# Final Independent Integrity Audit

Audit date: 2026-07-15  
Scope: `AGENTS.md`, `paper/paper.md`, `paper/references.bib`, `data/evidence_ledger.csv`, `data/source_index.csv`, `limitations.md`, `output/citation_audit.md`, central source texts, bibliography URLs, and DOI metadata. This audit did not modify the research files.

## Verdict

**PASS — the research-integrity release gate is cleared.**

All seven blockers in the previous audit have been resolved. The corrected paper is cautious about what the private genetic account can and cannot establish, and its conclusion is supported at the level claimed: one submitted tree was reported as unlike RD2's undisclosed Geisha comparator and very close to an undisclosed SL09 reference, but the public record does not establish an exact SL09 match, authenticated T2722 comparator, sample custody, population-wide identity, or a Kenya-to-Peru route.

No new material research-integrity blocker was found. This verdict covers the manuscript, bibliography, evidence ledger, source inventory, citation audit, and stated limitations. It does not certify HTML, DOCX, PDF, or website artifacts that have not yet been rebuilt and visually inspected from the final research files.

## Current mechanical results

- Paper length: **9,618 whitespace-delimited Markdown words before the References heading**, within the required 6,000–10,000 range.
- Paper citations: **51 unique citation keys**.
- Bibliography: **51 entries**, all keys unique. No cited key is missing, and no bibliography entry is uncited.
- Evidence ledger: **69 rows** with **69 unique claim IDs**, all marked `verified`. All 51 cited source keys occur in the ledger; every ledger key resolves to the bibliography and to an `included=yes` inventory row.
- Source inventory: **166 rows** with **166 unique source IDs**: 53 included and 113 excluded. Every excluded row has a reason, and no citation key points only to an excluded row.
- Duplicate checks: **zero duplicate canonical DOIs and zero duplicate canonical URLs**. The remaining identical Pruvot-Woehl title is an intentional audit-trail pair: the bad `qsz028` seed is explicitly excluded and the verified `qsz003` record is retained.
- BibTeX parses cleanly. Pandoc 3.9 cite processing succeeds without missing-key warnings and generates exactly one visible `References` heading followed by the CSL bibliography.
- All **15 DOI-bearing entries** resolved through Crossref. After normalizing HTML tags, braces, capitalization, and typographic dashes, their titles and publication years matched the cited works.
- A fresh URL pass returned HTTP 200 for 46 of 51 bibliography URLs, 403 for three publisher/news endpoints, one transient timeout, and a command-line 404 for Port2050. The 403/timeout results do not indicate incorrect metadata; the underlying DOI records or pages were independently inspectable.
- `git diff --check` passed.

## Resolution of the seven prior blockers

### 1. Claim-ledger completeness — resolved

The 17 formerly absent cited keys now have source-specific rows `ADD-001` through `ADD-017`, with locators, excerpts, paraphrases, strengths, and limitations. These include the 2024 ACE auction record, the two conflicting 2020 reports, Fischer and Traore on value formation, the WCR SNP database, the Geisha genome resource, official Inkawasi records, the second OpenAlex query, the CATIE negative-inspection record, and the processing/roast comparator.

Two further rows improve conceptual coverage: `ADD-018` separates Gesha/Geisha spelling from genotype, and `ADD-019` records the ambiguity between Indian Selection 9 and Kenyan Scott Laboratories SL.9. Every source cited by the final paper is now represented in the ledger.

### 2. Port2050 reproducibility — resolved for the limited claim

The bibliography, ledger, inventory, and paper now use the live retailer URL:

`https://port2050coffeeroasters.com/products/dwight-aguilar-sl09-inca-geisha`

The underlying page was opened directly in a browser during this audit. It displayed the product title `DWIGHT AGUILAR SL9 / INCA GESHA`, `FINCA NUEVA ALIANZA`, `Origin: Santa Tereza / Peru`, `Variety: SL9 / Inca Gesha`, and `Producer: Dwight Aguilar Masias`. The paper uses it only as Tier 3 evidence that this marketing language appeared; it does not use the page as genetic evidence.

Command-line retrieval still returns 404, and a locale-sensitive Safari visit redirected to the storefront home page. The bibliography and ledger disclose the command-line failure and absence of a retained HTML snapshot. Because the underlying product page was opened and inspected, the open-and-inspect rule is satisfied; because access is fragile and the page is mutable, it remains a reproducibility caution rather than a release blocker.

