02 / Reproducible handoff
Build a research handoff folder without customer documents
Keep three executable fixtures and a small manifest together, while avoiding the fiction that they were automatically extracted from one source.
A reproducible example should let someone repeat a finding without sending you a confidential report. The three tool repositories provide synthetic fixtures, but their filenames alone do not establish a relationship. Make the relationship explicit—and label what is only a demonstration.
Keep independent inputs independent
Clone the three public repositories into one working directory. Keep their original examples/ folders intact. The ledger fixture contains invented disclosure rows; the bilingual files are independently authored text; the PDF fixture contains blank geometry pages. They are not three automatically extracted views of one real filing.
handoff/
filing-evidence-ledger/examples/sample.json
bilingual-markdown-map/examples/source.md
bilingual-markdown-map/examples/zh-Hant.md
pdf-geometry-inspector/examples/sample.pdf
manifest.jsonUse a manifest for assumptions, not secrets
{
"kind": "synthetic-training-bundle",
"automatic_cross_format_extraction": false,
"review_questions": ["missing-period", "missing-image", "rotation"]
}Record each tool version or commit, the commands used and the output folder. Do not put API keys, customer names or private paths in the manifest. Re-run the CLIs rather than editing their JSON outputs by hand. The product sites provide the individual commands: evidence ledger, bilingual map and geometry inspection.
Give the recipient a stopping point
Expected findings are two incomplete evidence rows, one unpaired source chapter and an intentionally mixed three-page PDF. Those outcomes prove that the fixtures exercised the tools; they do not validate a customer's documents. Ask the recipient to reproduce the examples first and only then substitute authorized local data.
This layout uses ordinary folders, not a new workflow framework. No tool watches another folder or uploads results. Once you use real documents, make the source-to-excerpt and source-to-translation relationships manually and preserve the originals under your own access controls.