Preview corpus Phase 1 — file bootstrap

Tier-1 Reproducibility Survey

200 bioRxiv genomics preprints (Oct–Dec 2025) were scored for computational reproducibility readiness — whether a third party could, in principle, re-run the analysis from what the authors published. No code was executed in this stage; we read the manuscript and linked artifacts only.

What this is NOT

This is not a reproducibility success rate. Tier-1 produces an upper bound (“ceiling”) on readiness from documentation alone. A paper scored “partially ready” here may still fail when code is actually run (Tier-2).

N = 200 (201 reports − 1 withdrawn preprint) Frame 637 works · 31.4% sample Rubric v0.2.5 · Stage-1 prose-only

At a glance — how many papers at each readiness level?

Each paper is placed in one bucket based on the strongest environment-capture claim in the manuscript. Think of this as “how much reproducibility groundwork the paper itself documents.”

14.5%
29 of 200 papers · 95% CI 10.9–19.0%
Strong environment claim
Authors state the software environment is pinned (container, lockfile, or release snapshot). These papers enter code verification next.
67.0%
134 of 200 papers · 95% CI 61.4–72.1%
Partial documentation
Tool versions are named in the paper, but the full runtime environment is not closed (no reproducible install path).
18.5%
37 of 200 papers · 95% CI 14.5–23.4%
Weak or no pinning
Floating versions, missing pins, or no reproducibility path visible from the manuscript alone.

Readiness ladder (papers)

Three buckets sum to 100%. Most papers (67.0%) name tool versions but do not ship a closed environment.

Documentation depth per paper

152 papers (76.0%) have at least one analysis step partially documented (verdict RI). 48 papers (24.0%) have no step reaching even partial readiness.

The version axis alone caps half of all computational steps. Environment capture is the binding constraint on this corpus.

How the Tier-1 check works

GenomicsChain Tier-1 is a static, human-reviewed assessment aligned with NASEM (2019) reproducibility scope and ACM Artifact Badging v1.1 verdict tiers. Each paper becomes a structured report.json with a sign-off field. The field is in place; none of the published reports is signed yet.

1SamplePre-registered random draw of 200 genomics preprints from 637 works (Oct–Dec 2025).
2DecomposeSplit each paper into computational steps (median 18 per paper) from Methods prose.
3Score 4 axesData, Code, Version (environment), Params — each 0, 1, or 2 per step.
4VerdictWorst axis sets step verdict: RI (partial), NI (gaps), X (blocked). R★ needs code check (Stage 2).
5AggregateCross-paper statistics only — no named per-paper scores on this page until author sign-off.

Stage-2 funnel (what happens next)

  • 29 papers (14.5%) claim a strong enough environment pin to enter code verification.
  • 17 papers will be randomly audited from the remainder (10% stratified sample).
  • Total Stage-2 load: 46 papers — only then can we say “reproduced” vs “not.”

Standards & checkpoints

NASEM 2019 — Reproducibility Same data + code + methods → consistent results. Wet-lab replication and scientific truth are out of scope. Ref. [1]
ACM Artifact Badging v1.1 Verdict tiers map to Available / Functional / Evaluated. “Reproduced” requires execution (Tier-2), not prose alone. Ref. [2]
FAIRsoft scale precedent Large-scale tool metadata exists, but FAIRness ≠ re-execution. We measure installability and pin quality. Ref. [4], [5]
GenomicsChain rubric v0.2.5 Symmetric environment-closure gating, 13-type gap dictionary, two independent survey levers (env lock & in-code DB pin).

Four axes scored on every computational step

AxisScore 0Score 1Score 2 (Stage-1 ceiling = 1 for Version)
Data Missing or inaccessible ID present but gated / incomplete Public or legitimate controlled-access path documented
Code Not shared Reconstructable from description Exact public artifact linked
Version
(environment)
Unknown / unpinned Partial pins (tool names, floating reqs) Closed environment — lockfile or container digest (verified in Stage 2)
Params Not stated Partially stated All parameters for the step documented

Step verdicts (plain language)

RI — Partially ready

Every axis ≥ 1. Artifacts exist but gaps remain. Most common at Stage-1.

