Weather Data Reproducibility

Weather Data Reproducibility creates and verifies a compact provenance manifest for weather inputs and derived artifacts. It records what a run requested, what data was actually resolved, which subset was used, how it was processed, and which outputs were produced. It establishes lineage and integrity rather than forecast skill, observational truth, or scientific validity. The skill does not discover or download data by itself.

The source distinguishes four claims. Traceable means the history can be audited. Replayable means the identified inputs remain obtainable and the recorded procedure can run again. Bitwise reproducible means identical output bytes are expected in a pinned environment. Scientifically reproducible means equivalent methods support the same conclusion within a stated tolerance. A URL and run time alone do not justify the stronger claims.

Use it for auditable maps, soundings, time series, datasets, benchmarks, or releases, especially when inputs come from mutable URLs, mirrors, caches, or byte-range subsets. It can also help explain why two outputs differ or retain provenance before temporary weather files are deleted.

Do not label a run replayable when deleted inputs came from a mutable or short-lived source and no immutable copy remains. A manifest cannot restore unavailable data. Floating dependencies, lossy images, parallel reductions, platform behavior, licensed software, and hardware or provider requirements can limit exact reproduction.

Inputs include request intent, resolved data identities, selections, processing parameters, environment details, and completed artifacts. The output is a versioned JSON manifest with a declared reproducibility level and tolerance. Use null for an intentionally absent value and omit unknown meanings rather than filling fields with guesses.

For each source, record the exact locator, provider and dataset revision where available, identity metadata, content size and encoding, retrieval time, fallback provider, and applicable attribution. ETags are opaque identity values and are not always MD5 hashes. Compute a real content hash for materialized bytes.

Subset provenance includes GRIB inventories and hashes, selected rows, expressions and ordered inclusive byte ranges; Zarr arrays, chunks and slices; API parameters and raw-response identity; station, radar, or satellite selections; and requested versus selected coordinates and distance. Hash the assembled subset separately.

Time fields preserve their meanings in UTC: model initialization, forecast lead and valid time; observation, correction and ingestion time; radar or satellite scan intervals; file creation and retrieval time. Record relevant calendar handling. Processing records cover application and source revision, dirty status, libraries, runtime, result-changing arguments, unit conversions, quality decisions, interpolation, aggregation, seeds, and environment or image identity.

Hash artifacts after writers close. Finalize an atomic write with:

os.replace(temporary, path)

Replay validates schema, re-resolves input identities, checks sizes and hashes, reapplies selections, recreates the environment, and writes artifacts into a new directory. Compare bytes or scientific metrics according to the declared claim, report the first mismatch, and create a new manifest referencing the original instead of overwriting it.

The workflow pairs with model, observation, radar, or satellite access skills and covers GRIB2, NetCDF, Zarr, APIs, and object storage. Its local hashing and JSON example uses Python's standard library. Final artifacts and manifests precede cleanup of request-owned temporary files; shared caches must remain untouched.

Keep credentials, cookies, signed queries, private storage details, and unnecessary local identities out of shared manifests. Validate paths, retain TLS checks, and preserve access and licensing restrictions. A hash proves integrity, not that untrusted content is safe.

This skill supports weather analysts and developers who need to audit a processing run, compare outputs, or make a bounded and testable reproducibility claim.