Strust: an evidence-bounded project case study

The question

What would it take to trust a COBOL-to-Java conversion verdict?

My contribution

I founded Strust around a testing problem: run a COBOL reference and a Java candidate on controlled inputs, capture their observable behavior, and compare the results under an explicit contract.

The current case study audits the surviving v1 record. I preserved the historical paper, transcribed its ten-row result table, and linked five classes of claims to the artifacts needed to inspect them. The released analysis makes that arithmetic and evidence inventory reproducible.

What the evidence shows

The historical table records eight Pass rows and two Fail rows, giving a recorded pass rate of 80%. Its surrounding prose claimed 100%. The audit exposes that internal contradiction; the 80% figure is a count of document entries, not an independently validated conversion accuracy.

Within the locally declared evidence graph, two of nineteen required claim-artifact links survive. Recorded table arithmetic is the only fully auditable class among five modeled outcome claims. No single missing artifact would complete any of the four blocked claim classes.

Matrix linking five historical claim classes to ten artifact classes; only the historical paper and transcribed ledger are present — Case study figure 2 · Evidence inventory. Two of nineteen required claim-artifact links survive in the declared graph. Only recorded table arithmetic has complete coverage. This is an inventory of the preserved evidence, not an externally validated score of evidence quality.

Where the evidence stops

The recovered package does not preserve the executable COBOL/Java subjects, fixtures, raw traces, comparator policy, or execution environment. It cannot establish comparative accuracy, feasibility, performance, or enterprise scale.

The evidence graph uses locally declared requirements, not an external certification. Complete coverage would make a claim inspectable, not necessarily correct. The audit covers the recovered record; private or lost artifacts may once have existed.

A follow-up should freeze the comparison policy before evaluating blinded cases, retain raw outputs and environment identities, and report MATCH, DIVERGE, and INCONCLUSIVE separately. Rounding and encoding deserve explicit fixtures, but the historical failure notes alone do not establish their causes.

Read and reproduce

STRUST v1: An Evidence-Bounded Project Case Study

The package contains the recovered historical PDF, transcribed ledger, audit code and tests, generated figures, and a claim/checksum manifest. The audit runs without network access. It does not contain a runnable reconstruction of the historical conversion system.

Audit code & source (ZIP, 1.2 MB)

View citation and BibTeX

Weale, George. STRUST v1: An Evidence-Bounded Project Case Study. Author-hosted project case study; 2026 evidence audit of the 2025 project. https://georgeweale.com/pdfs/research/2026-strust-evidence-audit.pdf?v=affiliations-20260804

@misc{wealeStrustEvidenceAudit,
  author = {Weale, George},
  title = {{STRUST} v1: An Evidence-Bounded Project Case Study},
  note = {Author-hosted project case study; 2026 evidence audit of the 2025 project},
  url = {https://georgeweale.com/pdfs/research/2026-strust-evidence-audit.pdf?v=affiliations-20260804}
}