NI — Needs investigation

At least one axis is 0 but recoverable in principle (missing link, missing seed, etc.).

X — Blocked

Unrecoverable from published materials (no code, proprietary black box, etc.).

R★ — Declared closure

All axes at 2 in declared environment — but unexecuted. Requires Stage-2 code run; 0 steps at Stage-1.

References

The Tier-1 readiness metric is not a new private score: its scope, verdict tiers and graduated levels are taken from the standards below. Roles are stated so a reader can check what each source does and does not license us to claim.

  1. National Academies of Sciences, Engineering, and Medicine (2019). Reproducibility and Replicability in Science. Washington, DC: The National Academies Press. doi:10.17226/25303 Scope anchor. “Reproducibility” on this page means the NASEM sense — same input data, code and methods → consistent results. Replicability (new data) and scientific validity are out of scope.
  2. Association for Computing Machinery (2020). Artifact Review and Badging — Version 1.1 (24 August 2020). acm.org/publications/policies/artifact-review-and-badging-current Verdict mapping. R★ ≈ Artifacts Available + Evaluated-Functional; RI ≈ Available but not turnkey; NI ≈ below Evaluated; X ≈ not Available. ACM’s “exact reproduction not required, within an acceptable tolerance” is why stochastic steps are scored by distributional match. Version 1.1 matters: pre-2020 definitions had reproducibility and replicability the other way round.
  3. Essawy, B. T., Goodall, J. L., Voce, D., Morsy, M. M., Sadler, J. M., Choi, Y. D., Tarboton, D. G., & Malik, T. (2020). A taxonomy for reproducible and replicable research in environmental modelling. Environmental Modelling & Software, 134, 104753. doi:10.1016/j.envsoft.2020.104753 Graduated tiers (repeatable / runnable / reproducible / replicable). Our RI (“partially ready”) corresponds to their runnable.
  4. Martín del Pico, E., Gelpí, J. L., & Capella-Gutiérrez, S. (2024). FAIRsoft — a practical implementation of FAIR principles for research software. Bioinformatics, 40(8), btae464. doi:10.1093/bioinformatics/btae464 Scale precedent — automated indicator scoring applied to tens of thousands of bioinformatics tool entries, which is why a Tier-1-at-scale survey is feasible. Boundary: FAIRsoft scores software FAIRness, not paper-level re-execution. A high FAIRsoft score is not a readiness score.
  5. Barker, M., Chue Hong, N. P., Katz, D. S., et al. (2022). Introducing the FAIR Principles for research software. Scientific Data, 9, 622. doi:10.1038/s41597-022-01710-x FAIR4RS is a motivation, not a verification rubric — it does not supply the re-execution axis this survey scores.
  6. Nüst, D., & Eglen, S. J. (2021). CODECHECK: an Open Science initiative for the independent execution of computations underlying research articles during peer review to improve reproducibility. F1000Research, 10, 253. doi:10.12688/f1000research.51738.2 Closest existing service, and the model for the “verification, not validation” boundary: we check whether published materials support re-running the analysis, not whether the science is correct. CODECHECK is manual and per-paper; this survey is AI-assisted, not yet human-signed-off, and sample-wide.
  7. Siegel, Z. S., Kapoor, S., Nadgir, N., Stroebl, B., & Narayanan, A. (2024). CORE-Bench: Fostering the credibility of published research through a computational reproducibility agent benchmark. arXiv:2409.11363. arxiv.org/abs/2409.11363 Why human sign-off is mandatory. The best agent in that benchmark solved about a fifth of the hardest tier, so an unsupervised AI score would measure the agent, not the paper.
  8. Plesser, H. E. (2018). Reproducibility vs. replicability: a brief history of a confused terminology. Frontiers in Neuroinformatics, 11, 76. doi:10.3389/fninf.2017.00076 Terminology history — why this page states which definition it uses instead of assuming one.
  9. Kanwal, S., Khan, F. Z., Lonie, A., & Sinnott, R. O. (2017). Investigating reproducibility and tracking provenance — a genomic workflow case study. BMC Bioinformatics, 18, 337. doi:10.1186/s12859-017-1747-0 Depth precedent in genomics — a single workflow examined in detail. This survey trades that depth for breadth (N = 200), which is the gap it is designed to fill.

