the merged values into PSI/quality output. No real-sample frequency or downstream biological false-positive rate is claimed. **VT-003.** `R/vastdiff.R:176–177` indexes each requested set in table order. Requested pairs are preX/postX and preY/postY; the permuted table lists preX, preY, postY, postX. Event 1 has 45/5 reads in preX and 5/45 in postX, and zero coverage in both Y samples: exactly one eligible named pair. Event 2 has the same preX counts, but coverage only in postY: zero eligible named pairs. The permuted unpatched run reports event 1 all NA and event 2 dPsi `0.779547`, MV `0.65`. Reordering only the table columns restores event 1 and removes event 2, with the same numerical result for the eligible event. Both runs exit 0. Independent named-pair counting establishes the decisive wrongness without simulation. A separate [base-R oracle](repro/posterior_oracle.R), using only beta quantiles and numerical integration, gives marginal medians `0.8895478374` and `0.1104521626`, their difference `0.7790956747`, and 0.01-grid MV `0.65` at probability 0.95. The small median discrepancy is expected with default Monte Carlo size 1000. Reachability: [README paired analysis](https://github.com/vastgroup/vast-tools/blob/5598e1f1b0067611dc9e5abb285b4eacaf543ee3/README.md#L480-L485) and the current help at `vastdiff.R:57–58` explicitly associate the pairs by argument order. Both table orders are accepted by the actual input checker; the public wrapper reproduces the failure. **VT-004.** `R/vastdiff.R:226–235` simplifies singleton results, then indexes them in two dimensions. One event with 30/10 reads versus 10/30 should produce one completed event: both samples have 40 reads, above the default minimum of 10; posterior means are independently 31/42 and 11/42. Instead it exits 1 at line 234: `totalFirst[, x]: incorrect number of dimensions`, leaving only the header. Three such events with `-n 2` also exit 1, leaving two results. Two-event and default-chunk three-event controls exit 0 with all events. The public one-event command also fails with default PDF output enabled. Reachability: `--nLines` is documented in help (`vastdiff.R:97–99`) and defaults to 5000; no documented minimum event count excludes a singleton. This is a crash, not statistical rejection. The modulo-5000 condition is inferred from the chunk loop; a 5001-event run was not executed. **VT-005.** `Filter_and_clean_Combine.pl:221` tests the group string for truth. My config assigns sick to `0` and well to `1`; both names are present. The command exits 255 with `Sample (sick) from 0 not found`. Renaming the identifiers to control/treated succeeds, retaining the expected PSI 20/80 event. Independently, its sample SD is `42.4264068712`, above default 5, and each group has the requested one covered sample. Reachability: [README groups config](https://github.com/vastgroup/vast-tools/blob/5598e1f1b0067611dc9e5abb285b4eacaf543ee3/README.md#L592-L601) and help at `Filter_and_clean_Combine.pl:120–125` specify a group identifier with no prohibition on numeric labels. The error is real but obvious and avoidable by renaming, hence minor. **VT-006.** `Filter_and_clean_Combine.pl:520` divides by retained-event count after closing the result table. My two covered PSI 40/40 samples have independently computed SD 0, below default minimum 5; exactly zero events should survive. Unpatched `tidy --min_N 2` writes the correct header-only table, prints a zero-event total, then exits 255 with `Illegal division by zero`. The 20/80 control succeeds and retains one event. Reachability: [README's documented filtering](https://github.com/vastgroup/vast-tools/blob/5598e1f1b0067611dc9e5abb285b4eacaf543ee3/README.md#L590-L592) allows ordinary empty selections. A missing-event percentage is undefined on zero events; the selected event set is already correct, so this is a minor completion/reporting defect. ## Already reported or fixed? Independent read-only GitHub API checks used `gh api`: all 134 issue/PR entries (8 open), all 345 issue comments, 1328 master-history commit records, the recent-commit query since 2026-08-06, all 15 branch heads, and releases/tags. Raw responses are in [upstream-context/](upstream-context/). Matching source excerpts by branch/release are in [branch-source-review.json](upstream-context/branch-source-review.json). No matching public report or branch fix was found for any of these six specific failures. This is a bounded search result, not proof that the authors have never encountered them. Current master is still the pin; its only commit in the last three months adds Osa mappings. The current GitHub release is v2.5.1 (2020-06-26), tag commit `498bbecb2fcef3e98beb9d6a6cc7648a529453e5`, distinct from the pin despite the identical VERSION. The relevant vulnerable constructs also appear in that tag where the feature exists; standalone count/TPM exports were added later. Branch/tag review was source inspection, not execution of those versions. Per-claim distinctions from nearby upstream work: - VT-001: [#92](https://github.com/vastgroup/vast-tools/issues/92) explains generating counts with `-C`; it does not report rounded-zero count loss. No corresponding rounding/count-copy fix was found. - VT-002: [#68](https://github.com/vastgroup/vast-tools/issues/68) concerns grouping/configuration, and [#87](https://github.com/vastgroup/vast-tools/issues/87) concerns incomplete align output and EEJ memory/format problems. Neither describes name-order-dependent MIC missingness. The accumulator's old “check with Manu” comment signals unresolved policy, not a fix for order dependence. - VT-003: [#117](https://github.com/vastgroup/vast-tools/pull/117) and commit `6efb39c` fix other paired-mode problems; the `which` indexing survives unchanged. [#32](https://github.com/vastgroup/vast-tools/issues/32) fixed output-column reordering in `Add_to_FULL.pl`, a different module. No matching biological-pair order report was found. - VT-004: [#48](https://github.com/vastgroup/vast-tools/issues/48), [#54](https://github.com/vastgroup/vast-tools/issues/54), and [#60](https://github.com/vastgroup/vast-tools/issues/60) describe missing-replicate list indexing or plotting failures, not singleton chunk dimensions. No singleton fix was found. - VT-005: no matching numeric-zero group-label issue or existence-test fix was found; the truth test persists in every reviewed branch containing grouped tidy. - VT-006: commit `bf7422ddd` (2021-04-01) suppresses undefined per-type tally warnings in tidy, but does not guard the zero denominator. No matching empty-result fatal-error report or fix was found. Distribution check: no corresponding VAST-TOOLS Bioconda package/recipe was found (direct package/recipe lookups returned 404; recipe code searches found only hg19/hg38 annotation packages mentioning the tool; Anaconda name search returned unrelated packages). This command-line Perl/R repository is not itself a CRAN or PyPI package; similarly named Python astronomy tools are unrelated. The GitHub release and current source are therefore the relevant releases checked here. **Did this referee job find a new research result? No.** It independently confirms six software defects claimed by the hunter; no printed scientific benchmark, stated paper problem, or new-data prediction was tested. No author contact, issue, PR, fork, or patch publication occurred. Predictive-score baselines/SBS are not applicable to these deterministic correctness cases and posterior calculations.