GLOSSARY / SOURCE AND EVIDENCE
Use version words precisely.
A version number, data digest or test report identifies one part of a research artifact. It does not certify all the other parts.
- Application release
- The version of the packaged software and its interface. It can change while equations or sample data remain the same.
- Catalogue
- The architecture, scenario, lesson or study definitions and their input descriptions. The 59-study catalogue has no declared semantic version in this source; its registry records a content SHA-256 separately.
- Data kind
- A label such as synthetic, observed or user-supplied-unverified that describes where inputs came from. User-supplied does not mean independently checked.
- Dataset version
- An identity for a bundled set of example records. The hybrid preset bundle declares its own schema, data version and digest. Study examples currently have a digest, not a named dataset version.
- Model version
- An identifier for a particular equation and implementation family, such as
terminal-greedy-hourly-v1or the shared advanced study versionroadmap-screen-v1. - Input/result schema
- The allowed fields, units, bounds and output shape. The 59-study definitions publish numeric input bounds and units where supplied. A separate result
schema_versionis not declared in the current study envelope and is marked U in the registry. - Fixture
- A reproducible example input and output used by software. A generated default fixture is useful for regression; it is not automatically a hand calculation, measured dataset or independent scientific benchmark.
- Seed
- A specified starting value for a stochastic experiment’s random generator. A seeded simulation is reproducible under its stated software/model conditions; it is not calibrated by being repeatable.
- Convergence
- A solver-specific numerical stopping or feasibility condition. The common study result format does not require a shared convergence field, so each model card records what its default output exposes and marks unreviewed evidence unknown.
- Evidence digest
- A SHA-256 fingerprint that can detect content changes. It does not establish who created a file, whether its inputs are true, or whether the model is correct.
- Validation tier
- The kind of evidence available, from schema/software checks through independent reference cases and field comparison. Software tests in this repository are not independent scientific validation.
- U / NOT RUN
- U means the source inspected here does not establish that fact. NOT RUN means the named acceptance action was not executed in this documentation change.
Keep the layers separate
Before comparing two results, identify the application release, catalogue digest, data identity, model version, schema and time/unit basis. Then read the actual inputs and limitations. A matching label or digest alone cannot make unlike runs comparable.
See all 59 model cards, the capability guide, and the repository’s shared study input/result contract.