Standards alignment does not transfer their authority to our numbers. Tier-1 scores are contestable and supersedable; the rubric (v0.2.5), the sampling pre-registration and the 13-type gap dictionary are GenomicsChain’s own and are versioned separately.

Environment capture claims (detail)

Finer breakdown of what authors claim about pinning their software environment, from weakest (left) to strongest (right).

Capture claim distribution

Read: bar length = number of papers. Takeaway: 67.0% stop at “versions named in text”; only 14.5% claim a pin strong enough for code verification. Zero papers claim an immutable container digest.

Stage-2 funnel

Read: green = papers entering code check; grey = stops at Stage-1; mint = random audit sample from the rest. Takeaway: 14.5% advance; 85.5% terminate without code execution.

Claim typePapersShareMeaning
No pinning stated 10 5.0% The paper does not describe how to rebuild the software environment.
Floating versions (≥ / latest) 27 13.5% Versions use open ranges or “latest” — results may drift over time.
Exact tool names, environment not closed 134 67.0% Specific tool versions appear in text, but there is no single installable, closed environment.
Release snapshot (e.g. Bioconductor date) 8 4.0% Authors cite a dated release (e.g. Bioconductor version) plus interpreter — stronger, still unverified here.
Lockfile / conda env claimed 21 10.5% Authors point to a lockfile or equivalent — eligible for code verification (Stage 2).
Container with digest 0 0.0% Authors cite an immutable container digest — strongest claim; none observed in this sample.

Where documentation breaks down (step level)

4,005 computational steps across 200 papers. Each step is scored on four axes; the lowest axis caps the verdict.

Mean axis score (0–2)

Read: higher = better documentation on that axis. Takeaway: Version (environment) averages 0.57 — lowest of the four. At Stage-1, Version cannot reach 2 without code verification anyway.

Binding axis — what limits each step?

Read: for each step, the axis with the lowest score is the bottleneck (ties split evenly). Takeaway: Version alone caps ~49.9% of steps — environment capture is the corpus-wide constraint, not data sharing.

Verdicts & computational coverage

Step verdicts across all papers

Read: RI = partially ready, NI = needs investigation, X = blocked. R is always 0 at Stage-1. Takeaway: 54.8% of steps have a recoverable gap; 6.0% are blocked.

How much of each paper is computational?

Read: S/(S+W) = share of enumerated steps that are computational vs wet-lab (out of scope). Median 83% — most papers mix computation with lab work.

Median coverage
83%
Fully computational
26.0%
52 papers

Common documentation gaps

Gap types from the 13-type rubric dictionary — how many papers exhibit each gap (a paper can have multiple). Shares are computed over n = 199 papers: 1 of the 200 analyzed reports is excluded from gap tallies as malformed, so percentages in this table differ slightly from the paper-level figures elsewhere on this page.

GapPapersShare
Version not recorded 194 97.5%
Sample / cohort mapping missing 181 91.0%
Custom database or script not shared 171 85.9%
Stochastic step without seed 152 76.4%
Sub-pipeline outsourced 137 68.8%
Dead or broken link 131 65.8%
Wet-lab step (out of scope) 121 60.8%
Analysis code not deposited 120 60.3%
Provenance
Phase 1 file bootstrap · Generated 2026-08-05T10:30:13Z · API: /api/v1/tier1_survey/stats · Chain block #145

Canonical proof path (meeting anchor submit) is Phase 2. This preview reads from data/tier1_survey/collection.