Each test takes one claim from the DELPHI claim graph, writes down in advance what result would support it and what would refute it, and only then downloads real public data and runs the analysis. The rule is committed before the data are seen, so a refutation counts as much as a confirmation.
Loading tests.json…
Verdicts follow the pre-registered rule mechanically. "Mixed" means the measurement fell between the support and refute thresholds; "inconclusive" means the data could not be obtained or the test lacked power.
Every file is recorded with its download address, retrieval time, size and SHA-256 so that anyone can fetch the same bytes. Files over 1 MB are not kept in the repository; the fetch scripts re-download them.
| Dataset | Used by | Files | Size | Licence |
|---|
Each test lives in its own folder under claim-tests/ with four parts: result.json (the pre-registration, later filled with the measurement and verdict), fetch.py (downloads the data and records checksums), analysis.py (computes every number in the result) and derived/ (small tables the analysis wrote). The pre-registration is committed to git before fetch.py is first run, so the commit history shows the rule preceded the data.
Two cross-species tests exclude all birds before fitting, because the holdout thread keeps class Aves as an exam set that no thread may tune against. Bird rows can be scored once, later, with an --exam flag the builder does not run.
Rules in full: claim-tests/README.md. Nothing here is a nanopublication yet; tests become evidence nanopubs later through the normal test-then-production path.