# Submission claim audit — 2026-09-25

## Decision

The paper must be submitted around a narrow, auditable claim: **GWAS
Harmonizer is a browser-based, uncertainty-aware workflow that exposes
harmonization decisions and preserves row-level audit evidence.** It is not
yet defensible to claim universal numerical accuracy, universal 100% retention,
or a completed four-tool head-to-head benchmark across 54.9 million rows.

`Supplementary_Table_S10_four_tool_benchmark_55M_variants.md` and its Figure 3
are withdrawn from evidentiary use. They must not be cited by the abstract,
main text, cover letter, Zenodo record, website, or submission metadata. The
authoritative narrative source remains
`manuscript/GWAS_Harmonizer_revised_manuscript_2026-08-17.docx`; the Markdown
draft is a working aid only until it is reconciled with that document.

## Claim inventory

| Claim | Status | Traceable evidence | Submission wording / action |
|---|---|---|---|
| Controlled non-palindromic SNP transformations recovered the specified allele and numeric fields in 180,000 scenario-row evaluations. | Scoped support | `claim_evidence.csv` C01; controlled-perturbation manifest | State the scenario scope and selected 100,000 source variants. Do not generalize to all variants or all inputs. |
| Injected rsID blanks, conflicts and p-value inconsistencies were detected at the measured controlled-test rates. | Scoped support | `claim_evidence.csv` C02–C03 | Keep the precision/recall and sensitivity endpoints separate; describe the injected classes. |
| The clean FinnGen subset shows input concordance for GWAS Harmonizer and GWASLab. | Observational support | `corrected_comparison_results.json`; `claim_evidence.csv` C04 | Call it input concordance among unambiguous keys, not ground-truth accuracy. |
| Four public 100k-scale subset runs exercise build handling, reference alignment, palindromes, indels, rsIDs and audit accounting. | Scoped support | Terminal evidence packages cited in the working Markdown manuscript | Retain only figures that can be rechecked against each package's `harmonization_summary.json`, `paper_metrics.csv` and `evidence_manifest.json`. |
| A three-format full-file comparator provides supplemental feasibility/retention evidence. | Scoped support | `results/benchmarks/three_format_competitors_20260922/REPORT.md`; per-tool `run_metadata.json` | Keep outside the prespecified six-dataset benchmark. Do not claim native parser performance, controlled truth, or a four-tool Pan-UKBB comparison. |
| GWAS Harmonizer achieved 100% scientific accuracy/concordance across 54.9M variants. | Withdrawn | `claim_evidence.csv` C05; `FINDINGS.md` | Remove everywhere. The legacy rows total 55,009,416, mix independent studies, samples, fixtures and repeated perturbations, and cannot support this endpoint. |
| GWAS Harmonizer is faster/easier than comparators. | Pending | `claim_evidence.csv` C06/C09 | Omit until a prespecified matched runtime or usability protocol exists. |
| The odds-ratio/confidence-interval transformation path is validated. | Pending | Working manuscript §3.3 explicitly reports no exercised input | Run a documented GWAS-SSF ratio/CI validation before making the claim, or retain the limitation. |
| Full-file FinnGen and Suzuki establish the paper's six-dataset benchmark. | Pending | Paper context says only 100k subsets are currently in scope | Either complete and archive full runs with explicit study semantics, or describe the validation as subset-scale. |
| Pan-UKBB is a Harmonizer-vs-comparator result. | Pending / unavailable | `FORMAT_COVERAGE_20260915.md` | No local Harmonizer evidence bundle exists; exclude it from any Harmonizer comparison. |

## Immediate manuscript edits required

1. Replace the current abstract's “preliminary four-file exercise” paragraph
   only after its four packages and counts are rechecked against the cited
   evidence files.
2. Remove every reference to `Supplementary Table S10` / Figure 3 as a 54.9M
   head-to-head accuracy benchmark from the authoritative Word manuscript and
   live Google Doc. A new supplement can report the scoped controlled and
   observational endpoints above.
3. Retain the clearly stated limitations: no ratio/CI test, incomplete
   full-file coverage, no Pan-UKBB Harmonizer bundle, and no usability or
   matched-runtime evidence.
4. Before author review, create a compact Table S1 with source URL/access date,
   input checksum, exact parsed rows, semantic confirmation source, run ID,
   package checksum, and paper-use decision for each included dataset.

## Submission gates after claim freeze

- Confirm funding and conflict statements; select a repository license.
- Archive the exact release, tests, benchmark manifests and reconstruction
  instructions in Zenodo; add its DOI and a data-availability statement.
- Verify Chromium, Firefox and Safari, and document local/container deployment.
- Render the revised submission source and keep the Application Note within
  four pages (approximately 2,600 words, or 2,000 words plus one figure).