### 3. Inventory duplicates and contradictory screening — resolved

The duplicate municipal nursery and competition records were removed, leaving one canonical included row for each government page. The source inventory now reports no duplicate canonical DOI or URL and no duplicate citation key.

The Fischer working-paper record is retained as an excluded, redundant discovery record, while the peer-reviewed 2021 journal version is the included and cited record. The inventory no longer assigns contradictory substantive decisions to the same DOI.

### 4. WCR database metadata — resolved

`WCR2023Database` now records the official visible date **2023-08-15** and the exact headline **“World Coffee Research releases open-access Coffea arabica genetic fingerprinting database.”** The bibliography and source inventory agree. The inspected page directly supports the paper's 45-SNP, 23-variety, and more-than-30,000-sample statements.

### 5. References section — resolved

`paper/paper.md` now ends with a visible `# References` heading and citeproc reference container. An in-memory HTML conversion produced one `<h1 id="references">References</h1>` followed by one 51-entry bibliography.

### 6. Bibliography rendering — resolved

Pandoc now renders the previously malformed fields correctly, including:

- **Castillo Avilés, Gisselle Aimeth**;
- **Deacue Fields III**;
- the complete Texas A&M institution as one corporate entity; and
- compound Spanish family names in the Choque-Quispe record.

No new malformed author or organization was found in the generated reference list.

### 7. PROMPERÚ URL — resolved

The bibliography now uses the working official repository bitstream URL rather than the failing handle. A fresh download returned HTTP 200. The remote file and `raw/pdfs/PromPeru2019.pdf` were both 3,842,993 bytes and had the same SHA-256 hash:

`8f2adb9c61c9baef5f63d457b692e2a2d8801a85a367862ba07a23bd9fdfed97`

## Central scholarly checks

- Feran's live page contains the reproduced report wording that the fingerprint was previously unseen, belonged to the Ethiopian Legacy group, was very close to SL09, and was not close to Geisha. It does not disclose the Geisha or SL09 reference accession. The abstract, discussion, limitations, and conclusion now preserve that distinction and condition any T2722 implication on an authenticated comparator and sound sample custody.
- Pruvot-Woehl et al. support the eight-SSR authentication workflow, CATIE T.02722 reference, allelic-phenotype limitation, and the 39% exact/24% close/37% unrelated distribution for 88 Gesha-labelled submissions. The paper does not transfer that distribution into a classification of Peruvian material.
- Montagnon et al. support placing SL-09 among several cultivars in a broad Ethiopian Legacy cluster. The paper correctly treats cluster or nearest-reference similarity as weaker than exact identity.
- Jones's 1956 scan supports the unknown source block and historical morphology, yield, quality, and disease statements for Kenyan SL.9. The paper does not convert that sparse record into parentage or a Peru transfer route.
- WCR's Geisha catalog supports T2722-derived Panamanian Geisha as a distinct reference and warns that genetically different plants have circulated under the name. It also supports the Gesha/Geisha spelling discussion without making spelling an identity test.
- Competition, retailer, roaster, and government records are consistently used for labels, chronology, prices, scores, locations, and representations—not as scientific authentication. Regional Typica/Catimor and non-Peruvian sensory studies are consistently identified as context or mechanism rather than direct evidence about authenticated Inca Gesha/SL09 material.

## Residual cautions, not release blockers

1. **Port2050 remains fragile.** Browser access succeeded, but command-line access did not and no local snapshot exists. If the retailer removes the page, the claim should be restated as an access-dated historical observation or supported with a lawful archive.
2. **Two retained inventory sources are unused in the final paper.** `Velasquez2022` and `Scenery2026` are marked `included=yes` but are neither cited nor ledgered. No manuscript claim depends on them. If `included` is intended to mean “used in the final synthesis” rather than “retained as eligible background,” they should later be reclassified; this does not affect the current citation graph.
3. **One locator can be made more precise.** Ledger row `ADD-018` describes the WCR spelling passage as an “introductory summary,” while the live page displays that wording under `History`. The quotation, URL, access date, and paper citation are correct, so this is a minor locator-label cleanup rather than a support failure.
4. **Publication artifacts remain a separate gate.** The final HTML, DOCX, PDF, and research website must be rebuilt from this corrected manuscript and then checked for visible Abstract, Keywords, both central tables, correct names, live citation links, a References heading, and complete bibliography rendering.

## Release decision

The corrected research record may proceed to artifact generation and website publication. No unresolved citation, claim-calibration, bibliography, inventory-deduplication, or central-source issue requires another research rewrite. Public release should occur only after the final document and website builds complete their separate structural, visual, and link QA.